text
stringlengths 0
643k
⌀ | meta
stringlengths 137
151
⌀ |
|---|---|
# Introduction
Despite the clear indication that neutrinos do have mass which is reminiscent of the first established departure from the Standard Model (SM) of particle physics, it is still unclear whether they are of Dirac or Majorana type in nature. The latter possibility is related to the lepton number violation (while the first one conserves it) by heavy right-handed neutrino (RHN) mass(es) and is capable of explaining the smallness associated to the neutrino mass. Interesting consequences would follow, if we promote this lepton number symmetry (LNS) to a global one, say \(U(1)_{L}\), and consider it to be broken spontaneously so as to make RHNs massive. In this case, a massless Nambu-Goldstone boson (NGB) called Majoron results. While the mass of a Majoron is related to an explicit (soft) breaking of the LNS, analogous to the quark mass in the case of pion, such a particle carrying very suppressed (by the scale of symmetry breaking) interactions, in general, renders itself as a promising dark matter (DM) candidate. Previous studies with such Majoron field (\(\chi\)) as a freeze-out kind of DM have mostly been engineered through SM Higgs \((H)\) portal interaction of the form \(\lambda_{\chi H}\chi^2 H^\dagger H\), an explicit LNS breaking term, which helps in realizing the DM relic satisfaction. Such an interaction is suggestive of the scalar singlet DM scenario and obviously, there exists a one-to-one correspondence between the mass of the DM (\(m_\chi\)) and the LNS breaking parameter \(\lambda_{\chi H}\). However, contrary to the natural expectation that an explicit symmetry breaking parameter should be sufficiently small in t'Hooft's sense, it is observed that \(\lambda_{\chi H}\) needs to be large enough (\(\mathcal{O}(0.1)\) or so) in order to satisfy the correct relic density. Secondly, with the XENON-1T results, the entire parameter space of \(m_\chi\) ranging below TeV (except the Higgs resonance region) is basically ruled out in this minimal Majoron scenario. The Majoron also possesses a derivative coupling with its CP-even partner, similar to other pseudo-Nambu Goldstone Bosons (pNGB), through which it can annihilate into SM particles via \(s\)-channel mediated process by the CP-even scalar leading to a possible relic satisfaction. In case of pNGB as DM (not a Majoron), such a scenario can evade the direct detection experiment bounds as a result of the small momentum transfer of DM through the derivative coupling causing suppression in elastic scattering amplitude with nuclei. However, in case of Majoron as DM, it fails to explain the stability criteria as Majorons would decay to light neutrinos (\(\nu\)) via active sterile mixing connected to the neutrino mass generation in type-I seesaw. On the other hand, Majorons may also emerge as Feebly interacting massive particle (FIMP) DM where it is produced in the early Universe from the decay of some particle, a natural choice of which is the RHN (\(N\)). As the Majoron interaction strength with the SM particles is expected to be suppressed by the \(U(1)_{L}\) breaking scale (the seesaw scale), treating Majoron as FIMP type DM remains as a viable option. However, from the previous studies, it becomes apparent that with the tree-level coupling between the RHNs and the scalar singlet (the one responsible for breaking the LNS), sufficient production of Majorons cannot take place from the decay of RHNs. This is primarily because the decay rate (\(N \rightarrow \chi \nu\)) is proportional to the tiny light neutrino mass on top of the usual suppression by LNS breaking scale. However, the situation changes if a tiny LNS breaking Higgs-portal coupling \(\lambda_{\chi H}\) is introduced. The relic satisfaction[^1] of Majorons produced from the decay of the SM Higgs boson \(h \to \chi \chi\) requires the portal coupling \(\lambda_{\chi H} \sim O(10^{-10})\) which in turn fixes the Majoron mass \(m_\chi\) having a unique value \(\sim\) 3 MeV. Though such small portal coupling obeys the naturalness criteria, it suffers from a fine-tuned situation in terms of Majoron mass. Another possibility is to realize non-thermal Majoron production from the annihilation of SM Higgs via UV freeze-in framework where the Majoron mass is found to be relatively heavy (contrary to the light pNGB boson in UV freeze-in scenario ). In this work, with the aim of broadening the Majoron mass range (toward the lighter side) while the minimality and naturalness are retained, we propose a new production mechanism for Majoron as a FIMP-type DM. Instead of introducing any Higgs-portal coupling with Majorons, we introduce here a dimension-5 LNS breaking operator. Being suppressed by the cut-off scale, it can be regarded as a soft breaking of LNS which takes care of the production of Majorons via RHN annihilations. Also, we keep the framework minimal in the sense that no additional fields other than the two SM singlet right-handed neutrinos (related to neutrino mass generation) and a LNS breaking (SM singlet) red complex scalar field, are required. We primarily focus on the lighter side of Majoron mass \([\mathcal{O}(\rm{keV})-\mathcal{O}(\rm{GeV})]\), which remains attractive as this mass regime can be experimentally probed by several direct and indirect searches. For example, Majorons of MeV scale and beyond can be potentially detectable via mono-energetic neutrino flux in experiments like Borexino, KamLAND, and Super-Kamiokande (SK). Also, sensitive bounds from various \(\gamma\)-ray observations such as INTEGRAL, COMPTEL/EGRET, Fermi-LAT etc. are applicable in this mass regime.
# Majoron as Dark matter {#section2}
As we build our framework extending the original Majoron model where a Majoron can be considered as a dark matter, a brief discussion on the basic structure of it and related limitations are relevant to discuss first. The SM is extended by including two singlet right-handed neutrinos (\(\mathcal{N}_{R_{i=1,2}}\)) and a singlet complex scalar field (\(\Phi\)) such that a global lepton number symmetry \(U(1)_{L}\) prevails. As a result, a lepton number of \(L =-2\) units is assigned to \(\Phi\) while the SM lepton doublets (\({L}_\alpha\)) as well as the RHNs carry a lepton number \(L=1\). The renormalizable part of the Lagrangian (in the charged lepton diagonal basis) involving neutrinos is given by \[-\mathcal{L}\supset \frac{f_{i}}{2}\Phi \overline{\mathcal{N}^c_{R_i}} \mathcal{N}_{R_i}+y^\nu_{\alpha i}~\overline{L}_\alpha \tilde{H}\mathcal{N}_{R_i} +h.c., \label{lag}\] where, \(H\) is the SM Higgs doublet and \(y^\nu_{\alpha i}\) is the neutrino Yukawa coupling with \(\alpha=e,\mu,\tau\) and \(i =1,2\). The Majorana masses of RHNs are generated after the \(U(1)_L\) symmetry breaking via \(\Phi\) vacuum expectation value (vev) \(v_{\phi}\). Without loss of generality, we consider the generated RHN mass matrix \(M_R = f v_{\phi}/\sqrt{2}\) to be diagonal with \(f ={\rm{diag}} (f_1, f_2)\). Once the electroweak (EW) symmetry is broken by the Higgs vev \(v\) as \(\langle H \rangle = (0, v/{}\sqrt{2})^T\), active neutrino masses are generated through this minimal type-I seesaw mechanism \(m_\nu=m_D M_R^{-1} m_D^T\), where \((m_D)_{\alpha i}=y^\nu_{\alpha i}v/\sqrt{2}\) \((v=246~ \textrm{GeV})\). The \(U(1)_L\) symmetric scalar potential involving \(H\) and \(\Phi\) is given by, \[V(H,\Phi)=V_H-\frac{{\mu_\Phi^2}}{2}|\Phi|^2+ \frac{\lambda_\Phi}{2} |\Phi|^4+\lambda_{H\Phi}|H|^2|\Phi|^2, \label{pot}\] where \(V_H =-\mu_H^2 H^\dagger H+\lambda_H (H^\dagger H)^2\) is the usual SM Higgs potential and \(\lambda_{H\Phi}\) is the Higgs portal coupling of \(\Phi\). The stability of the potential is guaranteed by: \(\lambda_{\Phi},\lambda_H ~{\rm{and}}~ \lambda_{H\Phi}+\sqrt{2\lambda_H \lambda_\Phi}\geq 0.\) Once the \(U(1)_L\) global symmetry is broken spontaneously, \(\Phi = (v_\phi+\phi+i\chi)/\sqrt{2}\) (we use the linear representation throughout), a NGB (the CP-odd component) or the massless Majoron \(\chi\) results. The Majoron mass can be generated once an explicit LNS breaking term, \[{-\mathcal{L}_{\rm{LNB}} =-\frac{m^2}{4}(\Phi^2+{\Phi^*}^2)}, \label{soft-breaking}\] is introduced. Such a term breaks the \(U(1)_L\) to \(Z_2\) (under which \(\Phi \rightarrow-\Phi\)) and induces a mass for the Majoron as \(m_\chi^2= m^2\). This mass term is however expected to be small (soft breaking) in t'Hooft's sense as \(m \rightarrow 0\) limit enhances the overall symmetry of the framework. Furthermore, the portal coupling can also be considered to be small (\(\lambda_{H \Phi} \ll 1\)) which is technically natural from the point of view of enhanced Poincare symmetry in the limit \(\lambda_{H \Phi} \rightarrow 0\). Considering a large hierarchy between the scales of lepton number and EW symmetry breaking along with tiny Higgs portal coupling, the mixing generated between the CP-even parts of \(\Phi\) and \(H\) turns out to be negligible which results into the following masses of the scalar fields as \[{m_h^2 \simeq 2\lambda_H v^2,~ m_{\phi}^2 \simeq \lambda_\Phi v_\Phi^2,~ m_{\chi}^2 \simeq m^2},\] where \(m_h\) is the mass of the SM Higgs boson as 125 GeV.
## Seesaw Mechanism and Interactions of Majorons {#seesaw-mechanism-and-interactions-of-majorons .unnumbered}
Using the linear representation of the \(U(1)_L\) breaking field \(\Phi=\left( {v_\phi+\phi+i\chi}\right) /{\sqrt{2}}\), the following interaction of Majoron results via Eq. [\[lag\]](#lag){reference-type="ref" reference="lag"}, \[-\mathcal{L} \supset \frac{if_i}{2\sqrt{2}}\chi \overline{\mathcal{N}^c_{R_i}} \mathcal{N}_{R_i} + h.c.. \label{Majoron-1}\] Furthermore, after the EWSB, a mixing between left and right-handed neutrinos takes place via the Dirac mass \(m_D = y^\nu v /{\sqrt{2}}\) and consequently the mass terms involving \(\nu_L\) and \(N_R\) can be written as, \[-\mathcal{L}_{m}= \begin{pmatrix} \overline{\nu_L} & \overline{\mathcal{N}^c_R} \end{pmatrix}\begin{pmatrix} 0 & m_D \\ m_D^T & M_R \end{pmatrix} \begin{pmatrix} \nu_L^c \\ \mathcal{N}_R \end{pmatrix}, \label{mass1}\] where the generation indices are suppressed. This yields the light and heavy neutrino mass matrices \(m_{\nu}\) and \(M\) (\(\simeq M_R\), already diagonal) respectively in the rotated basis \((\tilde{\nu}_L^c~ ~\mathcal{N})^T\), where[^2] \(\tilde{\nu}_L^c =-i(\nu_L^c-\theta \mathcal{N}_R)\) with the active-sterile mixing matrix \(\theta = m_D^* M_R^{-1}\) and \(\mathcal{N} \equiv \mathcal{N}_R\). A further diagonalization of \(m_{\nu}\) by the \(U_{PMNS}\) leads to the diagonal mass matrix diag\((m_{\nu_j}, M_i)\) in the mass eigenstate basis \((n_j ~~\mathcal{N}_i)^T\). Defining the Majorana mass eigenstates of light (\(\nu_j = n_j + n^c_j\)) and heavy neutrinos (\(N_i =\mathcal{N}_i+\mathcal{N}^c_i\)) by \(\nu_j\) and \(N_i\) respectively, the interaction terms of Majoron (followed from Eq. [\[Majoron-1\]](#Majoron-1){reference-type="ref" reference="Majoron-1"}) with light and heavy neutrino mass eigenstates can be written as \[\begin{aligned}
-\mathcal{L}_{\chi \nu \nu} &=&-\sum_{j,k}\mathcal{L}_{\chi \nu_j \nu_k} \nonumber\\ &=&-\frac{\chi }{2\sqrt{2}} \sum_{j,k} \left(\sum_{i}if_i \overline{\nu_j} P_R \nu_k V^T_{ji} V_{ik} +h.c. \right),\nonumber\\ \label{L3} \end{aligned}\] and \[\begin{aligned}
-\mathcal{L}_{\chi N \nu} &=&-\sum_{i,j}\mathcal{L}_{\chi N_i \nu_j} \nonumber \\ & = & \frac{\chi}{2\sqrt{2}} \sum_{i,j} f_i \left(\overline{\nu_j} P_R N_i V^T_{ji} + \overline{N_i}P_R \nu_j V_{ij}\right) + h.c.,\nonumber\\ \label{L4} \end{aligned}\] where \(V = \theta^\dagger U\).
## Production and Decay of Majorons {#production-and-decay-of-majorons .unnumbered}
As we are looking for Majoron as FIMP, the interactions of Majorons mentioned above are suggestive of its primary production via \(N_i\to\chi \nu\), the associated decay width of which is given by, \[\Gamma_{N_i\to\chi \nu}= \frac{M_i^3}{32 \pi v_\phi^2} \sum_{j=1,2,3} |V_{ij}|^2. \label{d1}\] Such a decay width can be shown to be approximately proportional to light neutrino mass (\(\Gamma_{N_i\to\chi \nu}\simeq \frac{1}{32\pi}\left(\frac{M_i}{v_\phi} \right)^2 \sum_j m_{\nu_j}\)) and hence suppressed. Also, this particular decay channel of \(N_i\) opens up only after the electroweak symmetry breaking (EWSB) as the interaction in Eq. [\[L4\]](#L4){reference-type="ref" reference="L4"} contributing to this decay is proportional to the active-sterile neutrino mixing. Considering the RHNs are in thermal equilibrium at a temperature \(T > M_i\) and the branching of this decay remains sizeable enough compared to other decays of \(N_i\) (\(e.g. ~N_i\) to SM ones via neutrino Yukawa interactions), the relic contribution of Majoron can be expressed as, \[\Omega_\chi h^2 \approx \frac{1.09\times 10^{27}}{g_\star^\mathcal{S} \sqrt{g_\star^\rho}}m_\chi \sum_{i}\frac{g_i\Gamma_{N_i\to\chi \nu}}{M_{i}^2}, \label{omega}\] where \(g_\star^{\mathcal{S},\rho}\) is the effective number of degrees of freedom in the bath and \(g_i=2\) denotes the internal spin degrees of freedom of RHN. On the other hand, following Eq. [\[L3\]](#L3){reference-type="ref" reference="L3"}, Majoron can decay into light neutrinos having the decay width \[\begin{multlined} \Gamma_{\chi\to \nu \nu} = \frac{m_\chi}{16\pi v_\phi^2} \sum_{j} m_{\nu_j}^2\\ \simeq \frac{1}{10^{19}s}\left( \frac{m_\chi}{1~\textrm{MeV}}\right) \left( \frac{8\times10^8~\textrm{GeV}}{v_\phi}\right)^2 \left(\frac{\sum_{j}m_{\nu_j}^2}{2.6\times10^{-3}~\textrm{eV}^2} \right). \end{multlined} \label{d0}\] In getting the sum of the light neutrino mass-squared in Eq. [\[d0\]](#d0){reference-type="ref" reference="d0"}, the best fit values of atmospheric and solar mass-splittings are used from neutrino oscillation data where we consider a normal hierarchy of light neutrinos along with \(m_{\nu_1}=0\). The stability criteria of Majoron to be a viable dark matter candidate as given by \[\Gamma^{-1}_{\chi\to \nu\nu} > \tau_{U}, \label{stability}\] needs to be satisfied, where \(\tau_U\) is the lifetime of the universe \(\sim \mathcal{O}(10^{19})~\rm{sec}\). Using Eqs. [\[omega\]](#omega){reference-type="ref" reference="omega"} and [\[d0\]](#d0){reference-type="ref" reference="d0"}, the above relation can therefore be employed to provide a limit on the relic contribution of Majoron as given by \[\Omega_\chi h^2< 1.28\times10^{-9}\left(\frac{10^{19}~\rm{sec}}{\tau_U} \right).\] The contribution turns out to be insignificant in making up the dark matter relic \(\Omega_\chi h^2 \simeq 0.120\pm 0.001\). Apart from the interactions with neutrinos mentioned in Eqs. [\[L3\]](#L3){reference-type="ref" reference="L3"} and [\[L4\]](#L4){reference-type="ref" reference="L4"}, Majorons also have interaction with the CP-even scalar \(\phi\). However, in view of negligible portal coupling \(\lambda_{H\Phi}\), this \(\phi\) field can not be present in the thermal bath of the early Universe consisting of the SM fields. Since \(\phi\) has no such important role to play, it is generally assumed that the \(\phi\) field is heavy enough (\(m_{\phi} \sim v_{\Phi}\), \(i.e.\) with \(\lambda_{\Phi} \sim \mathcal{O}(1)\)) compared to the rest of the masses involved in the set-up and hence decoupled. Other possibility of \(\chi\) production is from the process \(H^{\dagger} {H} \rightarrow \chi \chi\) which is again proportional to tiny Higgs portal coupling and suppressed by \(m_{\phi}\). This turns out to be a possibility of having heavy Majorons \(\sim\) TeV. Here, however, we plan to explore lighter Majorons as DM.
# The Proposal
As we have discussed above, the minimal Majoron set-up cannot accommodate the required relic density of FIMP like DM as Majoron. In order to make it a viable option, we propose to extend this set-up by introducing an explicit dimension-5 \(U(1)_L\) breaking term[^3] \[{-\mathcal{L}_{{\rm{d}}_5} = \frac{\alpha_i}{2\Lambda} \left[\Phi^2+(\Phi^*)^2\right]\overline{\mathcal{N}^c_{R_i}}{\mathcal{N}_{R_i}}+h.c.}, \label{d5}\] where \(\Lambda\) is a cut-off scale. Note that this being a higher order term, the \(U(1)_L\) symmetry is softly broken. For simplicity, we consider the coefficients \(\alpha_i\) to be \(\mathcal{O}(1)\) and omit them for further discussion. This is similar to the mass-term of the Majoron introduced in Eq. [\[soft-breaking\]](#soft-breaking){reference-type="ref" reference="soft-breaking"} where the soft-breaking nature was invoked by considering \(m\) to be small. Such a term[^4] also contributes to the mass of the RHNs, after the spontaneous breaking of \(U(1)_L\) symmetry, as given by \[M_{i} = v_\phi\left( \frac{f_i}{\sqrt{2}}+\frac{ v_\phi}{\Lambda}\right). \label{RHN-mass}\] Note that the inclusion of such a term carries the potential of generating a sizeable population of Majorons via the four-point interaction between Majorons and RHNs \[-\mathcal{L}_{int} = \frac{1}{2\Lambda} \chi^2 \overline{N^c_i}N_i, \label{effective-int}\] induced by Eq. [\[d5\]](#d5){reference-type="ref" reference="d5"} after the \(U(1)_L\) symmetry is spontaneously broken. This term contributes to the production of Majorons via \(NN \rightarrow \chi \chi\) annihilation, somewhat similar to the UV freeze-in scenario as we elaborate below. Although suppressed by the cut-off scale \(\Lambda\), this process turns out to be crucial in producing Majorons as it could remain in operation at a very early Universe (after the \(U(1)_L\) symmetry breaking), contrary to the usual Majoron production from RHN decay \(N \rightarrow \chi \nu\) being effective after the EWSB at temperature \(T_{EW}\).\
## Majoron Production via annihilation {#majoron-production-via-annihilation .unnumbered}
As proposed above, the introduction of the effective interaction between RHNs and Majorons as in Eq. [\[effective-int\]](#effective-int){reference-type="ref" reference="effective-int"} opens up the new production channel for Majorons via annihilation of RHNs (see Fig. [\[fynmann\]](#fynmann){reference-type="ref" reference="fynmann"}) once the \(U(1)_L\) symmetry is broken at a temperature[^5] \(T_L \sim v_{\phi}\). It is interesting to note that the RHNs receive their masses from the spontaneous breaking of the \(U(1)_L\) symmetry around this temperature only. Now, these RHNs may or may not be in thermal equilibrium. First, we consider the RHNs to be in equilibrium (as case\[A\]) and thereafter we investigate the situation when the abundance of RHNs at \(T_L\) to be negligible (as case \[B\]) to begin with which is expected to be increased thereafter gradually by the neutrino Yukawa interaction. Following the discussion before, we assume the Higgs portal coupling \(\lambda_{H\Phi}\) negligible and consider \(\phi\) being heavy to be decoupled field.\
To study the evolution of the yield of Majoron as DM, we use the Boltzmann equation written in terms of the temperature \((T)\) and the comoving number density of \(\chi\) as \(Y_\chi=n_\chi/\mathcal{S}\) (here \(\mathcal{S}\) is the entropy density). A general form of this equation (irrespective of whether \(N_i\) and/or \(\chi\) are in equilibrium) can be written as follows\
\[\begin{multlined} \mathcal{H}T\frac{dY_\chi}{dT}=-2{\mathcal{S}}\sum_{i}\big\langle\sigma v\big\rangle_{N_i N_i\to \chi\chi} Y_{N_i}^{eq^2}\left(\frac{Y_{N_i}^2}{Y_{N_i}^{eq^2}}-\frac{Y_\chi^2}{Y_\chi^{eq^2}} \right)\\ -\sum_{i,j}\big\langle\Gamma_{N_i\to \chi \nu_j}\big\rangle Y_{N_i}^{eq}\left(\frac{Y_{N_i}}{Y_{N_i}^{eq}}-\frac{Y_\chi Y_{\nu_j}}{Y_\chi^{eq} Y_{\nu_j}^{eq}} \right)~\theta(T_{EW}-T). \label{BEbasic} \end{multlined}\] Here, \(\mathcal{H}\) is the Hubble rate given by \(\mathcal{H} =1.66\sqrt{g^\rho_\star(T)} T^2/M_{pl}\) and \(\mathcal{S}=\frac{2\pi^2}{45}g^{\mathcal{S}}_\star(T)\), where \(g^{\mathcal{S}}_\star(T)\) and \(g^\rho_\star(T)\) are the effective degrees of freedoms of relativistic species at temperature \(T\) having \(M_{pl}=1.22\times 10^{19}\) GeV as the Planck mass. At high temperature, \(g^{\mathcal{S}}_\star(T)=g^\rho_\star(T)=106.75\) follows from SM particle content. The yield of a particle species \(a\) in thermal equilibrium (\(Y_{a}^{eq}\)) is given by, \[Y_{a}^{eq} (T)= \frac{45 g_a}{4\pi^4 g^\mathcal{S}_{\star}(T)}\left( \frac{m_a}{T}\right)^2 K_2\left(\frac{m_a}{T} \right),\] where \(g_a=1(2)\) (for scalar (fermion)) and \(m_a\) are the internal degrees of freedom and mass of the particle species respectively. \(K_2\) is the modified Bessel's function of the second kind. \(\big\langle\sigma v\big\rangle\) and \(\big\langle\Gamma\big\rangle\) are thermally averaged cross section and decay width respectively, the estimate of which are described below in the following section. The relic density of \(\chi\) is obtained after substituting the freeze-in abundance \(Y_\chi(T_0)\) in, \[\Omega_\chi h^2=2.755\times10^8\left(\frac{m_\chi}{\rm{GeV}} \right)Y_\chi(T_0),\] where \(T_0\) is the present temperature.
# Dark Matter phenomenology
We now study the role of the higher dimensional interaction of Eq. [\[d5\]](#d5){reference-type="ref" reference="d5"} in obtaining the relic density of the Majoron field \(\chi\) which is expected to play the role of freeze-in DM. From the discussion of the previous section, we understand that this operator contributes to the production of Majorons from the annihilations of RHNs. The matrix element squared for the annihilation process \(N_{i}N_{i}\to \chi\chi\) is given by, \[|\overline{M}|^2_{N_{i}N_{i}\to \chi\chi}= \frac{1 }{\Lambda^2}\left(s-4M_i^2 \right), \label{ampl}\] leading to the cross section \[\sigma_{N_{i}N_{i}\to \chi\chi}=\frac{1}{16\pi s}\sqrt{\frac{s-4m_{\chi}^2}{s-4M_{{i}}^2}} |\overline{M}|^2_{N_{i}N_{i}\to \chi\chi}, \label{cross-sec}\] where \(\sqrt{s}\) is the centre of mass energy of the process at a temperature \(T\). We consider \(U(1)_L\) symmetry to be broken below the reheating temperature, \(T_R\) indicating \(T_L < T_R\). Since the origin of the Majoron field is intertwined with the breaking of this global symmetry, the initial temperature for studying the \(\chi\) abundance through the Boltzmann equation(s) (via Eq. [\[BEbasic\]](#BEbasic){reference-type="ref" reference="BEbasic"}) can be considered as \(T_L \sim v_{\phi}\) instead of the reheating temperature as in UV freeze-in scenario. In this way, we do not need to invoke a parameter outside the framework (such as \(T_R\)) in the parameter space scan. However, without any prior knowledge on \(T_R\) and keeping in mind that RHNs get their masses at \(T_L\), RHNs may or may not be in equilibrium at this temperature. With the above points in mind, below we proceed for the evaluation of the freeze-in relic density of Majoron as DM corresponding to two cases: \[A\] RHNs are in thermal equilibrium and \[B\] RHNs are not in thermal equilibrium at \(T_L\) and find out the relevant parameter space. The initial abundance of the \(\chi\) particle is considered to be zero at this temperature \(T_L\) in both cases. As we have already seen in section [2](#section2){reference-type="ref" reference="section2"} that the contribution \(N_i\to\chi\nu\) channel to the \(\chi\) abundance is almost negligible, we proceed further without it.
## RHNs: in thermal bath at \(T_L\)
With the consideration that the RHNs are already in thermal equilibrium and abundance of \(\chi\) remains vanishing at \(T_L\), the Majorons can be produced from the annihilation via dimension-5 operator whose yield can be estimated by the following Boltzmann equation for \(\chi\) \[\begin{multlined} \frac{dY_\chi}{dT} \simeq -\frac{2\mathcal{S}}{\mathcal{H}T} \sum_{i} \big\langle\sigma v\big\rangle_{N_iN_i\to\chi\chi} Y_{N_{i}}^{eq^2} \left[ 1-\frac{Y_\chi^2}{Y_{\chi}^{eq^2}}\right], \end{multlined} \label{BE1}\]\
which is a simplified form of Eq. [\[BEbasic\]](#BEbasic){reference-type="ref" reference="BEbasic"}. First term in \(r.h.s\) of Eq. [\[BE1\]](#BE1){reference-type="ref" reference="BE1"} refers to the production of \(\chi\) due to the annihilation of RHNs where the factor 2 in front arises due to production of two Majorons in the final state. The second term in the \(r.h.s\) related to the back reaction \(\chi\chi\to N_i N_i ~(\textrm{and}~\chi\nu\to N_i)\) is not important as \(\chi\) starts with a negligible abundance in general. However, for light Majoron (of \(\sim\) keV mass), the \(\chi\) abundance near freeze-in temperature can be comparable (though remains smaller) to its equilibrium abundance where this term would be significant. On the other hand, the equilibrium abundance of RHNs is given by \[Y_{N_{i}}^{eq} (T)=\frac{45 g_{i}}{4\pi^4 g^\mathcal{S}_{\star}(T)}\left( \frac{M_i}{T}\right)^2 K_2\left(\frac{M_i}{T} \right),\]\
with \(g_i=2\). The thermally averaged cross-section appearing in Eq. [\[BE1\]](#BE1){reference-type="ref" reference="BE1"}, can be expressed as \[\begin{multlined} \big\langle\sigma v\big\rangle_{N_iN_i\to\chi\chi}= \frac{1}{8M_i^4TK_2^2(M_i/T)}\times\\ \int_{4M_i^2}^{\infty}\sigma_{N_iN_i\to\chi\chi} (s-4M_i^2)\sqrt{s}K_1\left( { \sqrt{s}}/{T} \right) ds. \end{multlined} \label{sigmav}\] By integrating Eq. [\[BE1\]](#BE1){reference-type="ref" reference="BE1"} from \(T_L \sim v_{\phi}\) (from where \(\chi\) production takes place) to \(T_0\) (today), we can obtain the yield of \(\chi\). Note that the parameters involved in analyzing the Majoron or \(\chi\) production are: \(f_1, f_2\) (related to RHN masses \(M_1, M_2\)), \(v_\phi, m_{\chi}\) and the cut-off scale \(\Lambda\). For simplicity, we consider the couplings (real) \(f_1, f_2\) to be same (\(f_1 = f_2 = f\)) and fix \(f\) at some natural value, \(f = 0.01\) so that the degenerate RHN masses \(M_i = M\) always remain below \(v_\phi\). Thereafter, we employ the Boltzmann equation (Eq. [\[BE1\]](#BE1){reference-type="ref" reference="BE1"}) to perform a scan over the parameter space so as to satisfy the DM relic density by the \(\chi\) field. In doing so, we implicitly consider \(\Lambda\) being the cut-off scale is bigger than \(v_{\phi}\), however, bounded by the largest scale \(M_{Pl}\). Furthermore, the stability criteria via Eq. [\[stability\]](#stability){reference-type="ref" reference="stability"} must be satisfied. We also incorporate an upper limit on \(v_{\phi}\) as the maximum value for the reheating temperature is generally considered to be \(T^{\rm{max}}_R\sim10^{15}\) GeV (\(v_{\phi} < T_R\) as stated earlier). Our findings are displayed in Fig. [\[PS1-th\]](#PS1-th){reference-type="ref" reference="PS1-th"} by the coloured region corresponding to the correct relic satisfaction in the \(m_{\chi}-v_{\phi}\) plane. The colours in this relic satisfied \(m_{\chi}-v_{\phi}\) space are indicative of the corresponding \(\Lambda\) value in the right-side colour bar. The gradient of the colour (from blue to dark red side) is increased with the increase in \(\Lambda\) which saturates at \(M_{Pl}\). The lower boundary of the allowed parameter space follows from the stability constraint while the top-left disallowed region (white patch) is limited by a specific choice \(v_{\phi}/{\Lambda} < f/\sqrt{2}\), so that the dimension-5 operator's contribution to the RHN mass (see Eq. [\[RHN-mass\]](#RHN-mass){reference-type="ref" reference="RHN-mass"}) remains sub-dominant. On the other hand, the top-right disallowed region corresponds to \(\Lambda > M_{Pl}\). The leftmost region of the allowed parameter space is bounded by the model-independent limit on the Majoron mass \(m_{\chi} < \mathcal{O}(0.1)\) keV, signifying that the Majorons can never be able to reach thermal equilibrium. We find for Majoron as light dark matter in the range 0.1 keV-1 GeV, the Lepton number breaking scale \(v_{\phi}\) and \(\Lambda\) fall in the following range, \[\begin{aligned}
m_{\chi}: 0.4~ {\rm{keV}}-1~ {\rm{GeV}};\nonumber \\ v_\phi: 10^6~ {\rm{GeV}}-10^{15} ~{\rm{GeV}}; \nonumber \\ \Lambda: 10^{10} ~{\rm{GeV}}-10^{18}~ {\rm{GeV}}. \nonumber \end{aligned}\] In order to demonstrate the freeze-in of the Majoron, we include Fig. [\[BP1-th\]](#BP1-th){reference-type="ref" reference="BP1-th"} where the abundance of the \(\chi\) field is presented as a function of temperature \(T\). This corresponds to a specific benchmark set of points (BP) from the parameter space as included in Table [1](#BP-table){reference-type="ref" reference="BP-table"} (indicated by a dark dot on the allowed parameter space of Fig. [\[PS1-th\]](#PS1-th){reference-type="ref" reference="PS1-th"}). Note that as the RHNs are in thermal equilibrium from the beginning (indicated by the blue line), the maximum production of \(\chi\) (in the orange line) from the RHN annihilation takes place near the temperature \(T_L\) itself, where the \(U(1)_L\) is broken. Such a behaviour is similar to the usual UV freeze-in scenario. The horizontal grid line corresponds to the yield of \(\chi\) producing the correct relic abundance[^6] for \(m_{\chi} =10\) keV. Note that this result remains unchanged even if we consider non-degenerate RHNs. Since both the RHNs (even if \(M_1 \neq M_2\)) are considered to be in thermal equilibrium at \(T_L\), it does not affect the production of \(\chi\) from \(N_i\) annihilation which happens to be around \(T_L\). Next, we proceed to investigate the situation when RHNs are not in the thermal bath.
We incorporate a set of benchmark parameters in the second row of Table [2](#BP-table2){reference-type="ref" reference="BP-table2"} to mark the dependence on the other parameter \(\theta_R\) (which in general can be complex) through \(y^{\nu}\) matrix. As in this case, initially, the abundance of RHNs is negligible and it enters equilibrium via the neutrino Yukawa interaction, \(\theta_R\) plays important role in determining the strength of this interaction and in a way affects the production of Majoron as well from RHN annihilation. Here, we consider a complex[^9] \(\theta_R = 2.5 + 1.463i\). The corresponding evolution of \(N_{1(2)}\) and Majoron abundances are shown in Fig. [\[BP1-B3\]](#BP1-B3){reference-type="ref" reference="BP1-B3"} by blue dotted (solid) and orange lines respectively. To exhibit the effect of \(\theta_R\) on the abundance of \(\chi\), we now choose a real \(\theta_R = 2.5\) while keeping all other parameters unchanged and exhibit the evolutions of RHNs and \(\chi\) (by red dotted (solid) and green lines respectively) in the same figure. As we can see, this makes the Majoron under-abundant. With this particular choice of \(\theta_R~(= 2.5+1.463i)\) and \(f_1 = 0.009, f_2 = 0.1\), we represent the result of another parameter space scan in Fig. [\[PS2\]](#PS2){reference-type="ref" reference="PS2"} in \(v_{\phi}-m_{\chi}\) plane. We note that the allowed parameter space becomes broadened in comparison with Fig. [\[PS1-th\]](#PS1-th){reference-type="ref" reference="PS1-th"} and [\[PS1\]](#PS1){reference-type="ref" reference="PS1"}. This is due to the inclusion of additional parameter in the set-up in the form of different \(f_1\) and \(f_2\) along with complex \(\theta_R\). The role of a complex angle \(\theta_R\) can be significant in the context of the model-dependent constraints on Majoron (such as constraint on \(\chi\to\gamma\gamma\) channel etc.) as we discuss it in the following section.
# Constraints
As already discussed, the presence of the decay channel \(\chi\to\nu \nu\) can make the Majoron unstable. Hence, in order to make the Majoron a viable dark matter, the condition \(\tau_\chi> \tau_{U}\) needs to be employed which translates into the following expression from the recent analysis using *Planck* with CMB lensing and BAO data as, \[\begin{aligned}
\Gamma(\chi\to\nu\nu)< 1.29\times 10^{-19}~ \textrm{s}^{-1}. \nonumber \end{aligned}\] We have already incorporated this limit on the parameter space. We now turn our attention to other constraints applicable to the parameter space obtained in our set-up originating from supernova, cosmology as well as some indirect search experiments. The same tree-level coupling of Majoron with light neutrino mass eigenstates (see Eq. [\[L3\]](#L3){reference-type="ref" reference="L3"}) may also experience a constraint from supernova (SN). The basic idea stems from the fact that inside the supernova core, Majorons can be produced through the process \(\nu \nu \to \chi\) (mainly \(\nu_e\) takes part inside the SN core). Provided the neutrino-Majoron coupling parametrized by \(g_{\chi \nu \nu} \equiv \sum_i m_{\nu_i}/{2v_{\phi}}\) is large, such Majorons would affect the neutrino signal emitted from a core collapsing SN which is otherwise consistent with the binding energy released during SN explosion as measured for SN1987A. Such a consideration leads to a disallowed range of the coupling \(g_{\chi \nu \nu}\): (\(10^{-7}-10^{-5}\)) for \(m_{\chi} <\) 10 MeV and (\(10^{-9}-10^{-6}\)) for \(m_{\chi} <\) 200 MeV, as evaluated in. However, our parameter space scan indicates that \(v_{\phi} \gtrsim \mathcal{O}(10^{7})\) GeV and accordingly \(g_{\chi \nu \nu} \lesssim 2.5\times 10^{-18}\) (where \(\sum_i m_{\nu_i} \simeq 0.05\) eV is inserted). Hence, it remains way below the limit from supernova constraint. Also, this Majoron decay into light neutrinos may produce observable monochromatic neutrino fluxes which can be analyzed in dedicated experiments such as Borexino (testable at \(4\) MeV \(<m_\chi<60~\)MeV), KamLAND (testable at \(16\) MeV \(<m_\chi<60~\)MeV), SK (within \(30\) MeV \(<m_\chi<200~\)MeV and IceCube (up to 10 TeV). The study in translates these constraints as the lower bound on the lepton number breaking scale \(v_{\phi}\) against \(m_{\chi}\) above MeV. We have taken their limits from and imposed upon the parameter space obtained in both cases A and B.1 as shown in Figs. [\[constraint-A\]](#constraint-A){reference-type="ref" reference="constraint-A"} and [\[constraint-B\]](#constraint-B){reference-type="ref" reference="constraint-B"}. As a result, it restricts the obtained parameter space further. Apart from the model independent decay channel \((\chi\to\nu\nu)\), there exists another interesting decay channel of Majoron: \(\chi\to \gamma\gamma\) which can be induced in two-loops. Following, a simplified expression for the associated decay width turns out to be, \[\Gamma\left(\chi\to\gamma\gamma \right) \simeq \frac{\alpha^2 ({\rm{tr}}{\mathcal{K}})^2}{4096\pi^7}\frac{m_\chi^3}{v^2}\left| \sum_{f}N_c^f T_3^f Q_f^2 g\left(\frac{m_\chi^2}{4m_f^2} \right)\right|^2, \label{gamma-line}\] where \(f\) denotes all the SM fermions having color multiplicity as \(N_c^f=3 (1)\) for quarks (leptons), \(T_3^f\) is the isospin, \(Q\) is the corresponding electric charge in units of \(e=\sqrt{4\pi \alpha}\). The loop function is given by \[g(x)=-\frac{1}{4x}\left(\textrm{log}|1-2x+2\sqrt{x(x-1)}| \right)^2, \nonumber\] having \(x ={m_\chi^2}/{4m_f^2}\). Here \({\mathcal{K}}\) is a model-dependent parameter given by \({\mathcal{K}}=m_D m_D ^\dagger/\left(vv_\phi\right)\), a dimensionless hermitian coupling matrix. Such \(\gamma\) emission lines can be probed by many experiments such as INTEGRAL \(\gamma\) observatory, COMPTEL/EGRET, Fermi-LAT. In addition, the diffuse X-ray background observed by HEAO was looking for the emission line over \(3-48\) keV range, while other X-ray telescopes like \(chandra\) and XMM cover the energy range \(0.3-12\) keV. Similarly, the \(\gamma\)-ray line emission of \(20\) keV-\(7\) MeV can be covered by INTEGRAL SPI observations. Energies above \(7\) MeV is covered by a combination of COMPTEL and EGRET \(\gamma\)-ray telescopes. Fermi \(\gamma\)-ray telescope recently searches the emission line within \(7-200\) GeV energy range. We first evaluate the \({\mathcal{K}}\)-matrix for each \(m_{\chi}\) from the allowed parameter space (using \(v_{\phi}\) and neutrino Yukawa coupling \(y^\nu\) via Eq. [\[CI\]](#CI){reference-type="ref" reference="CI"}) of Figs. [\[PS1\]](#PS1){reference-type="ref" reference="PS1"} and [\[PS2\]](#PS2){reference-type="ref" reference="PS2"}. After plugging this \(\mathcal{K}\) into Eq. [\[gamma-line\]](#gamma-line){reference-type="ref" reference="gamma-line"}, respective \(\Gamma (\chi \to \gamma\gamma)\) is plotted against \(E_{\gamma} = m_\chi/2\) in top (for case B.1) and bottom (case B.2) panels of Fig. [\[decay-gamma\]](#decay-gamma){reference-type="ref" reference="decay-gamma"} for the entire allowed parameter space obtained in our proposal where the limits from different X-ray and \(\gamma\)-ray observations are also shown. The differently coloured shaded patches on the upper side of each plot indicate the excluded regions for different experiments, the description of which is given in the inset of each figure. It can be seen that the Majoron mass below MeV can easily evade these bounds making the cosmological bound most effective in the sub-MeV mass regime. Whereas, relatively larger values of \(v_{\phi}\) appear to be somewhat constrained for \(m_{\chi} \sim 3.5-13\) MeV mass for case B.1. Alongside, Majoron masses beyond 100 MeV, in this case, fall within the sensitivities of the observations by COMPTEL and EGRET. For case B.2, it can be seen from Fig. [\[decay-gamma\]](#decay-gamma){reference-type="ref" reference="decay-gamma"} that the entire \(m_\chi\gtrsim 0.8\) MeV region turns out to be within the exclusion limits of \(X\)-ray and \(\gamma\)-ray emission line bounds for this particular choice of \(\theta_R\). However, this conclusion is not very robust as it can be altered for a different choice of \(\theta_R\), and hence model dependent. This stems from the fact that the decay width \(\Gamma (\chi \to \gamma\gamma)\) depends on the \(\mathcal{K}\) matrix which is governed by the choice of complex \(\theta_R\). In case of real \(\theta_R\) however, the \(\mathcal{K}\) remains independent of \(\theta_R\). For this reason, we do not incorporate a separate figure for Case A (with real \(\theta_R\)) which would result in a similar plot as the top panel of Fig. [\[decay-gamma\]](#decay-gamma){reference-type="ref" reference="decay-gamma"}.
# conclusion
In this work, we relooked into the original Majoron model, the existence of which is due to the spontaneous breaking of global lepton number symmetry. Its feeble coupling with the SM fields renders itself as a natural candidate for FIMP-like dark matter. Earlier studies showed that though it is plausible to have Majorons as FIMP-like dark matter, the relic satisfaction remains a challenge and Majoron mass can either be 2.7 MeV or heavy \(\sim\) TeV. In an effort to explore the possibility of having Majorons as a viable FIMP type DM candidate in a broader range (and below GeV) so as to make it interesting from the observational point of view, we include a higher order explicit lepton number breaking term in the Lagrangian involving RHNs and the spontaneous symmetry breaking scalar field. While such an interaction presents itself as a natural soft symmetry-breaking term (due to the suppression by the cut-off scale), it introduces new feeble interaction for Majorons, thereby introducing new production channel for Majorons via annihilation of the RHNs. Our analysis shows that an adequate region of parameter space consistent with Majoron as dark matter satisfying correct relic abundance in the keV-GeV mass range can be obtained. While this entire mass regime is important in the context of supernova cooling, the above-MeV region is also fascinating to be probed by monochromatic neutrino lines. Even though the bounds from such experiments turn out to be more stringent than the naively considered lifetime of the Universe, the parameter space undergoes a rich dark matter phenomenology and finally provides compelling evidences in the form of gamma-ray lines that can be probed by ongoing and future experiments indicating some tantalizing connection between dark matter and neutrino physics.
[^1]: Direct detection bound is automatically evaded due to the presence of such a small portal coupling, contrary to the freeze out case, involved.
[^2]: \(M_R\) being diagonal, \(N\) coincides with \(\mathcal{N}_R\).
[^3]: The usual dimension-5 operator contributing to light neutrino mass, \(\frac{c_{ij}}{\Lambda} L_iHL_jH\), may also be present. However, assuming \(c_{ij}\) to be small enough, we ignore this term for the analysis without loss of any generality.
[^4]: Disappearance of any cross generation term is ensured in the mass-diagonal basis of RHNs.
[^5]: We consider the reheating temperature after inflation (\(T_R\)) to be bigger than \(T_L\), though smaller than \(\Lambda\).
[^6]: Even with below 10 keV mass, these Majorons won't form small structures warm dark matter as they (Bosonic DM) are non-thermally produced in our case.
[^7]: Recall that RHNs become massive at \(T_L\) only.
[^8]: This particular choice of \(\theta_R\) with degenerate RHN mass corresponds to the same decay widths for both the RHNs, hence \(N_1\) and \(N_2\) reach equilibrium simultaneously. A general picture is demonstrated in the next subsection.
[^9]: Note that a complex \(\theta_R\) can be the source of CP violation in leptogenesis. This particular choice of \(\theta_R\) is related to correct baryon asymmetry generation via leptogenesis with \(M_{1(2)}\) considered here. | {'timestamp': '2022-12-19T02:10:07', 'yymm': '2212', 'arxiv_id': '2212.08404', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08404'} |
null | null |
# Introduction
The study of entanglement properties in many-body systems had a vast contribution to our understanding of quantum phases of matter. In particular, the structure of entanglement in pure ground states of lattice systems with local interactions is characterized by an area law. In one-dimensional systems, logarithmic violations of the area law signal the criticality of the state, with a prefactor that is universal and described by conformal field theory (CFT). Despite the enormous progress for pure states, the characterization of entanglement in mixed states remains a challenge. The core of the problem is finding an efficiently computable measure of entanglement, with a prospective candidate being the logarithmic negativity. In the quantum field theory framework, the entanglement negativity between two segments can be evaluated in ground and thermal states of a 1D chain. These predictions are nicely confirmed by numerical calculations in harmonic and spin chains. Results for higher dimensional lattices also exist, and the role of negativity in finite-temperature phase transitions has been addressed. The studies have also been extended to the out-of-equilibrium scenario, and the time evolution of entanglement negativity was investigated in quantum quenches. Interestingly, for free-particle chains it has been observed that there is an intimate proportionality relation between the negativity and the Rényi mutual information (RMI) with index \(\alpha=1/2\). Namely, to leading order one finds that the two quantities are related by a factor two, which has been substantiated in a number of other quench scenarios. In particular, the results of suggest that the relation should hold for rather generic unitary dynamics, provided that the initial state is pure. On the other hand, recent counterexamples were found in the context of non-unitary dissipative dynamics, where the leading order behaviour of negativity and RMI were found to be different. However, this cannot be a completely generic property of mixing dynamics, as the proportionality has been recovered in other non-unitary cases and even for random mixed states. Furthermore, the inequivalence of entanglement negativity and RMI has also been pointed out in quench dynamics of irrational CFTs. The above examples clearly demonstrate, that the relation between these two quantities is far from being understood. In this paper we shall provide an example for a mixed state of a free-fermion chain, where the relation between negativity and RMI breaks down. Namely, we consider a nonequilibrium steady state (NESS) that emerges under unitary dynamics, after two half-chains at different temperatures are joined together. In the very same setting, the standard mutual information was shown to violate the area law, and similar logarithmic violations have been found in the NESS of various other free-fermion related chains. In sharp contrast, in the analogous NESS of a harmonic chain both the negativity and the mutual information were found to satisfy the area law, which is the result dictated by free-boson CFT. Here we study the RMI with \(\alpha=1/2\) as well as the negativity between two adjacent or disjoint intervals in the fermionic NESS, and show that they both scale logarithmically in the segment size or cross ratio, respectively. However, the prefactors are described by different functions of the temperatures, and hence the two quantities provide inequivalent information about the NESS. Our results are obtained from the Rényi generalization of the negativity via an analytic continuation, using the techniques introduced in. The analytical results are in perfect agreement with our numerical calculations. Furthermore, we also provide numerical evidence that the negativity between adjacent intervals grows logarithmically in time during the dynamics, governed by the exact same prefactor found for the NESS. The rest of the manuscript is structured as follows. In Section [\[sec:model\]](#sec:model){reference-type="ref" reference="sec:model"} we introduce the model and describe the setup. The results for the Rényi mutual information as well as for the entanglement negativity are presented in Sec. [\[sec:rmi\]](#sec:rmi){reference-type="ref" reference="sec:rmi"} and [\[sec:neg\]](#sec:neg){reference-type="ref" reference="sec:neg"}, respectively. Our findings are discussed in Sec. [\[sec:disc\]](#sec:disc){reference-type="ref" reference="sec:disc"}, followed by two appendices with some technical details of the calculations.
# Model and setup[\[sec:model\]]{#sec:model label="sec:model"}
We consider free fermions on an infinite chain, described by the Hamiltonian with fermionic creation and annihilation operators satisfying the canonical anticommutation relations \(\{ c^\dag_m,c_n\}=\delta_{m,n}\). We are interested in the nonequilibrium dynamics of the hopping chain that emerges due to an inhomogeneous initial state, characterized by two different temperatures \(\beta_{l}\) and \(\beta_{r}\) on the left and right side The half-chain Hamiltonians \(H_l\) and \(H_r\) have the same form as [\[H\]](#H){reference-type="eqref" reference="H"}, but with sums running over sites \(-\infty\) to \(-1\) and \(1\) to \(\infty\), respectively. The state \(\rho_0\) is thus a tensor product of two Gibbs states at different temperatures, which for \(t>0\) evolves under the unitary dynamics The main focus of our studies is the nonequilibrium steady state \(\rho_\infty\), and its entanglement properties. For the initial state [\[rho0\]](#rho0){reference-type="eqref" reference="rho0"} it has been shown that the NESS is locally well defined by the requirement for any observable \(\mathcal{O}\) supported on a finite set of sites. Furthermore, since \(\rho_\infty\) is a Gaussian state, it is fully characterized by its correlation matrix where the occupation function is given by In other words, the right-moving fermionic modes are thermalized at the temperature of the left-hand side bath and vice-versa. In fact, the NESS can be understood in a simple hydrodynamic picture, where the non-interacting modes propagate ballistically and carry the information of their initial occupation to large distances. The conventional way of characterizing entanglement is to consider a bipartition of the system into a subsystem \(A\) and its environment \(B\), and consider the reduced density matrix \(\rho_A = \mathrm{Tr}_B (\rho_\infty)\). However, since the NESS is a mixed state, the von Neumann entropy is not a proper measure of entanglement. Instead, one could consider a tripartite scenario with subsystems \(A_1\), \(A_2\) and \(B\), and calculate the mutual information which is a measure of the total correlations between \(A_1\) and \(A_2\). For two adjacent intervals, \(\mathcal{I}(A_1:A_2)\) satisfies an area law in thermal equilibrium for arbitrary local Hamiltonians. However, for the NESS at hand it has been shown in, that the mutual information violates the area law and scales logarithmically in the subsystem size. This behaviour is due to the jump singularity in the occupation function [\[nq\]](#nq){reference-type="eqref" reference="nq"} at \(q=0\), between while a second jump between the values \(1-a\) and \(1-b\) occurs at \(q=\pm \pi\). The mutual information quantifies only the total (quantum + classical) correlations between two subsystems in the NESS. In the following we shall extend the calculations to obtain the logarithmic negativity, which is a proper measure of entanglement. As a first step towards this goal, we study the Rényi mutual information, which was shown to have a very close relation to the negativity in the context of quantum quenches. The techniques introduced in the next section will be directly applicable to the calculation of the negativity.
# Rényi mutual information[\[sec:rmi\]]{#sec:rmi label="sec:rmi"}
The Rényi mutual information is defined as via the Rényi entropy It is completely analogous to the standard mutual information [\[mi\]](#mi){reference-type="eqref" reference="mi"}, which is reobtained in the limit \(\alpha \to 1\). For generic \(\alpha\), however, the RMI is not even a proper measure of correlations. Indeed, for the NESS of a transverse Ising chain emerging from the initial state [\[rho0\]](#rho0){reference-type="eqref" reference="rho0"} it has been shown, that the RMI may become negative for indices \(\alpha > 2\). On the other hand, for \(0<\alpha<2\) the subadditivity of the Rényi entropy was proven for arbitrary fermionic Gaussian states, ensuring the positivity of the RMI. Our main goal is to derive the asymptotics of the RMI with index \(\alpha=1/2\) between two (adjacent or disjoint) intervals of size \(\ell \gg 1\). The strategy is to first express the RMI with integer index \(n>1\) and then perform an analytic continuation. Indeed, for integer values of the Rényi index, the entropy can be calculated as where \(Z_{A,k}\) is the generating function of the full counting statistics (FCS) of the particle number \(\hat N_{A}\) in \(A\) Here \(\openone_A\) is the identity matrix on subsystem \(A\), and the phases must be evaluated at the discrete values Note that the index \(k\) takes integer/half-integer values for \(n\) odd/even, and \(C_A\) is the reduced correlation matrix with elements [\[cmn\]](#cmn){reference-type="eqref" reference="cmn"} restricted to \(m,n \in A\). While originally obtained via the replica trick in a field theoretic setting, Eq. [\[sn\]](#sn){reference-type="eqref" reference="sn"} can also be derived immediately using the determinant formula for the Rényi entropy of Gaussian states, see Appendix [\[app:det\]](#app:det){reference-type="ref" reference="app:det"}. In order to obtain the RMI in [\[rmi\]](#rmi){reference-type="eqref" reference="rmi"}, we need to evaluate the entropies for the intervals \(A_1\) and \(A_2\), which involves the calculation of determinants [\[Zk\]](#Zk){reference-type="eqref" reference="Zk"} of a Toeplitz matrix with a symbol [\[nq\]](#nq){reference-type="eqref" reference="nq"} that has a jump singularity. This can be performed using standard Fisher-Hartwig techniques as in. However, for disjoint intervals this is not any more true for the subsystem \(A_1 \cup A_2\). To overcome this problem, one can apply a trick that allows us to deal with Toeplitz determinants, at the expense of modifying the occupation function [\[nq\]](#nq){reference-type="eqref" reference="nq"}. Indeed, the main argument is that the logarithmic contribution of the RMI we are interested in should depend only on the parameters \(n_{q\to 0^-}=a\) and \(n_{q\to 0^+}=b\) defined in [\[ab\]](#ab){reference-type="eqref" reference="ab"}. Therefore, we introduce the piecewise constant occupation function and the corresponding correlation matrix Note that we also regularize the problem by considering sites on a ring of finite size \(L\), such that the momenta \(q\) are integer multiples of \(2\pi/L\) and the corresponding eigenstates are \(\varphi_q(j) \sim \mathrm{e}^{iqj}\) up to normalization. Clearly, the correlation matrix [\[Cpmn\]](#Cpmn){reference-type="eqref" reference="Cpmn"} will not reproduce the correct extensive part of the entropy in the NESS, which, however, anyway cancels out in the RMI. We thus proceed with the calculation of the determinant in [\[Zk\]](#Zk){reference-type="eqref" reference="Zk"} by exchanging \(C_A \to C'_A\). Let us first rewrite where Then \(C^0_A\) is nothing else but the reduced correlation matrix of a Fermi-sea ground state with all the negative momenta occupied. This is known to have the same eigenvalues as the overlap matrix where the sum runs over the subsystem \(A\) and the momentum indices are restricted to the Fermi sea, \(p,q \le 0\). The FCS can thus be rewritten as where and \(\openone\) is the identity over the Fermi sea in momentum space. It is easy to see that the matrix \(M\) with elements defined in [\[M\]](#M){reference-type="eqref" reference="M"} has a Toeplitz structure for *arbitrary* subsystems \(A\). For simplicity, we consider two disjoint intervals of equal length \(A_1=\left[1,\ell \right]\) and \(A_2=\left[d+\ell+1,d+2\ell \right]\), by keeping the ratios \(d/L\) and \(\ell/L\) fixed. Introducing \(\theta=2\pi j/L\) and taking the thermodynamic limit \(L \to \infty\), one obtains the integral with a piecewise constant symbol \(\phi_A(\theta)\) that depends on the parameters \(a,b\) and the phase \(\lambda_k\). In particular, for a single interval one has where and the two jump singularities are located at On the other hand, for the composite subsystem \(A_1 \cup A_2\) we have the symbol with additional jump locations at One can now apply the Fisher-Hartwig theorem to evaluate the Toeplitz determinant in [\[Zk2\]](#Zk2){reference-type="eqref" reference="Zk2"}. Indeed, let us rewrite the symbol in the form \(\phi_A(\theta)=\prod_s g_s(\theta)\) where and \(s\) labels the FH singularities. The logarithm of the \(L/2 \times L/2\) Toeplitz determinant then reads \[\begin{aligned}
&\ln \det (\openone+M) = \frac{L}{2} \sum_{s} i \, \theta_s \beta_s + \sum_s \mathcal{C}_s \nonumber \\ &+ 2\sum_{s_1<s_2} \beta_{s_1}\beta_{s_2} \ln \left[L \sin \Big(\frac{\theta_{s_2}-\theta_{s_1}}{2}\Big)\right], \label{FH} \end{aligned}\] where the constant \(\mathcal{C}_s\) depends on the jump only via \(\beta_s\), and we assumed \(\sum_s \beta_s=0\). Let us first consider the case of adjacent intervals, where both symbols have two FH singularities with \(s=0,1\). In particular, for \(\phi_{A_1}(\theta)\) the jumps at \(\theta_0=0\) and \(\theta_1\) are characterized by \(\beta_1=-\beta_0=\beta(\lambda_k)\), where Applying the FH formula [\[FH\]](#FH){reference-type="eqref" reference="FH"} and adding the contribution of the second determinant in [\[Zk2\]](#Zk2){reference-type="eqref" reference="Zk2"} then gives \[\begin{aligned}
\ln Z_{A_1,k} &= \frac{\ell}{2} \ln \left[(1-a+\mathrm{e}^{i\lambda_k} a)(1-b+\mathrm{e}^{i\lambda_k}b)\right] \nonumber \\ &-2\beta^2(\lambda_k) \ln \left [L \sin ( \pi \ell/L)\right] + 2 \, \mathcal{C}(\lambda_k), \end{aligned}\] where \(\mathcal{C}(\lambda_k)\) is a subleading constant. The result for \(A_1 \cup A_2\) simply follows by substituting \(\ell \to 2\ell\). It is easy to see that the extensive term cancels out in the RMI, and after carrying out the sum over \(k\) as in [\[sn\]](#sn){reference-type="eqref" reference="sn"} and taking the limit \(L \to \infty\) one arrives at where the prefactor reads Before testing the validity of [\[sign\]](#sign){reference-type="eqref" reference="sign"}, let us first observe some of its properties. First of all, due to the definition of the \(\lambda_k\) in \(\eqref{lamk}\), the expression is real as it should. Moreover, it is easy to check that [\[sign\]](#sign){reference-type="eqref" reference="sign"} is invariant under the simultaneous exchange of \(a \to 1-a\) and \(b \to 1-b\), which follows from the symmetry of the Rényi entropy under a particle-hole transformation. This property is necessary to account for the second jump at \(q\to\pm\pi\) of the actual occupation \(n_q\) in [\[nq\]](#nq){reference-type="eqref" reference="nq"}. The actual scaling of the RMI [\[rmi\]](#rmi){reference-type="eqref" reference="rmi"} in the NESS can then be obtained by calculating using the matrix elements in [\[cmn\]](#cmn){reference-type="eqref" reference="cmn"} for increasing sizes \(\ell\) of the intervals \(A_{1,2}\). We fixed \(\beta_l=0\) and calculated the scaling of the RMI for several values of \(n\) by varying the inverse temperature \(\beta_r\). The fitted values of the logarithmic prefactor are shown in Fig. [\[fig:sign\]](#fig:sign){reference-type="ref" reference="fig:sign"}, together with the analytical prediction [\[sign\]](#sign){reference-type="eqref" reference="sign"}, with an excellent agreement. It is also straightforward to handle the case of disjoint intervals, where the symbol [\[phi12\]](#phi12){reference-type="eqref" reference="phi12"} has four jumps, characterized by \(\beta_1=\beta_3=\beta(\lambda_k)\) and \(\beta_0=\beta_2=-\beta(\lambda_k)\). One then has \[\begin{aligned}
&\ln Z'_{A_1 \cup A_2,k} = \ell \ln \left[(1-a+\mathrm{e}^{i\lambda_k} a)(1-b+\mathrm{e}^{i\lambda_k}b)\right] \nonumber \\ &-2\beta^2(\lambda_k) \ln \left[\frac{f^2(\ell) f(2\ell+d)f(d)}{f^2(\ell+d)}\right] + 4 \, \mathcal{C}(\lambda_k) \,, \label{lnZ12} \end{aligned}\] where we introduced Thus in the limit \(L\to\infty\) the RMI becomes where the parameter \(\eta\) is nothing but the cross ratio of the disjoint intervals Note that this is exactly the result conjectured in. The result can be checked against the numerical data obtained in the NESS, by fixing the distance \(d\) of the intervals and varying \(\ell\). The RMI with \(n=2\) is plotted against \(\eta\) in Fig. [\[fig:rmidi\]](#fig:rmidi){reference-type="ref" reference="fig:rmidi"} for \(\beta_l=0\) and various \(\beta_r\), while the lines show the FH result in Eq. [\[Indi\]](#Indi){reference-type="eqref" reference="Indi"}. The agreement is good, although some finite-size corrections can be seen for small values of \(d\).
## Analytic continuation[\[sec:rmiac\]]{#sec:rmiac label="sec:rmiac"}
Our next goal is to perform an analytic continuation of the RMI to arbitrary Rényi index \(\alpha\). We follow the route that has been applied in to construct an integral representation of the sum [\[sign\]](#sign){reference-type="eqref" reference="sign"}. Let us first note, that the discrete values [\[lamk\]](#lamk){reference-type="eqref" reference="lamk"} appearing in the sum are actually related to the zeroes of the polynomial Indeed, it easy to see that and thus the summand of [\[sign\]](#sign){reference-type="eqref" reference="sign"} reads Analogously to [\[betalam\]](#betalam){reference-type="eqref" reference="betalam"}, we can define the function and by the residue theorem we thus have where the integration contour \(\Gamma\) encircles the roots \(z_k\). To evaluate the contour integral above, one should note that the function \(\beta(z)\) has a branch cut along \(x\in\left[a,b\right]\), while it is analytic outside of that interval. The integration contour must avoid this branch cut, running infinitesimally close to it on both sides, whereas the contour can be closed on a circle of radius \(R \to \infty\), as depicted in Fig. [\[fig:cont\]](#fig:cont){reference-type="ref" reference="fig:cont"}. Since \(\beta(z) \to 0\) on the latter part of the contour, the only contribution to the integral comes from the branch cut. Indeed, using [\[betabc\]](#betabc){reference-type="eqref" reference="betabc"}, one needs only the term in \(\beta^2(z)\) that changes sign when crossing the branch cut. Furthermore, one can rewrite as the derivative of the Rényi entropy density This yields the integral representation of the prefactor The result is an analytic function of the index \(n\) and can thus be continued to arbitrary non-integer values. It is instructive to check how [\[signint\]](#signint){reference-type="eqref" reference="signint"} relates to the result in, obtained by a direct FH calculation with the NESS correlation matrix [\[cmn\]](#cmn){reference-type="eqref" reference="cmn"}, using the method of. In fact, by partial integration one obtains \[\begin{aligned}
\sigma_n &=\frac{1}{\pi^2} \int \limits_{a+\epsilon}^{b-\epsilon} \mathrm{d} x \, s_n(x) \frac{b-a}{(x-a)(b-x)} \nonumber \\ &+\frac{1}{\pi^2} \left[s_n(b)+s_n(a)\right] \ln \frac{\epsilon}{b-a} \,, \label{signpint} \end{aligned}\] which is exactly the expression found in for \(n=1\). Note that one has to introduce an infinitesimal \(\epsilon \to 0\) to avoid the logarithmic singularity of the integral, which is exactly canceled by the boundary term on the second line of [\[signpint\]](#signpint){reference-type="eqref" reference="signpint"}. The prefactor \(\sigma=\lim_{n\to1}\sigma_n\) can be evaluated in terms of the dilogarithm function \(\Li{x}\) as \[\begin{aligned}
\sigma = \frac{1}{\pi^2} \left[ a \, \Li{\frac{a-b}{a}} \right. + &(1-a) \Li{\frac{b-a}{1-a}} \nonumber \\ + b \, \Li{\frac{b-a}{b}} + & \left. (1-b) \Li{\frac{a-b}{1-b}} \right]. \end{aligned}\] Another case we are interested in is the RMI with index \(\alpha=1/2\), where the derivation of a closed form expression is more involved and can be found in appendix [\[app:sig\]](#app:sig){reference-type="ref" reference="app:sig"}. The result can again be written in terms of dilogarithms, but with complex arguments as \[\begin{aligned}
\sigma_{1/2} &= \frac{2}{\pi^2} \mathrm{Re} \left[\frac{1}{2}\ln^2 \left(\frac{1+i \, x_a}{1+i \, x_b}\right) \right. \nonumber \\ &\left. +\Li{\frac{1+ i \, x^{\phantom{-1}}_a}{1+i \, x^{-1}_a}}+\Li{\frac{1+ i \, x^{\phantom{-1}}_b}{1+i \, x^{-1}_b}} \right. \nonumber \\ &\left.-\Li{\frac{1+ i \, x^{\phantom{-1}}_b}{1+i \, x^{-1}_a}}-\Li{\frac{1+ i \, x^{\phantom{-1}}_a}{1+i \, x^{-1}_b}} \right], \label{sig12} \end{aligned}\] where the parameters in the arguments are defined as The formula [\[sig12\]](#sig12){reference-type="eqref" reference="sig12"} is checked against the numerical data in Fig. [\[fig:sig12\]](#fig:sig12){reference-type="ref" reference="fig:sig12"} with an excellent agreement.
# Entanglement negativity[\[sec:neg\]]{#sec:neg label="sec:neg"}
We now move forward to our main goal of evaluating the logarithmic negativity between two intervals \(A_1\) and \(A_2\) in the NESS. Throughout this section we will use the notation \(A=A_1\cup A_2\), and \(\rho_A=\mathrm{Tr}_B (\rho_\infty)\) with \(B = \bar A\). The negativity is originally defined via the partial transpose of the reduced density matrix \(\rho_A\) which, however, in general leads to a non-Gaussian state. For free-fermion systems one could, instead, define the negativity via the partial time reversal operation. This leads to the Gaussian operators where \(T_2\) denotes the partial time reversal with respect to \(A_2\). They are characterized by their correlation matrices where we used a block notation and one has \((C_\pm)_{mn}=\mathrm{Tr\,}(O_\pm c_m^\dag c_n)\). The fermionic logarithmic negativity is then defined via the trace norm as and was shown to be a proper entanglement measure. In fact, it also provides a natural upper bound to the standard negativity, as defined via the partial transpose. To calculate \(\mathcal{E}\), we follow a similar strategy to the one applied for the RMI. Namely, we use the replica trick and define the quantity for arbitrary *even* integer \(n\). If we are able to calculate \(\mathcal{E}_n\) analytically, the logarithmic negativity follows from the replica limit One thus needs the analytic continuation \(n\to1\) of a sequence \(\mathcal{E}_n\) on even integers. Analogously to the RMI in [\[sn\]](#sn){reference-type="eqref" reference="sn"}, this object can be decoupled into a sum where we defined the twisted partition function Here \(\hat N_{A_1}\) and \(\hat N_{A_2}\) are the particle-number operators of the respective subsystem. Note that both \(\lambda_k\) and the reversed phase \(\pi-\lambda_k\) are located on the upper half plane for \(k>0\), and take values as in [\[lamk\]](#lamk){reference-type="eqref" reference="lamk"} with half-integer \(k\). The partition function \(\tilde Z_{k}\) can be computed using trace formulas for the products of Gaussian operators as where \(\openone_{A_1}\), \(\openone_{A_2}\) and \(\openone_{A}\) are the identity matrices on the respective subsystem. Clearly, due to the appearance of the reversed phase, [\[detzt\]](#detzt){reference-type="eqref" reference="detzt"} is not the determinant of a Toeplitz matrix. We will show, however, that after exchanging \(C_A \to C'_A\), a dual Toeplitz matrix can again be constructed, which is similar to the overlap matrix. To this end, we first note that for a correlation matrix \(C^0_A\) of a Fermi sea state as in [\[C0mn\]](#C0mn){reference-type="eqref" reference="C0mn"}, one has the identity where \(D=\mathrm{diag}(d_j)\) is an arbitrary *diagonal* matrix and the modified overlap matrix is defined as In other words, since [\[trdc\]](#trdc){reference-type="eqref" reference="trdc"} is satisfied for arbitrary \(n\), the matrices \(DC^0_A\) and \(\tilde M\) have the same eigenvalues. The proof of the identity simply follows by writing out the left hand side of [\[trdc\]](#trdc){reference-type="eqref" reference="trdc"} as and reordering the sums. One can now replace \(C_A \to C'_A\) as defined in [\[CpA\]](#CpA){reference-type="eqref" reference="CpA"} and insert it into [\[detzt\]](#detzt){reference-type="eqref" reference="detzt"}, which leads to where the matrices are given by \[\begin{aligned}
&D = E^{-1}(a-b) (\mathrm{e}^{i\lambda_k \openone_{A_1}} \mathrm{e}^{i (\pi-\lambda_k) \openone_{A_2}}-\openone_A) \,, \\ &E = (1-b)\openone_A + b \, \mathrm{e}^{i\lambda_k \openone_{A_1}} \mathrm{e}^{i (\pi-\lambda_k) \openone_{A_2}}\,. \end{aligned}\] The matrix \(D\) is thus diagonal, with matrix elements given by Finally, we can apply [\[trdc\]](#trdc){reference-type="eqref" reference="trdc"} and replace \(DC^0_A\) by \(\tilde M\) in the first determinant of [\[detzt2\]](#detzt2){reference-type="eqref" reference="detzt2"}. In the limit \(L\to \infty\), this becomes a Toeplitz determinant of the matrix with the Fisher-Hartwig symbol where \(\Phi(\lambda)\) is given in [\[Phi\]](#Phi){reference-type="eqref" reference="Phi"}, while the jump locations \(\theta_1\) and \(\theta_2,\theta_3\) are defined in [\[theta1\]](#theta1){reference-type="eqref" reference="theta1"} and [\[theta23\]](#theta23){reference-type="eqref" reference="theta23"}, respectively. We are now ready to evaluate the determinant using the FH theorem [\[FH\]](#FH){reference-type="eqref" reference="FH"}. Let us first consider the case of adjacent intervals where \(\theta_2=\theta_1\) and \(\theta_3=2\theta_1\). One has thus three FH singularities described by the functions In turn, the Rényi negativity [\[rnegsum\]](#rnegsum){reference-type="eqref" reference="rnegsum"} is given by where the prefactor of the logarithmic contribution is Note that, in general, one has an extensive term which receives contributions also from the second determinant in [\[detzt2\]](#detzt2){reference-type="eqref" reference="detzt2"}, and its prefactor can be evaluated as Here we used the fact that \(\pi-\lambda_k\) has the same set of allowed values as \(\lambda_k\). Furthermore, taking twice the real part is equivalent to including the negative \(k\) values in the sum. Using the factorization in [\[pnz\]](#pnz){reference-type="eqref" reference="pnz"}, one has then and hence the extensive prefactor vanishes in the limit \(n \to 1\). On the other hand, the logarithmic prefactor can be obtained using [\[betaneg\]](#betaneg){reference-type="eqref" reference="betaneg"} as Finally, let us consider the case of disjoint intervals. The symbol [\[phitk\]](#phitk){reference-type="eqref" reference="phitk"} has then four FH singularities, characterized by Similarly to [\[lnZ12\]](#lnZ12){reference-type="eqref" reference="lnZ12"}, the different pairings of the \(\beta_s(\lambda_k)\) are now multiplied by different logarithmic factors, and the prefactor [\[signegn\]](#signegn){reference-type="eqref" reference="signegn"} splits into two parts. In particular, the first sum including the \(\ln^2[\Phi(\lambda_k)]\) is multiplied by \(\ln \ell\), whereas the second mixed term is multiplied by \(-\ln \eta\), with the cross-ratio defined in [\[cr\]](#cr){reference-type="eqref" reference="cr"}.
## Analytic continuation
The final step to obtain the entanglement negativity is to carry out the analytic continuation [\[lnlim\]](#lnlim){reference-type="eqref" reference="lnlim"}. This can be performed in a very similar fashion as for the Rényi mutual information in Sec. [\[sec:rmiac\]](#sec:rmiac){reference-type="ref" reference="sec:rmiac"}. Namely, we shall find an integral representation of the sum [\[signegn\]](#signegn){reference-type="eqref" reference="signegn"}, which is analytic in \(n\) and thus the limit \(n \to 1\) can be carried out. First of all note that, using [\[Phi\]](#Phi){reference-type="eqref" reference="Phi"}, the first part of the sum is actually related to the RMI prefactor [\[sign\]](#sign){reference-type="eqref" reference="sign"} as and thus will vanish in the limit \(n \to 1\). We shall thus focus only on the mixed term in [\[signegn\]](#signegn){reference-type="eqref" reference="signegn"}, where both \(\lambda_k\) and \(\pi-\lambda_k\) appear. The main idea that was already applied in is to consider the complex roots \(\tilde z_k\) of the polynomial which can be expressed via the positive phases \(\lambda_k\) as In the integral representation the roots should appear as poles. It should be noted, however, that there is a one-to-one correspondence between the phases \(\lambda_k\) and the roots \(\tilde z_k\) only for \(n/2\) even. Indeed, for \(n/2\) odd the phase \(\lambda_k=\pi/2\) appears, which corresponds to \(\tilde z_k^{-1}=0\), i.e. \(\tilde p_n(z)\) actually has only \(n/2-1\) roots. The contribution of the missing pole thus has to be added in this case. For the moment we shall assume \(n/2\) to be even, and comment on the other case later on. Inverting the relation [\[ztk\]](#ztk){reference-type="eqref" reference="ztk"} requires some care. In order to have all the phases on the upper half plane, one needs The integral representation of the sum thus reads where the arguments are given by \[\begin{aligned}
&\Phi(z)= \frac{1-a+ i\sqrt{\frac{z}{z-1}} a}{1-b+ i\sqrt{\frac{z}{z-1}} b} \,, \\ &\bar\Phi(z)= \frac{1-b+ i\sqrt{\frac{z-1}{z}} b}{1-a+ i\sqrt{\frac{z-1}{z}} a} \,. \end{aligned}\] Note that, apart from the poles at \(\tilde z_k\), the integrand has a branch cut along \(\left[0,1\right]\) due to the square root in the argument of the logarithms. Thus the contour \(\tilde \Gamma\) has to be chosen in a similar fashion to Fig. [\[fig:cont\]](#fig:cont){reference-type="ref" reference="fig:cont"}, such that it goes around the branch cut and encircles all the poles, with its outer radius taken to be very large. The first contribution to the contour integral comes from the path running along the branch cut. Indeed, for \(z=x+i 0^\pm\) and \(x \in \left[0,1\right]\) one has and the integral reads where \[\begin{aligned}
&\Phi(x)= \frac{1-a+ \sqrt{\frac{x}{1-x}} a}{1-b+ \sqrt{\frac{x}{1-x}} b} \,, \\ &\bar\Phi(x)= \frac{1-b-\sqrt{\frac{1-x}{x}} b}{1-a-\sqrt{\frac{1-x}{x}} a} \,. \end{aligned}\] Note that a factor two appears since the contributions are equal on both sides of the branch cut, and by dropping the factor \(i\) one now needs the imaginary part of the integral. Since \(\Phi(x)>0\) for \(x\in[0,1]\), an imaginary part can only appear when \(\bar \Phi(x)<0\). This is the case on the interval \(x \in [\tilde a, \tilde b ]\) where One has then \(\mathrm{Im} \ln [\bar \Phi(x)]= \pi\) and one arrives at We also need the contribution over the large circle of the contour \(\tilde \Gamma\), parametrized as \(z=R \mathrm{e}^{i\varphi}\) with \(R\to\infty\), which turns out to be nonvanishing. Indeed, the square-roots in \(\Phi(z)\) and \(\bar\Phi(z)\) converge to one and thus the logarithms in the integrand yield a constant. On the other hand on has and thus the asymptotics depends on the parity of \(n/2\). In fact, this is simply due to the missing root for \(n/2\) odd, where the contribution from \(\lambda_k=\pi/2\) has to be added to the contour integral [\[cintneg\]](#cintneg){reference-type="eqref" reference="cintneg"}. Adding the two pieces removes the parity dependence, and with \(\mathrm{d} z/z=i \mathrm{d}\varphi\) one obtains the contribution on the circle which is valid for arbitrary \(n\). Collecting all the contributions, the integral representation of the sum [\[signegn\]](#signegn){reference-type="eqref" reference="signegn"} reads \[\begin{aligned}
\tilde \sigma_n &= \frac{1}{\pi^2} \int_{\tilde a}^{\tilde b} \mathrm{d} x \frac{\tilde p'_n(x)}{\tilde p_n(x)} \ln \left(\frac{1-a + a \, \sqrt{\frac{x}{1-x}}}{1-b + b \, \sqrt{\frac{x}{1-x}}}\right) \nonumber \\ &-\frac{n}{2\pi^2} \, \mathrm{Re} \ln^2 \left(\frac{1-a + a \, i}{1-b + b \, i} \right) +2(1-n)\sigma_n \,. \label{signegint} \end{aligned}\] The result can now be analytically continued to \(n\to 1\). In fact, as shown in Appendix [\[app:sig\]](#app:sig){reference-type="ref" reference="app:sig"}, the integral for \(\tilde \sigma = \lim_{n\to1} \tilde \sigma_n\) can be evaluated in a closed form. After a lengthy calculation one obtains \[\begin{aligned}
\tilde \sigma &= \frac{1}{\pi^2}\mathrm{Re} \left[ \frac{1}{2}\ln^2 \left(\frac{1+i \, \tilde x_a}{1+i \, \tilde x_b}\right) \right. \nonumber \\ & \left. + \Li{\frac{1+ i \, \tilde x^{\phantom{-1}}_b}{1-i \, \tilde x^{-1}_a}} + \Li{\frac{1+ i \, \tilde x^{\phantom{-1}}_a}{1-i \, \tilde x^{-1}_b}} \right. \nonumber \\ & \left.-\Li{\frac{1+ i \, \tilde x^{\phantom{-1}}_a}{1-i \, \tilde x^{-1}_a}}-\Li{\frac{1+ i \, \tilde x^{\phantom{-1}}_b}{1-i \, \tilde x^{-1}_b}} \right], \label{signeg} \end{aligned}\] where the parameters are defined as
## Numerical results
In the following we shall test the analytical results of the previous section by calculating the entanglement negativity \(\mathcal{E}=\ln \mathrm{Tr\,} \sqrt{O_+O_-}\) in the NESS. Using trace formulas for the product of Gaussian operators, this can be obtained directly as \[\begin{aligned}
\mathcal{E} = &\ln \det \left[ \sqrt{\frac{\openone+G_\times}{2}} + \sqrt{\frac{\openone-G_\times}{2}} \right] \nonumber \\ + \frac{1}{2} &\ln \det \left[\frac{\openone+G_+G_-}{2}\right], \label{lndet} \end{aligned}\] where \(\openone\equiv\openone_A\) and we introduced the matrix and \(G_\pm = 2C_\pm-\openone\) with the correlation matrices \(C_\pm\) defined in [\[cpm\]](#cpm){reference-type="eqref" reference="cpm"}. Thus one has to evaluate determinants of matrices that are completely determined via the reduced correlation matrix \(C_A\). We performed numerical calculations using [\[lndet\]](#lndet){reference-type="eqref" reference="lndet"} for various pairs of inverse temperatures. For adjacent intervals the analytical results of the previous section suggest the logarithmic scaling which we indeed observe in the numerics. We thus fitted our numerical data and compared the prefactor \(\tilde \sigma\) to the analytical formula [\[signeg\]](#signeg){reference-type="eqref" reference="signeg"}, with the result shown in Fig. [\[fig:signeg\]](#fig:signeg){reference-type="ref" reference="fig:signeg"}. The agreement is excellent, confirming that the logarithmic term in the negativity is determined only by the values \(a\) and \(b\) on both sides of the jump in the occupation function. We have also checked the disjoint case, where the analytic continuation suggests the result in terms of the cross ratio [\[cr\]](#cr){reference-type="eqref" reference="cr"}. Note that, analogously to the result [\[Indi\]](#Indi){reference-type="eqref" reference="Indi"} for the RMI, we expect that the constant term vanishes for disjoint intervals. This is indeed what we observe, as demonstrated in Fig. [\[fig:negdi\]](#fig:negdi){reference-type="ref" reference="fig:negdi"} for fixed \(\beta_l=0\) and various \(\beta_r\) values. Remarkably, compared to the case of the RMI in Fig. [\[fig:rmidi\]](#fig:rmidi){reference-type="ref" reference="fig:rmidi"}, the corrections to [\[lndi\]](#lndi){reference-type="eqref" reference="lndi"} remain very small even for short distances \(d=2\).
## Comparison of \(\tilde \sigma\) and \(\sigma_{1/2}\)
We are now ready to compare the RMI with \(\alpha=1/2\) to the entanglement negativity. In fact, previous results on quantum quenches in free-fermion systems have indicated, that the two quantities are related as \(\mathcal{E} = \frac{1}{2}\mathcal{I}_{1/2}\). In the present case, however, one can see that an analogous relation between the prefactors \(\tilde \sigma\) and \(\frac{1}{2}\sigma_{1/2}\) does not hold. Despite the very similar structure of their analytical expressions in [\[sig12\]](#sig12){reference-type="eqref" reference="sig12"} and [\[signeg\]](#signeg){reference-type="eqref" reference="signeg"}, the first obvious difference is in their respective variables [\[xab\]](#xab){reference-type="eqref" reference="xab"} and [\[txab\]](#txab){reference-type="eqref" reference="txab"}. It is instructive to rewrite these variables in terms of the inverse temperatures. Substituting [\[ab\]](#ab){reference-type="eqref" reference="ab"} one obtains and analogously for \(x_b\) and \(\tilde x_b\) by exchanging \(\beta_r \to \beta_l\). Interestingly, there is a simple factor two difference, which might suggest that the two prefactors become comparable only after an additional rescaling of the temperatures. This is indeed how the pairs \(\beta_l\) and \(\beta_r\) were chosen in Figs. [\[fig:sig12\]](#fig:sig12){reference-type="ref" reference="fig:sig12"} and [\[fig:signeg\]](#fig:signeg){reference-type="ref" reference="fig:signeg"}, resulting in a qualitatively similar behaviour of the prefactors. However, having a closer look at the vertical scale, one immediately sees that the values of \(\frac{1}{2}\sigma_{1/2}\) are much larger than those of \(\tilde \sigma\). The reason of this mismatch is that the arguments of the dilogarithm functions [\[sig12\]](#sig12){reference-type="eqref" reference="sig12"} and [\[signeg\]](#signeg){reference-type="eqref" reference="signeg"} are still not the same in terms of their natural variables. In order to compare the prefactors directly, we shall fix \(x_b=\tilde x_b=-1\) (setting the left reservoir to zero temperature \(\beta_l \to \infty\)), and plot them against the variable \(\tilde x_a\), as shown in Fig. [\[fig:negvsrmi\]](#fig:negvsrmi){reference-type="ref" reference="fig:negvsrmi"}. In the regime \(\tilde x_a<0\) allowed for positive temperatures, \(\frac{1}{2}\sigma_{1/2}\) is always larger than \(\tilde \sigma\), as observed already in Figs. [\[fig:sig12\]](#fig:sig12){reference-type="ref" reference="fig:sig12"} and [\[fig:signeg\]](#fig:signeg){reference-type="ref" reference="fig:signeg"}. We have, however, also plotted the prefactors for \(\tilde x_a>0\), corresponding to negative temperatures according to [\[xbeta\]](#xbeta){reference-type="eqref" reference="xbeta"}. This should be understood as an additional particle-hole transformation in the right bath, allowing for values \(0<a<1/2\) in [\[nqpc\]](#nqpc){reference-type="eqref" reference="nqpc"}. Interestingly, the two curves cross each other at \(\tilde x_a \approx 0.577\), and \(\tilde \sigma\) becomes larger than \(\frac{1}{2}\sigma_{1/2}\). The prefactors coincide again in the limit \(\tilde x_a=1\), where the occupation [\[nqpc\]](#nqpc){reference-type="eqref" reference="nqpc"} with \(a=0\) and \(b=1\) is just a shifted Fermi sea. Hence, in this limit one reproduces the ground-state result \(\tilde \sigma=\frac{1}{2}\sigma_{1/2}=\frac{1}{4}\), which follows also from CFT calculations.
## Time evolution of the negativity
Finally we investigate how the negativity builds up during the time evolution leading to the NESS. In particular, we are interested in \(\mathcal{E}(t)\) evaluated in the time-evolved state [\[rhot\]](#rhot){reference-type="eqref" reference="rhot"}. Since the density matrix \(\rho_t\) is Gaussian at any time \(t\), the calculation of the negativity can easily be generalized. In fact, one only needs the time-evolved correlation matrix where the propagator has matrix elements \(U_{mn}=i^{n-m}J_{n-m}(t)\) given by the Bessel functions. The initial correlation matrix \(C(0)=C_l(0) \oplus C_r(0)\) is a direct sum of thermal correlation matrices of the half-chains, corresponding to the factorized initial state [\[rho0\]](#rho0){reference-type="eqref" reference="rho0"}. Note that, although in [\[Ct\]](#Ct){reference-type="eqref" reference="Ct"} one needs the product of infinite matrices, the terms that contribute are heavily restricted by the light-cone behaviour of the propagator. With the expression [\[Ct\]](#Ct){reference-type="eqref" reference="Ct"} at hand, one can simply define the time-dependent matrices \(G_\pm(t)\) and \(G_\times(t)\) and use the corresponding formula [\[lndet\]](#lndet){reference-type="eqref" reference="lndet"} to evaluate \(\mathcal{E}(t)\). We are interested in the entanglement negativity of two intervals placed symmetrically around the junction. The results are shown on Fig. [\[fig:lnegt\]](#fig:lnegt){reference-type="ref" reference="fig:lnegt"} for some fixed initial temperatures, both for adjacent as well as disjoint intervals. Although the lack of translational invariance prevents us from using FH techniques, we expect that the result is similar to the one found for the mutual information. Namely for adjacent intervals one has a logarithmic growth for short times \(t < \ell\), with the exact same prefactor that governs the system-size scaling in the NESS. On the other hand, for times \(t>\ell\) after the front has crossed the boundaries of the intervals, one expects a relaxation towards the NESS value. This is indeed what we observe on the left of Fig. [\[fig:lnegt\]](#fig:lnegt){reference-type="ref" reference="fig:lnegt"}, where [\[lnegt\]](#lnegt){reference-type="eqref" reference="lnegt"} is shown by the red solid line, with the constant parameter determined by fitting. The NESS values are indicated by the horizontal lines for \(\ell=20,40\), and the data shows a clear convergence towards them. The disjoint case on the right of Fig. [\[fig:lnegt\]](#fig:lnegt){reference-type="ref" reference="fig:lnegt"} shows qualitatively similar features, with a delayed increase and decrease of \(\mathcal{E}(t)\) at times \(t=d/2\) and \(t=\ell+d/2\), with \(d/2\) being the distance from the junction. The precise description of \(\mathcal{E}(t)\) between these times is, however, beyond our reach.
# Discussion[\[sec:disc\]]{#sec:disc label="sec:disc"}
We have studied the scaling of entanglement in a current-carrying NESS of free fermions. Due to the Gaussianity of the state, both the RMI as well as the Rényi generalization of the entanglement negativity can be obtained via determinants involving the reduced correlation matrix. Moreover, the analytic continuation in the Rényi index can be worked out explicitly. One finds a logarithmic scaling for both quantities, which is a consequence of the jump singularity in the NESS occupation function, and the prefactors can be determined by FH techniques. In particular, we find different prefactors for the negativity as well as for the RMI with index \(\alpha=1/2\). To derive our analytical result we assumed that the only important details of the otherwise continuous occupation function are the values at the two jump locations. In our particular case, the second jump at momentum \(q=\pi\) is related to that at \(q=0\) by a particle-hole symmetry. This is equivalent to the change of variables \(x_{a,b} \to-x_{a,b}\) and \(\tilde x_{a,b} \to-\tilde x_{a,b}\), which is a trivial symmetry of the prefactors [\[sig12\]](#sig12){reference-type="eqref" reference="sig12"} and [\[signeg\]](#signeg){reference-type="eqref" reference="signeg"}. A more general situation would be to consider initial states with different chemical potentials in the two reservoirs, which would brake the particle-hole symmetry. We conjecture that the corresponding negativity prefactor would be a sum of the prefactors \(\tilde \sigma /2\) evaluated at the corresponding jump parameters, and similarly for the RMI. Another natural extension of the result would be to consider the NESS of an XY chain, which is described by block-Toeplitz matrices with similar jump singularities. The result for the RMI can then be obtained by using a generalized FH conjecture proposed in. It would be interesting to extend these techniques to the calculation of the negativity. Note, however, that for spin models one has the additional difficulty, that the partial transpose produces a non-Gaussian state, and the negativities in the spin and fermion representations are not equivalent. Nevertheless, the quench examples studied so far indicate that these two representations actually produce very similar results, at least to leading order. Whether this holds true for the mixed NESS at hand remains an open question. In a broader context one might wonder about the origin of the discrepancy between the \(\alpha=1/2\) RMI and the negativity. In the field theory context, this was found to occur only in irrational CFTs and was associated to a breakdown of the quasiparticle picture. This is clearly not the case here, as the free-fermion chain has perfectly well defined quasiparticles. However, there is actually no proper CFT description of the NESS studied here, as the logarithmic scaling of the negativity and RMI is clearly related to the nonvanishing dispersion at \(q \to 0\) and is thus a lattice effect. On the other hand, discrepancies between \(\mathcal{E}\) and \(\frac{1}{2}\mathcal{I}_{1/2}\) have also been observed in the case of dissipative free-fermion dynamics. The property this scenario has in common with ours is the global state being mixed. Whether in such cases the RMI is still a good measure of correlations and provides some relevant information about the quantum state remains to be further explored. | {'timestamp': '2022-12-19T02:12:44', 'yymm': '2212', 'arxiv_id': '2212.08499', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08499'} |
null | null |
# Introduction
Observations by the Large Area Telescope onboard the *Fermi Gamma-ray Space Telescope* have led to the detection of gamma-ray pulsations from nearly 300 pulsars[^1] (The Fermi-LAT Collaboration, 2023, in prep.). These fall into two main classes: canonical pulsars that are still rapidly spinning-down from their initial rotation periods; and millisecond pulsars (MSPs) that have been spun-up (or "recycled" @Bhattacharya+1991) to rates of hundreds of rotations per second by accretion from an orbiting star. Both pulsar classes have characteristic gamma-ray emission properties (curved spectra and low flux variability over time) that are distinct from those of other gamma-ray emitting objects. These properties can be used to identify promising pulsar-like gamma-ray sources, which can then be targeted with radio telescopes to detect pulsations and confirm their nature. The few-arcmin localisation regions of unassociated *Fermi*-LAT sources enable targeted, long and repeated observations of promising sources, and a higher detection efficiency than can be achieved when surveying a broad region of the sky. This method has proven highly successful at discovering new MSPs; more than a quarter of the 400 MSPs known in the Galactic field have been discovered in *Fermi*-LAT sources[^2] in a global effort coordinated by the *Fermi* Pulsar Search Consortium. These searches have been particularly effective in finding exotic but elusive "spider" binary pulsars whose long radio eclipses due to diffuse intra-binary material make them easily missed in single-pass untargeted surveys. Around two-thirds of the known spider binaries in the Galactic field were found by targeting *Fermi* sources. Many new MSPs remain to be found amongst the LAT sources: pulsars make up around 6% of the identified or associated[^3] sources in the recent 12-yr iteration of the *Fermi*-LAT Fourth Source Catalog, while over 2000 sources remain unassociated. Observed gamma-ray and radio fluxes from pulsars are not strongly correlated with one another, and so prospects remain high for detecting radio pulsars even in faint new gamma-ray sources that have only recently been detected thanks to the exposure that has accumulated during *Fermi*'s ongoing all-sky survey. This also means that new radio MSPs discovered within faint new *Fermi*-LAT sources can still be bright enough to be valuable astrophysical tools, and indeed several new MSPs found using this method have been added to pulsar timing array (PTA) projects aiming to detect gravitational waves. The potential importance of detecting new MSPs in *Fermi*-LAT sources is illustrated by the fact that several previous discoveries found in this manner now mark the extreme edges of the MSP population, and are therefore the best current probes for several fundamental astrophysics questions. These include: the fastest known Galactic MSP; the pulsar binary system with the shortest known orbital period; and a group of massive black-widow MSPs that probe the maximum neutron star mass. One crucial benefit of finding a new MSP within a gamma-ray source is that gamma-ray pulsations can often be detected and timed directly in the *Fermi*-LAT data. Initial timing solutions can be refined and extrapolated backwards to the start of the LAT data (which currently spans more than 14 years), providing long and precise rotational ephemerides without the need for lengthy radio timing campaigns. Illustrating the potential scientific benefits of this capability, gamma-ray timing of the recently discovered black-widow PSR J1555-2908 may have revealed a second, planetary mass object in a long period orbit around the inner binary system. The LAT data have also even been recently exploited to build a gamma-ray PTA, whose sensitivity to a stochastic gravitational wave background may reach that of current radio PTAs within a decade. While searches for pulsations in the *Fermi*-LAT data itself can reveal new MSPs, trials factors and computational costs limit these searches to the brighter pulsars. The detection of gamma-ray pulsations from MSPs in binary systems (which most are) is also impossible without prior orbital constraints. Radio surveys and initial timing therefore remain critical for expanding the population of Galactic MSPs. The efforts of searching for new MSPs in *Fermi*-LAT sources have recently been bolstered by new radio telescopes. These bring capabilities of observing in new parameter spaces or with greater sensitivity. In this paper, we present the first results from one such new survey of unassociated *Fermi*-LAT sources, using the MeerKAT radio telescope. The full MeerKAT array is around five times more sensitive than the Murriyang Parkes telescope , the next most sensitive radio telescope in the Southern Hemisphere. The Transients and Pulsars with MeerKAT (TRAPUM) project is a large survey project using MeerKAT to search for new pulsars. All TRAPUM observations target sky locations in which pulsars are particularly likely to lie: globular clusters; nearby galaxies; supernova remnants, pulsar wind nebulae and other TeV sources; and GeV gamma-ray sources. A separate dedicated L-band survey of the Galactic plane (MMGPS-L) is also ongoing using the same instrumentation and processing infrastructure. To date, the TRAPUM and MMGPS-L searches have discovered more than 150 new pulsars[^4], the majority of which are MSPs. Here we present the first results from TRAPUM's survey of unassociated *Fermi*-LAT sources, which led to 9 of these MSP discoveries. The paper is organised as follows: Section [2](#s:survey){reference-type="ref" reference="s:survey"} describes the survey setup (recording and processing infrastructure, target selection and observation strategy); Section [3](#s:results){reference-type="ref" reference="s:results"} presents the new discoveries and subsequent investigation (localisation, timing and multi-wavelength follow-ups) and an estimate of the survey's sensitivity; and finally a brief discussion and conclusions follow in Sections [4](#s:discussion){reference-type="ref" reference="s:discussion"} and [5](#s:conclusions){reference-type="ref" reference="s:conclusions"}.
# Survey Properties {#s:survey}
## MeerKAT and the TRAPUM Processing Infrastructure
MeerKAT is a radio interferometer located in the Karoo, South Africa, consisting of 64 antennas with 13.5-m effective diameter. Here we give a brief description of MeerKAT and the TRAPUM infrastructure used to perform our pulsar search observations. For a full technical description of the instrument, we refer the reader to. At the time of data taking (between 2020 June and 2021 February), two receivers were available: the L-band receiver operating between 856--1712 MHz and the Ultra High Frequency (UHF) receiver between 544--1088 MHz. This survey was conducted at L-band, but follow-up observations of new pulsars were also made at UHF. At the L-band centre frequency, a coherent tied-array beam produced using all 64 MeerKAT antennas has a typical full-width at half maximum on the order of a few arcseconds. Furthermore, the data rate of complex voltages from the antennas is too high to record while observing. Efficiently searching the several-arcmin localisation region of an unassociated *Fermi*-LAT source in a single pointing therefore relies on the ability to form and record the Stokes intensities from a large number of coherent beams simultaneously. This capability is provided by the Filterbanking BeamFormer User Supplied Equipment (FBFUSE), a 32-node, GPU-based, software beamformer developed by the Max Planck Institute for Radio Astronomy. FBFUSE coherently sums the channelised complex voltages in real-time, using sky position-dependent complex weights computed by the purpose-built `Mosaic` software[^5] from a delay model provided by the MeerKAT Science Data Pipeline. FBFUSE can produce channelised time series for up to 864 coherent beams, as well as one incoherent beam produced by summing the (real-valued) Stokes intensities from each antenna. FBFUSE's beamforming algorithm requires a multiple of four antennas, and not all antennas are available for all observations. As such, we used either 56 or 60 antennas during our observations, depending on availability. The channelised data from each coherent beam are then recorded onto a distributed 3.5-PiB file system accessible from the Accelerated Pulsar Search User Supplied Equipment (APSUSE) instrument, a second, 60-node computer cluster, with two NVIDIA GeForce GTX 1080 Ti GPUs per node, on which the pulsar search takes place. The number of beams that can be stored is limited by the data rate at which APSUSE can record, and so down-sampling in time from the native data rate is necessary for a large number of beams to be recorded. All of our observations used the 4096-channel MeerKAT F-engine channeliser mode, but beamformed spectra were down-sampled in time by a factor of 16 from the native time resolution (\(4096/856\,{\rm MHz} = 4.785\,\upmu\)s) to give 76-\(\upmu\)s time resolution. With these time and frequency resolutions, up to 288 coherent beams could be formed and recorded. Additionally downsampling in frequency by a factor of two allowed for 480 coherent beams with 2048 frequency channels. Both of these recording modes were used in our survey. The full set of filterbank files for 480 coherent beams constitute 46 TiB of data per hour of observation, allowing at most 73 hours of data in this format to be stored for processing on APSUSE. It was therefore necessary to process these data quickly, identify promising candidates for further follow-up, and delete the raw data to ensure there was sufficient storage space for other TRAPUM projects to continue observing. Only the filterbank files for beams in which promising candidates were identified were retained for later use.
## Observing strategy
The very high gain and low system temperature of the MeerKAT array (\(G = 2.8\) K Jy\(^{-1}\) for the full array and \(T_{\rm sys} = 18\) K, @Bailes2020+MeerTIME) enable the detection of pulsars with low flux densities, even with short observation lengths. This motivated a strategy involving short pointings towards as many sources as possible. We chose 10 minute observations, for which we still obtain flux density limit estimates of around \(95\upmu\)Jy that compare favourably against previous surveys of Southern sources with longer observations (see Sections [3.4](#s:sensitivity){reference-type="ref" reference="s:sensitivity"} and [4](#s:discussion){reference-type="ref" reference="s:discussion"}). Previous surveys for MSPs within unidentified *Fermi*-LAT sources have revealed the importance of observing sources more than once to mitigate non-detections due to scintillation or unfavourable orbital phases (e.g. due to spider eclipses or "jerk" effects @Andersen2018+Jerk). In this paper we describe the first two passes of this survey, both performed at L-band. At least two further passes are planned for each source at the UHF band, where the lower frequency will provide additional sensitivity to pulsars with steep spectra, but where propagation effects and dispersive smearing are larger. During each observation, FBFUSE can be configured to form coherent beams that are distributed within the primary field-of-view either at pre-specified locations, or automatically using an optimal hexagonal tiling pattern. In this latter mode, used for all of our survey observations, the beam spacing is defined by an overlap parameter, which is the fractional sensitivity level of a 2-dimensional Gaussian function fit to the tied-array beam response, simulated by `Mosaic` at the centre of the frequency band, at the point midway between two neighbouring beams. The simple Gaussian model of the earlier versions of `Mosaic`, used for all our survey observations, tended to overestimate the true overlap when using the full MeerKAT array, and so e.g. points midway between beams in a tiling with an intended \(50\%\) overlap actually only achieved \(\sim\)`<!-- -->`{=html}40\(\%\) sensitivity. In the first pass of all our targeted sources, we used the 4096-channel mode, with a maximum of 288 coherent beams. The beam tiling patterns for each source were configured with a desired overlap of \(50\%\). We prioritised observing sources with larger positional uncertainties when they were at lower elevations, where the coherent tied-array beams covered a larger solid angle, to ensure that the full localisation region could be covered. For the sources with the largest uncertainty regions, observations were scheduled to ensure a sufficiently low elevation that the beam tiling covered a circle with a radius at least as large as the semi-major axis of the elliptical 95% confidence region (hereafter \(r_{\rm 95\%}\)) from . No significant sensitivity penalty due to ground spillover is incurred for elevations above \(30\degr\). For better-localised sources, the outer beams were well outside the *Fermi* source region. Further development of FBFUSE after our survey began provided the capability to alter the tiling overlap between sources within one observing block. For 75% of the observations in the second pass, we used the 2048-channel, 480-beam mode and adjusted the tiling overlaps for each source to maximise sensitivity while ensuring that the coherent beams covered a circular region with radius at least \(r = \sqrt{\log(0.01)/\log(0.05)} \,r_{\rm 95\%}\), the semi-major axis of an approximate 99%-confidence ellipse. An example of the resulting tiling pattern is shown in Figure [\[f:tiling_example\]](#f:tiling_example){reference-type="ref" reference="f:tiling_example"}. All sources had \(r_{\rm 95\%} < 5\arcmin\), such that the tiled region covered a very small patch around the centre of the \(\sim\)`<!-- -->`{=html}1\(\deg\) primary beam, meaning no significant sensitivity loss occurs for coherent beams that are not located at the central pointing position.
## Search Pipeline
The filterbank data from all coherent tied-array beams and incoherent beams were searched by a dedicated pipeline built around the `peasoup`[^6] GPU-accelerated pulsar search code, which performs an FFT-based acceleration search via time-domain resampling with incoherent harmonic summing (@Barr2020+peasoup, and described in detail in @Morello2019+HTRUrepro). The data were de-dispersed at trial dispersion measures (DMs) up to \(300\) pc cm\(^{-3}\) with spacing \(\Delta\textrm{DM}=0.06\) pc cm\(^{-3}\). This range covers the maximum DMs predicted by the Galactic electron density model of in the direction of any of our target sources, and 70 of our 79 target sources have maximum predicted DMs that are less than half of this range (i.e. below 150 pc cm\(^{-3}\)). The acceleration search covered a range of \(a \pm 50\) m s\(^{-2}\), which is slightly less than twice as large as the maximum acceleration seen from any known fully-recycled binary MSP (\(26.1\) m s\(^{-2}\) from PSR J0024\(-\)`<!-- -->`{=html}7204V, @Ridolfi2016+47Tuc) in the ATNF Pulsar Catalogue. Higher accelerations have been observed from relativistic binaries, such as pulsars with massive white dwarf companions or double neutron star systems, but these are all mildly-recycled systems, which typically do not have sufficient spin-down power to emit gamma-ray pulsations, although the Double Pulsar is one notable exception. At each trial DM, the spacing between acceleration trials was chosen according to the scheme described in, which ensures that signals lying between two acceleration trials suffer a total smearing that is not more than 10% larger than the other (unavoidable) smearing effects due to intra-channel dispersion smearing and finite time resolution. After performing the acceleration search, a clustering algorithm was used to search for clusters of candidates from neighbouring coherent beams with similar periods and DMs. Candidates with periods close to known radio-frequency interference (RFI) signals were excluded. This clustering step also aims to distinguish between astrophysical and terrestrial RFI signals, based on how the signal-to-noise ratio (S/N) drops off in neighbouring beams as a function of the angular offset from the beam in which the signal is detected most strongly. With the exception of side-lobe detections of very bright pulsars, signals from astrophysical point sources should only be detected in single beams, or in a small number of beams close to their sky positions. Candidates are therefore rejected if they are detected in many non-neighbouring coherent beams with S/Ns that drop off too slowly with angular offset, as this is indicative of a terrestrial interference signal. The data from the beam containing the strongest detection of each candidate cluster were then folded using the `PulsarX` software[^7], and folded candidates were scored by the Pulsar Image-based Classification System (PICS) machine-learning classifier. Candidates surpassing a conservatively low PICS-score threshold of \(10\%\), typically a few hundred per pointing, were then grouped for visual inspection. As mentioned previously, the full set of raw data could not be stored indefinitely, except for data from beams containing high-confidence pulsar candidates. Reduced data products, i.e. candidate lists, and folded data for every candidate, were retained from all observations.
## Target Selection {#s:targets}
We built a list of observing targets by identifying pulsar-like unassociated *Fermi*-LAT gamma-ray sources within the catalogue. Several studies have had success in using machine-learning classification techniques to identify pulsar candidates from the population of unassociated sources detected by the *Fermi* LAT. These methods rely on the fact that gamma-ray pulsars have characteristic spectral properties that distinguish them from other gamma-ray emitting source classes: 1) gamma-ray pulsar spectra typically have significant curvature, in that they deviate from a simple power-law spectrum due to a sub-exponential cutoff at photon energies above a few GeV; and 2) gamma-ray pulsars are very stable emitters on long time scales. The Random Forest algorithm has been shown to perform well for the purpose of classifying *Fermi*-LAT sources, and so we employed this method to rank sources from . We used five parameters from for the classification: `PLEC_SigCurve`, the significance of the log-likelihood improvement when fitting the source spectrum with a (curved) sub-exponentially cutoff power-law spectrum typical of pulsars, compared to a simple power-law; `Variability2_Index`, the chi-squared value of the energy flux measured in 2-month time bins; `Signif_Avg`, the point-source significance (brighter sources can have higher curvature and variability significance, including this parameter in the ranking accounts for this); `PLEC_Index` and `PLEC_Expfactor`, the photon index and the pre-factor \(a\) in the exponential cutoff term \(\exp(-aE^b)\), where \(b=2/3\) was used for all unassociated sources and all but six bright pulsars in . These final two parameters encode the energy at which the gamma-ray spectrum peaks, which for pulsars typically lies between \(0.5-4\) GeV, but which can be at much higher energies for active galactic nuclei (AGN) with curved spectra. We combined the gamma-ray source classes listed in into three broad categories: AGN, pulsars, or "other" (which contains e.g. supernova remnants and pulsar wind nebulae). The classifier was trained to identify these classes using sources with highly likely or confirmed associations listed in . To evaluate the classifier performance, we removed 33% of the sources, chosen at random, trained the classifier on the remaining population, and compared the classifier results to the known association classes. For identifying pulsars, the classifier had a precision of 82% (i.e., 82% of sources predicted to be pulsars were in fact pulsars) and a recall of 71% (i.e. 71% of pulsars in the sample were correctly identified as such). For each unassociated source in the Random Forest algorithm estimates the probability, \(P({\rm psr})\), of this source being a pulsar. We used this list of unassociated sources, ranked by their predicted pulsar probability, and made further cuts to reduce the number of sources to observe. First, only sources with declinations below \(+20^{\circ}\) were included. Next, we removed sources with Galactic latitudes within \(\left|b\right| < 10^{\circ}\) of the Galactic plane. This is because gamma-ray pulsars close to the Galactic plane tend to be slowly-spinning young pulsars, which have much narrower radio beams, and are therefore more likely to be radio-quiet gamma-ray pulsars undetectable to our survey. The Galactic plane is also being surveyed with MeerKAT with similar sensitivity as part of the MMGPS-L survey up to \(\left|b\right|<5^{\circ}\). Our survey was planned before the MMGPS-L survey region had been finalised, and hence our Galactic latitude cut aimed to avoid redundant observations. We then removed sources whose 95% confidence regions had semi-major axes larger than 5 arcmin, which is the largest region that could be covered in a single pointing using 288 coherent beams overlapping at 50% sensitivity. This semi-major axis cut removes around 40 sources that pass our other cuts, but several of these will be covered in future UHF observations, where the coherent beams are wider. Finally, we removed any sources with \(P(\rm psr) < 12\%\), which leaves 79 sources to search, while retaining 95% of the total sum of all \(P(\rm psr)\) values. Summing the classifier probabilities gives a (crude) prediction that 38 pulsars (not necessarily all detectable in radio surveys) should exist within these sources, with only 2 expected pulsars in the remaining sources that did not pass the probability threshold. The sources that were searched, along with the pulsar probabilities predicted by the classifier, are listed in Table [\[t:targets\]](#t:targets){reference-type="ref" reference="t:targets"}.
# Results {#s:results}
From our two-pass survey, nine candidate signals were identified in the final visual inspection step as likely being new millisecond pulsars, and all of these were later confirmed by additional detections (see Section [3.1](#s:followup){reference-type="ref" reference="s:followup"}). The pulse profiles for the newly-discovered MSPs are shown in Figure [\[f:pulse_profiles\]](#f:pulse_profiles){reference-type="ref" reference="f:pulse_profiles"}. Our survey also independently detected PSR J0312\(-\)`<!-- -->`{=html}0921, a black-widow MSP discovered recently in a GBT observation of 4FGL J0312.1\(-\)`<!-- -->`{=html}0921 (Tabassum et al., 2023, in prep.). Shortly after our first observations, another two target sources, 4FGL J0802.1\(-\)`<!-- -->`{=html}5612 and 4FGL J1231.6\(-\)`<!-- -->`{=html}5116, were identified as young gamma-ray pulsars by the direct detection of gamma-ray pulsations in the *Fermi*-LAT data by *Einstein\@Home*[^8]. Neither pulsar was detected in our survey, but this is not surprising, as only a very small fraction of young pulsars discovered in gamma-ray pulsation searches have been detected in radio observations.
## Follow-up Observations {#s:followup}
To confirm the pulsar nature of the detected candidates, we performed dedicated follow-up observations with MeerKAT at both L-band and UHF, and checked for archival data from previous search observations of the *Fermi*-LAT sources in which they were found. Three pulsars (J1757\(-\)`<!-- -->`{=html}6032, J1803\(-\)`<!-- -->`{=html}6707 and J1823\(-\)`<!-- -->`{=html}3543) were re-detected in archival data from the Parkes radio telescope. The five high-confidence pulsar candidates detected in the first pass were removed from the scheduled second survey pass, so that they could be observed in a dedicated set of confirmation observations along with candidates from the second pass. In these confirmation observations we employed a very dense tiling (with \(90\%\) overlap) with a smaller number of beams around the location of the coherent beam in which the candidate was detected most strongly in the initial survey observations. This dense tiling ensured high sensitivity for re-detecting these candidates, while also enabling us to more precisely localise each pulsar using the method described below (see Section [3.2](#s:localisation){reference-type="ref" reference="s:localisation"}). Eight of the pulsar candidates (excluding PSR J1036\(-\)`<!-- -->`{=html}4353, discussed below) were re-detected in these confirmation observations. One pulsar, PSR J1709\(-\)`<!-- -->`{=html}033 was only seen in the UHF observation, most likely due to unfavourable scintillation that was seen during observations in which this pulsar was detected. One pulsar, PSR J1036\(-\)`<!-- -->`{=html}4353 was not included in the confirmation observations, as it had not been immediately identified as a candidate due to a bug in the folding pipeline that caused it to be folded with the wrong acceleration sign. Instead, it was identified at a later date when the folded archives were corrected. It was confirmed in a later UHF observation as part of the next stage of this survey, which will be presented in a future paper.
## Localisation {#s:localisation}
Following the detection of a pulsar in our survey, the sky position could be estimated to a much higher precision than the size of a coherent beam by triangulation using the measured S/Ns in neighbouring beams. The method with which we localised candidates from our survey is described in Bezuidenhout et al. (2022, RASTI, submitted), based on the concept outlined in, which works as follows. Given the model of the coherent beam point spread function (PSF) provided by `Mosaic`, one can compute the expected ratio of the S/Ns that would be recovered in two neighbouring beams for a pulsar at any given point nearby. This expected ratio will match the observed ratio, within uncertainties, for a strip of positions between the two beams. These strips can then be computed for each pair of beams, and the location of the candidate can be inferred from where these strips all cross one another. This procedure is implemented by the `SeeKAT` package[^9]. We used `SeeKAT` on each of the confirmation observations, which had the most dense beam tilings. The best-fitting positions and uncertainties were in good agreement across the three observations (one at L-band, two at UHF), and so we combined these results by summing together the log-likelihood surfaces from each observation. The best-fitting positions and uncertainties for each pulsar are given in Table [\[t:pulsars\]](#t:pulsars){reference-type="ref" reference="t:pulsars"}, and an example of the localisation log-likelihood surface is illustrated in Figure [\[f:seekat\]](#f:seekat){reference-type="ref" reference="f:seekat"}. The `SeeKAT` positional uncertainties for new pulsar candidates are typically on the order of a few arcseconds, sufficiently precise to enable the identification of potential counterparts in multi-wavelength catalogues.
## Timing
Follow-up timing campaigns have begun for all newly detected pulsars. Initially, all timing observations were performed at other telescopes. The pulsars visible from the Northern hemisphere are followed up mainly using the Nançay (for PSRs J1526\(-\)`<!-- -->`{=html}2744 and J1823\(-\)`<!-- -->`{=html}3543) and Effelsberg (for PSRs J1709\(-\)`<!-- -->`{=html}0333 and J1906\(-\)`<!-- -->`{=html}1754) telescopes at L-band, while all other pulsars (and, initially, also those observed at other telescopes) are followed up at the Murriyang Parkes telescope, using the Ultra-wide-band Low (UWL) receiver, covering a frequency range from 0.7 to 4 GHz. Depending on the discovery S/N at MeerKAT or, when available, the S/N at the improved position, we follow-up the TRAPUM discoveries for 1 to 2 hrs. Observations are carried out in search mode. Whenever possible the pulsars have been observed with a pseudo-logarithmic cadence to help achieve phase coherence in our timing analysis. Phase-connected timing solutions for two MSPs, described in detail below, were obtained using the `Dracula` algorithm[^10]; this was necessary given the sparsity of the timing data in both cases. Obtaining a phase-connected timing solution for the remaining seven pulsars has required a dedicated timing campaign using the Pulsar Timing User Supplied Equipment (PTUSE) system developed for the MeerTime project, which additionally provides coherent de-dispersion and full Stokes polarisation information. This MeerKAT timing campaign, along with follow-up searches for gamma-ray pulsations from these seven pulsars, will be presented in a dedicated paper (Burgay et al., 2023, in prep.). For these seven pulsars, we give a preliminary timing solution in Table [\[t:pulsars\]](#t:pulsars){reference-type="ref" reference="t:pulsars"}, obtained using `PRESTO`'s `fit_circular_orbit.py` routine to fit a sinusoidal modulation to the observed barycentric spin periods from multiple observations without requiring phase-alignment across observations. We also plot the orbital properties of the 8 newly-detected binary MSPs (excluding PSR J1709\(-\)`<!-- -->`{=html}0333 which appears to be an isolated pulsar) according to these ephemerides, in comparison to the overall MSP population, in Figure [\[f:mass_vs_PB\]](#f:mass_vs_PB){reference-type="ref" reference="f:mass_vs_PB"}.
### PSR J1526\(-\)`<!-- -->`{=html}2744 {#s:J1526}
Using the `SeeKAT` position as a starting point, we obtained a phase-connected timing solution using TOAs produced from the original L-band observation, the two UHF observations, as well as several Nançay and Parkes observations taken in 2021 as part of our dedicated timing campaign. This timing solution provided precise constraints on the pulsar's orbital semi-major axis and time of ascending node, as well as a refined position, but the spin-down rate was not significantly measurable and was highly correlated with the declination. We used the radio timing solution to search for gamma-ray pulsations in the photon arrival times measured by the *Fermi* LAT. For this we used `SOURCE`-class gamma-ray photons from within a \(3\degr\) region-of-interest around the `SeeKAT` position, with energies greater than 100 MeV and with a zenith angle below \(105\degr\) according to the "Pass 8" `P8R3_SOURCE_V2` instrument response functions [^11]. To increase sensitivity to faint pulsations, we used `gtsrcprob` to compute photon probability weights, using the spectral and spatial parameters for nearby sources from the DR2 catalogue, as well as the `gll_iem_v07.fits` Galactic and `iso_P8R3_SOURCE_V2_v1.txt` isotropic diffusion emission models. Folding the gamma-ray data over the validity interval of the radio timing solution did not yield a significant detection. To search for gamma-ray pulsations in earlier data it was necessary to search over a 5-dimensional parameter space (\(\alpha\), \(\delta\), \(\nu\), \(\dot\nu\) and \(P_{\rm orb}\)). The pulsar's orbital semi-major axis and time of ascending node were already constrained precisely enough by the initial radio ephemeris that only one trial was required in these dimensions. The search was performed using the weighted \(H\)-test, a statistic that normally performs an optimal incoherent sum of the Fourier power in the first 20 harmonics. In this case, we only summed over 3 harmonics, as detecting power in higher harmonics requires increased search grid density in each dimension, but gamma-ray pulsars have most power in lower harmonics. This search detected a significant pulsed signal with \(H = 108.9\), and with phase-connected pulsations visible from the start of the LAT data. Following this detection, we derived a precise 13-yr gamma-ray timing solution by varying the timing parameters to maximise the unbinned Poisson log-likelihood of the weighted photon phases using a template pulse profile consisting of two wrapped Gaussian peaks whose parameters were also free to vary in the fit. The best-fitting parameter values and uncertainties are given in Table [1](#t:J1526){reference-type="ref" reference="t:J1526"}. The gamma-ray photon phases according to the best-fitting timing model and template pulse are shown in Figure [\[f:J1526_fermi\]](#f:J1526_fermi){reference-type="ref" reference="f:J1526_fermi"}. We also tested for but did not significantly detect proper motion (\(\left|\mu\right| < 49\,\textrm{mas\,yr}^{-1}\)) or eccentricity (\(e < 8\times10^{-4}\)).
The nature of the companion star in this system is currently unclear. The short orbital period (\(4.9\) hr) and low minimum companion mass (\(0.08\,M_{\odot}\)) suggest a heavy black-widow or light redback companion, but no radio eclipses typical of these systems have been seen. Many black-widow binaries have detectable optical counterparts, and so we searched for an optical counterpart to this system using the ULTRACAM high-speed multi-band imager on the 3.5-m New Technology Telescope (NTT) at ESO La Silla. The longest observation lasted 3.5 hr, covering orbital phases between 0.1--0.8. We did not find any counterpart at the pulsar's timing position, with \(3\sigma\) magnitude upper limits of \(i_s = 23.2\),\(g_s = 24.0\), and \(u_s=23.0\) in the deepest 5-minute stacked image, which was obtained close to the companion's superior conjunction, where a heated black-widow companion would appear at its brightest. It therefore seems likely that the companion is a light-weight (\(M_{\rm min} = 0.083\,M_{\odot}\)) white dwarf. If confirmed, e.g. through the detection of a non-variable optical counterpart below our ULTRACAM detection threshold, this would be the shortest orbital period of any known fully-recycled Galactic MSP--WD system, and the second shortest orbit of any PSR--WD binary system, after the relativistic binary PSR J0348+0432. For the observed \(\textrm{DM}=30.95(3)\) pc cm\(^{-3}\), predicts a distance of 1.3 kpc. The gamma-ray energy flux from this source above 100 MeV is \(G_{100} = 2.5\pm 0.4\times10^{-12}\) erg cm\(^{-2}\) s\(^{-1}\), corresponding to a gamma-ray luminosity of \(L_{\gamma} = 5.3\times10^{32}\,{\rm erg\,s}^{-1}\). This can be compared to the pulsar's spin-down power \(\dot{E} = 4 \pi^2 I \nu \dot{\nu} = 9.1\times10^{33}\,{\rm erg\,s}^{-1}\), for an assumed pulsar moment-of-inertia \(I=10^{45}\) g cm\(^{2}\), corresponding to an efficiency \(\eta = L_{\gamma} / \dot{E} \approx 6\%\), which is typical of gamma-ray MSPs. Using the timing solution obtained above, we carried out a search for continuous gravitational waves from this pulsar. Continuous gravitational wave emission is expected due to a non-axisymmetric deformation (\(\epsilon\)) of the neutron star and the dominant signal frequency is twice its rotational frequency. Our search was therefore targeted at a frequency \(f\) \(\approx\) \(803.5\) Hz and spin down \(\dot{f}\) \(\approx\) \(-1.14 \times 10^{-15}\) Hz/s and directed at the timing position. An *a priori* estimate of the detectable strain amplitude using Advanced LIGO data yields a value about an order of magnitude larger than the spin-down upper limit amplitude, \[h_0^{sd} = \left ( \frac{5}{2}\frac{GI |\dot{\nu}|}{c^{3}d^{2}\nu} \right) ^ {1/2}\] where we again assume \(I=10^{45}\) g cm\(^2\). This is the strain amplitude of the signal assuming that all the rotational kinetic energy lost by the pulsar (\(\dot{E}\)) was converted into gravitational waves. We therefore do not expect a detection, but, in the spirit of leaving no stones unturned , we carried out the search. We used all of the Advanced LIGO data from the Hanford and Livingston detectors collected during the O1, O2 and O3 runs. The gamma-ray timing solution allows us to coherently combine these datasets using a single template, thus achieving the maximum possible sensitivity this search could have to date. Loud detector glitches in the data were removed through gating. The frequency range relevant to this search is free of known lines in the detectors. The multi-detector, matched-filtering \(\mathcal{F}\)-statistic was used for the analysis. Our search result has a p-value of %, estimated using off-source data. This result implies a non-detection and based on it we set a 95% confidence upper limit of \(1.25 \times 10^{-26}\) on the intrinsic gravitational wave amplitude. Our upper limit is a factor of \(\approx\) \(20\) larger than the spin down upper limit, and the uncertainty on this upper limit is not more than \(\hULpercentUncertainty\)% including calibration uncertainties. Translating the amplitude upper limit into an upper limit on the ellipticity of the source, we constrain the ellipticity of J1526\(-\)`<!-- -->`{=html}2744 to be \(<2.45\times 10^{-8}\), which is close to the minimum ellipticity proposed for millisecond pulsars by. We also searched in a band of \(f\) and \(\dot{f}\) accounting for mismatches between the phase of the gravitational wave signal and the phase locked to the electromagnetic observations. Such mismatches could result from a differential rotation between the parts of the star emitting the gravitational wave and the electromagnetic pulsations or if the star was biaxial and consequently freely-precessing. We searched in a bandwidth of 0.4% of the spin parameter values following, and estimated the p-values of the results as done in. The results from this band search were also consistent with expectations from Gaussian noise.
### PSR J1803\(-\)`<!-- -->`{=html}6707 {#s:J1803}
PSR J1803\(-\)`<!-- -->`{=html}6707 was the first pulsar to be found in our survey, and was quickly confirmed by a detection in 1 hr of archival Parkes data from 2015. Detection of the pulsar in this observation required a significant jerk term, indicating that the pulsar was likely to be in a short-period binary system. Pulsations were only detected for \(\sim\)`<!-- -->`{=html}30 mins in one of two 1 hr dedicated follow-up observations with Parkes in December 2020, indicative of a wide eclipse typical of black-widow or redback systems. Establishing an orbital timing solution (given in Table [2](#t:J1803){reference-type="ref" reference="t:J1803"}) revealed a \(\sim\)`<!-- -->`{=html}9 hr orbit and a minimum companion mass of 0.26 \(M_{\odot}\) (assuming a pulsar mass of 1.4 \(M_{\odot}\)), as well as significant variations in the orbital period requiring several orbital frequency derivative terms to describe the orbital phase throughout the 1 yr of phase-connected timing data. These features are characteristic of redback binary systems. We used the UHF MeerKAT observations to estimate the sky position of this pulsar using `SeeKAT`. The refined position is coincident with the location of a star in the *Gaia* DR3 catalogue. We also observed this star with ULTRACAM, which revealed optical variability with the same 9 hr periodicity, confirming that this is indeed the optical counterpart. The optical light curve, shown in Figure [\[f:ultracam\]](#f:ultracam){reference-type="ref" reference="f:ultracam"}, varies by around 1.5 mag, with a single peak indicative of significant heating via irradiation from the pulsar but the counterpart is detectable at all orbital phases, as is typical for nearby redback companions which tend to have hot surfaces (\(T > 4000\) K) even on the non-irradiated side. Light curve modelling to estimate properties of the companion such as its temperature, irradiation, radius and the binary inclination angle will be performed in a dedicated follow-up project (Phosrisom et al. 2023, in prep.). As with PSR J1526\(-\)`<!-- -->`{=html}2744, we folded the gamma-ray data using the radio ephemeris to check for gamma-ray pulsations. Significant pulsations are detected within the 1 yr interval in which the radio ephemeris is valid, with a weighted \(H\)-test of \(H=51\), corresponding to a \(5.6\sigma\) detection using the false-alarm probability calibration from. These pulsations are shown in Figure [\[f:J1803_fermi\]](#f:J1803_fermi){reference-type="ref" reference="f:J1803_fermi"}. However, these pulsations quickly disappear when extrapolating to earlier data. This is not unexpected, as Taylor-series models for orbital period variations lack predictive power, and the spin-down rate is not measured precisely enough in the radio data to fold the 13 yr of *Fermi*-LAT data. We have not been successful in extending the ephemeris by timing the gamma-ray data, likely as a result of the pulsar's faint gamma-ray flux and the large amplitude of the orbital phase variations. A longer radio timing baseline will hopefully solve this problem by providing a more precise spin-down rate measurement, and a longer phase-connected radio ephemeris that will fold more *Fermi*-LAT data to provide better statistics with which to build a gamma-ray pulse profile template to search with. Nevertheless, this detection confirms the association between 4FGL J1803.1\(-\)`<!-- -->`{=html}6708 and PSR J1803\(-\)`<!-- -->`{=html}6707. The gamma-ray energy flux from this source above 100 MeV is \(G_{100} = 5.0\pm 0.5\times10^{-12}\) erg cm\(^{-2}\) s\(^{-1}\), corresponding to an efficiency \(\eta = 1.7\%\) for the assumed distance of 1.4 kpc (for \(\textrm{DM}=38.382\) pc cm\(^{-3}\)), which is again fairly typical of gamma-ray MSPs. Using data from the first *eROSITA* all-sky survey, the X-ray counterpart of PSR J1803\(-\)`<!-- -->`{=html}6707 was independently detected (positional match within \(\sim\) 5\(^{\prime\prime}\)) in a pilot search for likely X-ray counterparts of unassociated *Fermi-LAT* sources. Figure [\[f:eROSITA_Image\]](#f:eROSITA_Image){reference-type="ref" reference="f:eROSITA_Image"} depicts the X-ray counterpart as seen in eRASS1. The formal detection significance in the \(0.2-2.3\,\mathrm{keV}\) bandwidth is \(4.6\sigma\) for a vignetting corrected exposure time of 167.63s and a count rate of \((7.2\pm2.4)\times10^{-2}\,\mathrm{ct\,s^{-1}}\) obtained from all seven telescope modules. The source is not detected in the \(2.3-8.0\,\mathrm{keV}\) bandpass. The detected X-ray photons do not support a detailed spectral analysis. *eROSITA*'s temporal resolution is 50 ms which prevents the detection of a periodicity at the millisecond level as well. Assuming a power-law spectrum with a photon index of 2.0 and an absorption column density of \(6\times10^{20}\,\rm{cm^{-2}}\), the *eROSITA* counting rate implies an unabsorbed X-ray flux of \(F_{\rm X} = (1.5\pm0.5)\times10^{-13}\,\mathrm{erg\,s^{-1}\,cm^{-2}}\) in the \(0.2-10\,\mathrm{keV}\) band. For the @YMW16 distance we obtain an X-ray luminosity of \(L_{\rm X} \sim 3.5\times10^{31}\,\mathrm{erg\,s^{-1}}\) and an X-ray efficiency of \(L_{\rm X}/\dot{E} \sim 4.7\times 10^{-4}\), a typical value for X-ray detected MSPs. A significant contribution by the \(g\sim20\) optical counterpart to the detected X-ray flux is not expected. The X-ray-to-optical flux ratio of the system is around 1, much greater than that expected for coronal X-ray emission from stars. It is interesting to note that, considering the \(\gamma\)-ray flux of the likely counterpart, the \(\gamma\)-ray to X-ray flux ratio of this source is comparatively small at \(F_{\gamma}/F_{\rm X} \sim 35\). This value is located at the lower end of the observed distribution for high-energy pulsars, which extends over a range of around \(10-10^4\). It is therefore likely that the X-ray flux contains a contribution from, or is dominated by, an intra-binary shock, as is commonly seen in redback binary systems. Follow-up observations are planned to further investigate the X-ray emission from this system.
### PSR J1036\(-\)`<!-- -->`{=html}4353
This pulsar is both the fastest-spinning (with \(\nu \approx 595\) Hz), and the most highly-accelerated (with \(a \approx 15\) m s\(^{-2}\)) pulsar found by our survey. Follow-up observations have revealed this to also be a redback system. A preliminary orbital solution revealed a \(\sim\)`<!-- -->`{=html}6.2 hr orbit and a minimum companion mass of \(0.23\) M\(_{\odot}\). Several non-detections have occurred during our timing campaign in the half of the orbit in which the pulsar is behind its companion, likely due to eclipses by material from the redback companion. A single *Gaia* source lies within its `SeeKAT` localisation region. Like PSR J1803\(-\)`<!-- -->`{=html}6707, we observed this source with ULTRACAM, revealing a double-peaked light curve, shown in Figure [\[f:ultracam\]](#f:ultracam){reference-type="ref" reference="f:ultracam"}, very typical of redback binary systems. Modelling of this data will also be presented in a future paper (Phosrisom et al. 2023, in prep.).
### The other new MSPs
Of the remaining MSPs, one (PSR J1709\(-\)`<!-- -->`{=html}0333) appears to be isolated, while three (PSRs J1623\(-\)`<!-- -->`{=html}6939, J1823\(-\)`<!-- -->`{=html}3543 and J1858\(-\)`<!-- -->`{=html}5422) have companion masses and orbital periods that are consistent with helium white dwarf (He-WD) companions, but the remaining two warrant further mention. The first, PSR J1906\(-\)`<!-- -->`{=html}1754 lies at an unusual position in the orbital parameter space shown in Figure [\[f:mass_vs_PB\]](#f:mass_vs_PB){reference-type="ref" reference="f:mass_vs_PB"}, with a minimum companion mass (\(M_{\rm c} > 0.05\,M_{\odot}\)) much lower than that of typical white-dwarf companions, but with an orbital period (\(P_{\rm orb} = 6.5\,\textrm{d}\)) much longer than a typical black-widow system. Of the MSP binaries with \(P_{\rm orb} > 1\,\textrm{d}\), only PSR J1737\(-\)`<!-- -->`{=html}0811 has a similarly low minimum companion mass. The low minimum mass could indicate a very face-on binary inclination angle, with \(i < 14.5\deg\) required for \(M_{\rm c} = 0.2\,M_{\odot}\). However, such orbits are *a priori* unlikely, as only around 3% of orbits should have a lower inclination, assuming a random distribution of viewing angles. The final pulsar, PSR J1757\(-\)`<!-- -->`{=html}6032, has a larger minimum companion mass (\(M_{\rm c} > 0.45\) M\(_{\odot}\)) that suggests it is perhaps more likely to have a CO-WD companion, similar to that of the relativistic binary PSR J1614\(-\)`<!-- -->`{=html}2230.
## Sensitivity {#s:sensitivity}
As our observations were not flux-calibrated, we estimated the flux density thresholds (\(S\)) above which a pulsar should have been detectable to our survey. To estimate this, we used the pulsar radiometer equation, \[S = \frac{\rho \, (T_{\rm sys} + T_{\rm sky})}{\beta \, G\, n_{\rm ant} \, \, \sqrt{n_{\rm pol} T_{\rm obs} B}} \sqrt{\frac{w}{1-w}}\, \label{e:radiometer}\] where we have assumed: \(G=0.044\) K Jy\(^{-1}\) is the gain per antenna; \(\rho=9\) was the S/N threshold required for a candidate to be folded for visual inspection; \(n_{\rm ant}\) is the number of antennas; \(n_{\rm pol}=2\) polarisations are summed; \(T_{\rm obs}\) is the exposure time; \(B=700\,\)MHz is the estimated useable bandwidth after RFI excision; and \(T_{\rm sys} = 22.5\) K is the combination of the receiver temperature, atmospheric and ground spillover contributions. The sky temperature, \(T_{\rm sky}\), towards each source is taken from the 408 MHz all-sky map of, reprocessed by, and scaled to the central frequency of 1284 MHz according to an assumed spectral index of \(-2.6\). The resulting flux-density limits are given in Table [\[t:targets\]](#t:targets){reference-type="ref" reference="t:targets"}. Estimated sensitivity also depends strongly on the assumed pulse duty cycle, \(w\). For the new MSPs, this varies from \(5\%\) (PSR J1803\(-\)`<!-- -->`{=html}6707) to \(35\%\) (PSR J1906\(-\)`<!-- -->`{=html}1754); we have assumed \(w=15\%\) for our estimated sensitivities. The remaining factor in Equation [\[e:radiometer\]](#e:radiometer){reference-type="ref" reference="e:radiometer"}, \(\beta\), accounts for various losses incurred during observing and searching. This includes constant fractional losses due to the 8-bit voltage digitisation (5%) and beamforming efficiency (5%); sensitivity losses due to using incoherent harmonic summing (15% for our assumed 15% duty cycle, @Morello2020+FFA); finite time resolution and intra-channel dispersion smearing (both \<1%). In addition to these fractional losses, there are random losses due to the location of a signal between FFT bins (averaging 8%), DM trials and acceleration trials. We estimate these losses for an assumed \(\nu=500\) Hz, \(\textrm{DM}=100\) pc cm\(^{-3}\) pulsar, via a Monte Carlo procedure-generating signals with random offsets from the nearest search trial and evaluating the S/N losses. The final loss factor to consider is that due to the angular offset between a pulsar and the centre of the nearest coherent beam. To evaluate this, for each source we drew random locations from the localisation probability densities and evaluated the sensitivity at that location of the nearest coherent beam according to the simulated PSF model from `Mosaic`. Combining all of these losses provided a Monte Carlo distribution for \(\beta\). For each source we quote three estimated sensitivities: \(S_{\rm opt}\) is the optimum sensitivity, assuming only the constant losses described above, and therefore estimates a fundamental flux density limit for our survey; while \(S_{\rm 50}\) and \(S_{\rm 95}\) are the flux density limits obtained using Equation [\[e:radiometer\]](#e:radiometer){reference-type="ref" reference="e:radiometer"} using the median and 95th centile values, respectively, of the \(\beta\) values obtained from our Monte Carlo estimates. Average sensitivities were \(S_{\rm opt} \approx 55\upmu\)Jy, \(S_{\rm 50} = 120\,(90)\,\upmu\)Jy, and \(S_{\rm 95} = 150\,(110)\,\upmu\)Jy for the first (second) pass, respectively. The flux density threshold for the incoherent beam (for pulsars lying outside the region tiled by coherent beams) is approximately \(\sqrt{n_{\rm ant}} \approx 8\) larger than that for a coherent beam.
# Discussion {#s:discussion}
In this paper, we have presented the first MeerKAT survey for new radio pulsars in unassociated *Fermi*-LAT sources. While our strategy of targeting pulsar-like gamma-ray sources is certainly not novel, the capabilities of this next-generation radio telescope do lead to significant advantages over previous surveys. Foremost among these is the extremely high sensitivity of the full MeerKAT array, surpassed only by the Arecibo and FAST telescopes, but unprecedented in the Southern Hemisphere. In Section [3.4](#s:sensitivity){reference-type="ref" reference="s:sensitivity"}, we estimated typical 95% flux density upper limits of \(\sim\)`<!-- -->`{=html}100 \(\upmu\)Jy, which can be compared to the nominal \(\sim\)`<!-- -->`{=html}200 \(\upmu\)Jy sensitivity (not including many of the loss terms that we consider) that was achieved with typical hour-long pointings in the similar survey of *Fermi*-LAT sources performed at Parkes by. Of the newly discovered pulsars, only PSR J1803\(-\)`<!-- -->`{=html}6707 is bright enough that it could perhaps have been discovered in hour-long Parkes observations, but such a discovery would have been complicated by jerk effects from its short orbital period and high acceleration. All of the new pulsars have eventually been detected in dedicated Parkes observations, but with low S/Ns that would be hard to detect without prior knowledge of the DM and spin period. A key factor for our survey, as also discussed for their survey of *Fermi*-LAT sources performed with the Arecibo telescope, is that these competitive flux-density thresholds are achieved with only short 10-min observations, bringing several benefits. First, these short observation times allow for more sources to be observed within a given observing time budget, while MeerKAT's rapid (\(\sim\)`<!-- -->`{=html}30 s) slew time ensures that costly overheads are not incurred by doing so. Second, short observations enable the search processing to be performed quickly (the computing cost of an acceleration search scales with at least \(T^3 \log(T)\) for integration time \(T\)), allowing storage space to be freed quickly enough that observation scheduling is not limited by this factor, and allowing us to search up to high accelerations. Third, short observation times enable the detection of very short-period binary MSPs. For very short orbits, the assumptions that go into an acceleration search (that the orbital motion within the observation can be approximated with a constant acceleration) break down as higher-order "jerk" terms become significant, with sensitivity to binary MSPs only maintained for observations lasting less than 10% of an orbit. Our short 10 min observation strategy mitigates the worst of the jerk effects, but still means that sensitivity is lost for binaries with periods shorter than \(\sim\)`<!-- -->`{=html}100 min. Only three Galactic field MSPs with shorter orbital periods than this are known, although all three are gamma-ray MSPs. Additionally, short observations mean that multiple passes can be performed to minimise missed discoveries due to the time-varying effects that can contribute to a pulsar's detectability in a given observation. Interstellar scintillation introduces time-and frequency-dependent variations in the observed spectrum, which can lead to a pulsar being undetectable over large frequency ranges. Indeed PSR J1709\(-\)`<!-- -->`{=html}0333 remained undetected in our L-band confirmation observation, despite a dense beam tiling that covered its now-known position. Our UHF confirmation observations reveal the reason for this-in both observations, separated by 70 min, the pulsar was only detected in \(\sim\)`<!-- -->`{=html}20\(\%\) (\(\sim\)`<!-- -->`{=html}100 MHz) of the bandwidth, indicative of scintillation. Subsequent detections with the Parkes UWL and the Effelsberg L-band receivers revealed similar behaviour. MSPs in black-widow and redback binaries can also be undetectable for large fractions of an orbit, usually (but not exclusively) around the pulsar's superior conjunction, as a result of diffuse intra-binary plasma dispersing, scattering and absorbing radio pulsations. The jerk terms that limit the sensitivity of an acceleration search to short period binaries also have an orbital phase dependence, and so sensitivity also depends on the orbital phase at which observations take place. Our short exposure, two-pass survey partially mitigates these effects. The case of PSR J1526\(-\)`<!-- -->`{=html}2744 further illustrates this: we were unable to detect this pulsar in dedicated follow-up observations with Parkes and Nançay prior to obtaining an orbital solution without performing jerk searches, a computationally expensive technique that has only recently started being employed in radio surveys of *Fermi* sources. One additional time-dependent effect that is specific to this survey is the fact that the shape of a coherent tied-array beam on the sky depends on the elevation at which a source is observed, with beams being more elongated for sources at lower elevations due to the smaller projected baselines between antennas. The beam tiling pattern used to cover a given source region therefore depends on the exact sidereal time and array configuration. Indeed, four of the nine pulsars discovered here were only detected in the second observation of their respective *Fermi* sources. This was partially due to the improved flexibility of FBFUSE's tiling patterns that was developed between the first and second passes, and our use of a coarser frequency resolution but larger number of coherent beams, which allowed us to cover a larger solid angle around each source with a more sensitive tiling pattern. Two of these pulsars, PSRs J1709\(-\)`<!-- -->`{=html}0333 and J1036\(-\)`<!-- -->`{=html}4353, lay just outside the corresponding *Fermi*-LAT source 95% localisation regions that we aimed to tile with coherent beams in the first pass, but were detected in the second pass where we targeted a larger nominal 99% confidence region. The other two pulsars that were only detected in the second pass, PSRs J1623\(-\)`<!-- -->`{=html}6936 and J1757\(-\)`<!-- -->`{=html}6032, lay in a less sensitive location in the coherent beam tiling pattern in their first observation, midway between three neighbouring beams, where sensitivity was \(\sim\)`<!-- -->`{=html}50% of that at the centre of a beam, but were at a more favourable location in the second observation. One final advantage to our survey, over projects using single-dish telescopes, is the rapid and precise localisation that can be obtained from a multi-beam detection using `SeeKAT`. This has two significant scientific benefits: rapid localisations enable immediate multi-wavelength follow-up and catalogue searches; and precise knowledge of a pulsar's location greatly assists in obtaining a phase-connected timing solution to fully exploit the scientific potential of a new MSP discovery. In the absence of an interferometric localisation, it often takes lengthy timing campaigns to reveal the location of a pulsar, as astrometric parameters can be highly degenerate with a pulsar's spin-down rate (and sometimes orbital parameters) until a data set spanning several months has been obtained. The `SeeKAT` positions for PSRs J1803\(-\)`<!-- -->`{=html}6707 and J1036\(-\)`<!-- -->`{=html}4353, were precise enough to unambiguously link these pulsars to their *Gaia* counterparts, establishing these as redback binary systems, and providing sub-milliarcsecond astrometric uncertainties that are sufficiently precise that these parameters no longer needed to be fit for when building a timing solution. For PSR J1526\(-\)`<!-- -->`{=html}2744, the \(\sim\)`<!-- -->`{=html}2\(\arcsec\) positional uncertainty represents a reduction of a factor of \(\sim\)`<!-- -->`{=html}10,000 in the number of sky positions that had to be searched to detect gamma-ray pulsations, greatly decreasing the required computational effort. With only two pointings towards our targeted sources there is still a high chance that some otherwise detectable pulsars among our targets may have been missed due to scintillation or eclipsing. We have therefore planned another two observations towards the sources observed here using the UHF receiver. All 9 pulsars discovered here were also detected at a higher S/N in our UHF observations, illustrating the sensitivity gains that can be made by observing at lower frequencies to exploit typical MSP spectra that decrease with frequency, and MeerKAT's extremely low RFI environment at UHF. It is important to consider what prospects remain for further detections among *Fermi*-LAT sources. To obtain a crude estimate for the number of pulsars remaining in this sample, we can simply sum the probabilities of each source being a gamma-ray pulsar that the Random Forest classifier computed in Section [2.4](#s:targets){reference-type="ref" reference="s:targets"}, obtaining \(\sum P(\textrm{psr}) = 38.4\). Subtracting from this sum the 9 new pulsars discovered here, plus the 3 others mentioned in Section [3](#s:results){reference-type="ref" reference="s:results"} that were recently discovered by other projects, gives \(\sim\)`<!-- -->`{=html}26 pulsars remaining in this sample, and around 190 pulsars in the full catalogue. However, this is certainly an overestimate, as many of these sources have been searched already with other telescopes, and therefore our target list represents the remainder from a larger, unbiased sample from which more easily detected pulsars have already been removed. There are hints of this effect in our results: five of the nine new pulsars had fairly low pulsar probabilities \(P(\rm psr) < 25\%\), perhaps because the more promising sources have already been surveyed extensively by other telescopes, while the lower probability targets tend to be fainter, newer sources many of which have not been searched before. This also suggests that further searches of a larger number of apparently less promising *Fermi*-LAT sources below the \(P(\rm psr)\) threshold that we used here may yet bear fruit. Furthermore, even if this is an accurate estimate of the number of gamma-ray pulsars in our sample, it does not mean that there are 26 detectable radio pulsars in this sample. A large fraction of the young gamma-ray pulsars detected by the *Fermi*-LAT remain undetected in radio observations (and indeed the two recently-discovered young gamma-ray pulsars that were covered in our survey were not detected), presumably due to the viewing angle missing the radio beam. While we deliberately avoided sources at low Galactic latitudes in order to reduce the chance of young pulsars entering our sample, we note that there are 17 sources in our list at lower Galactic latitudes than 4FGL J0802.1\(-\)`<!-- -->`{=html}5612, one of the two recently discovered young pulsars in our sample. Although MSPs tend to have wider radio beams, and hence are less likely to be "radio-quiet", a handful of gamma-ray MSPs have now been discovered in the *Fermi*-LAT data, but have remained undetected in deep radio follow-up searches, and similar objects may exist within our targets. Nevertheless, it is certain that there are still several detectable MSPs lurking in our target list, and iterative releases bring new, albeit fainter, unassociated sources to target. Indeed, since our survey began, likely MSP binaries have been discovered in optical and X-ray searches of three of our target sources, 4FGL J0540.0\(-\)`<!-- -->`{=html}7552, 4FGL J0940.3\(-\)`<!-- -->`{=html}7610, and 4FGL J1120.0\(-\)`<!-- -->`{=html}2204, while a fourth, 4FGL J1702.7\(-\)`<!-- -->`{=html}5655, has recently been identified through the detection of gamma-ray eclipses. 4FGL J1120.0\(-\)`<!-- -->`{=html}2204 in particular appears to have a white dwarf companion in a fairly long (15 hr) orbit, but has remained undetected in many previous long radio observations, as well as in our 10-minute pointings. Its non-detection is therefore unlikely to be explained by orbital or eclipsing effects, suggesting that deeper searches, in addition to our continuing "shallow" survey, may be necessary to uncover the nature of many of the remaining unidentified but pulsar-like *Fermi*-LAT sources.
# Conclusions and Future Work {#s:conclusions}
We have presented the first results from TRAPUM's survey for new pulsars in unassociated *Fermi*-LAT sources using the MeerKAT radio telescope, discovering nine new MSPs, of which eight are in binary systems. Our results continue the trend of short-period binaries being discovered at a far higher rate within *Fermi*-LAT sources than in untargeted surveys, with the discovery of two new redback binaries (PSRs J1036\(-\)`<!-- -->`{=html}4353 and J1803\(-\)`<!-- -->`{=html}6707) with optically bright companion stars and radio eclipses, and a third (PSR J1526\(-\)`<!-- -->`{=html}2744) that is possibly the most compact known MSP-WD binary system in the Galactic field. Two other MSPs have preliminary orbital solutions that mark their companions as outliers among the known MSP-binary population, PSR J1757\(-\)`<!-- -->`{=html}6032 appears to have a less typical CO-WD companion, while PSR J1906\(-\)`<!-- -->`{=html}1754 has a much longer orbital period than any known black widow, but a minimum companion mass that is far lighter than that of a typical WD companion. We obtained phase-connected timing solutions for two of the new MSPs, PSRs J1526\(-\)`<!-- -->`{=html}2744 and J1803\(-\)`<!-- -->`{=html}6707, using timing observations at Parkes and Nançay, that enabled us to detect gamma-ray pulsations in the *Fermi*-LAT data. Orbital period variations in PSR J1803\(-\)`<!-- -->`{=html}6707 precluded further extrapolation of the timing solution, but for PSR J1526\(-\)`<!-- -->`{=html}2744, gamma-ray timing provides a full 14-yr timing solution. This in turn enabled us to search for continuous gravitational waves from this pulsar in the Advanced LIGO O1, O2 and O3 data, although none were detected and the strain upper limit remains well above the pulsar's spin-down luminosity budget. A dedicated timing campaign is underway at MeerKAT to obtain phase-connected timing solutions, and afterwards gamma-ray pulsations, from the remaining seven discoveries (Burgay et al., 2023, in prep.), as are optical follow-up observations and modelling of the new redback binaries (Phosrisom et al., 2023, in prep.). Our results also emphasise the promise of continued radio surveys of *Fermi*-LAT sources: several high-confidence pulsar candidates still remain within our target list but have eluded detection in ours and earlier surveys, perhaps due to scintillation, eclipses or simple intrinsic faintness. Indeed, our TRAPUM survey continues to make discoveries beyond the first results presented here. Two further passes of the sources surveyed here are currently underway using MeerKAT's UHF receiver, from which we have discovered 8 additional MSPs already[^12]. We have also performed several hour-long observations of a group of high confidence redback candidates that were previously identified in optical observations, detecting a further three new MSPs that will be presented in a dedicated paper (Thongmeearkom et al., 2023, in prep.). We will continue our initial survey strategy of observing a large number of sources with short (10-minute) observations, expanding our target list to include more sources from the most recent catalogue, but deeper observations may be required to fully explore the most promising gamma-ray sources. | {'timestamp': '2022-12-19T02:13:36', 'yymm': '2212', 'arxiv_id': '2212.08528', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08528'} |
# Introduction {#sec_intro}
On Earth, gravity-induced turbulent convection holds a central place as its ability to vigorously transport heat in nature and industrial applications. The main challenge of convective turbulence studies is to explore the basic properties of heat transport, namely, the "constitutive law". The paradigmatic setup to study convective flows is the Rayleigh-Bénard (RB) convection, i.e., a fluid layer heated from below and cooled from above. The heat transport properties are then related to the scaling behavior between the dimensionless heat flux (characterized by the Nusselt number \(\Nun\)) and the dimensionless temperature difference (characterized by the Rayleigh number \(\Ra\)), i.e., \(\Nun \sim \Ra^\beta\) where \(\beta\) is the scaling exponent. Decades of studies on RB setup show the emergence of universal scaling exponent in the constitutive law. Typically, one theoretically arguments \(\beta = 1/3\) from the elegant theory of marginal stability, or \(\beta \approx 0.3\) from experimental observations in the classical regime, and \(\beta = 1/2\) in the ultimate regime predicted by a mixing length model assuming that the heat flux is fully controlled by turbulence. Both heat transport scaling relations are extensively examined by various experimental and numerical investigations. Grossmann and Lohse developed the unified theory, which well explains the observed heat transport scaling behaviors and successfully describes the existing experimental and numerical data. In microgravity, as the gravity is however almost absent, gravity-induced convection becomes too feeble to transport matter and heat. Vibration, omnipresent in science and technology, has been shown to be an attractive way to operate fluids, modulate convective patterns, and control heat transport by creating an "artificial gravity", e.g., vibration shapes liquid interfaces in arbitrary direction, vibration levitates a fluid layer upon a gas layer, vibration selects patterns through the parametric response, vibration significantly enhances or suppresses heat transport depending on the mutual direction of vibration and temperature gradient. Vibroconvection, resulting directly from a non-isothermal fluid subjected to the external vibration, is very pronounced under microgravity conditions and provides a potential mechanism of heat and mass transport in absence of gravity-induced convection. Elucidating the potential constitutive law of vibroconvective turbulence and its underlying mechanism is not only of great importance in microgravity science, but also provides practical guiding significance for space missions and microfluidic technologies. In past decades, due to the difficulty of microgravity experiments, the experimental studies on vibroconvection at low gravity were limited. A known experiment was carried out with the ALICE-2 instrument onboard MIR station, which revealed the vibrational influence on the propagation of a temperature wave from a heat source in near-critical fluids. The other known experiment was conducted in the parabolic flights during the 46-th campaign organized by the European Space Agency, which reported the first direct experimental evidence of vibroconvection in low gravity. There are extensive theoretical and numerical investigations of vibroconvection under weightlessness conditions. In the limiting case of high-frequencies and small amplitudes, the averaging technique was applied to theoretically deduce the dynamical equation of the mean flows. Based on the averaged equations, the onset and bifurcation scenarios of vibroconvection were widely investigated in square, rectangular, and cubic enclosures. The synchronous, subharmonic and non-periodic responses to external vibration were observed in vibroconvection from a parametric study over a wide range of frequencies and amplitudes. The parametric and Rayleigh-vibrational instability were examined in vibroconvection in the absence of gravity. The heat transport enhancement near the onset of vibroconvection were also investigated. However, the basic properties of constitutive law in vibroconvective turbulence have been rarely addressed. All together, one key question of whether heat transport with scaling behavior and universal exponents can exist in vibroconvective turbulence remains unclear. Here, due to the limitation of microgravity experiments, we theorectically and numerically unveil the emergence of unified constitutive law and underlying mechanism of vibroconvective turbulence. We carried out a series of direct numerical simulations on vibroconvection in a wide range of amplitudes and frequencies. We find that the thermal columns, generated by vibration-induced "artificial gravity", are the major structures responsible to transport heat, which are different from that in gravity-induced convection, like the large-scale circulation in classical RB convection. We also find that a unified scaling law in vibroconvection indeed exists, and the scaling behavior is determined by the competition between the thermal boundary layer (TBL) and oscillating boundary layer (OBL) induced by extermal vibration. The heat transport scaling transits from \(\Nun \sim \Am^{-1}\Ren_\mathrm{os}^{2}\) in the OBL-dominant regime to \(\Nun \sim \Am^{-1}\Ren_\mathrm{os}^{4/3}\) in the TBL-dominant regime, where \(\Am\) is the vibration amplitude and \(Re_\mathrm{os}\) is the oscillational Reynolds number. It is concluded that vibroconvective turbulence in microgravity defines a unified constitutive law with distinct underlying mechanism of gravity-induced convective turbulence on Earth.
# Results
## Flow structure {#subsec:flowstructure}
The vibroconvection setup in our study is the convective flows in an enclosure heated from below by a hot wall and cooled from above by a cold wall, and subjected to the harmonic vibration in horizontal direction. All simulations reported here are of the Navier-Stokes equations of vibroconvection under Boussinesq approximation and performed in a rectangular enclosure of aspect-ratio of \(W\):\(D\):\(H\) = \(1\):\(0.3\):\(1\) in three-dimensional cases and of \(W\):\(H\)=\(1\):\(1\) in two-dimensional cases, where \(W\), \(D\), \(H\) are respectively the width, depth and height of convection cell. Figure [\[fig:flowstructure\]](#fig:flowstructure){reference-type="ref" reference="fig:flowstructure"}a shows the typical snapshots of flow structures in vibroconvection with different dimensionless frequencies \(\omega=10^5\), \(10^6\), and \(10^7\) at fixed dimensionless amplitude \(\Am=0.01\) and fixed Prandtl number \(\Prn = 4.38\). Here, the Prandtl number, dimensionless frequency and amplitude are defined as \(\Prn = \nu/\kappa\), \(\omega = \Omega H^2/\nu\) and \(\Am = \alpha \Delta A/H\), where \(\nu\), \(\kappa\), \(\alpha\) and \(\Delta\) are respectively the kinematic viscosity, thermal diffusivity, thermal expansion coefficient and temperature difference between hot and cold plates; \(\Omega\) and \(A\) are the angular frequency and pulsating displacement of the external vibration. It is seen that the shaking by external vibration strongly destabilizes the conductive state and generates large distortion of temperature field in bulk regions by creating an artificial gravity. With increasing \(\omega\), it is vibration-induced artificial gravity that becomes strong enough to destabilize thermal boundary layer and trigger abundant thermal plumes. Those plumes are transported into bulk regions and self-organized into columnar structures. This indicates that the feature of main structures responsible for heat transport in microgravity vibroconvection is different from that in the gravity-induced RB convection. To quantitatively analyze the feature of columnar structures, we extract the instantaneous temperature field in bulk zones and calculate the power spectrum \(P(k)\) of temperature fluctuations by applying the Fourier transform in the vibrational direction as shown in Fig. [\[fig:flowstructure\]](#fig:flowstructure){reference-type="ref" reference="fig:flowstructure"}b. It is found that there exists a characteristic wave number \(k_{m}\), at which the wave number distribution function \(P(k)\) is maximal. Indeed, \(k_{m}\) characterizes the number of columnar structures in vibroconvection. We then plot the variation of \(k_{m}\) as functions of \(\omega\) in Fig. [\[fig:flowstructure\]](#fig:flowstructure){reference-type="ref" reference="fig:flowstructure"}c. It is shown that \(k_m\) monotonically increases with increasing \(\omega\), indicating that more columnar structures are formed under stronger vibrational driving force. This is consistent with the fact that larger heat transport enhancement occurs at larger \(\omega\). Further, to examine the role of thermal boundary layer in vibroconvective heat transport processes, we decompose the globally averaged thermal dissipation rate \(\epsilon_T = \kappa \lvert \nabla T \rvert^2\) into their boundary layer (BL) and bulk contributions, and then plot the variation of relative contributions as functions of \(\omega\) in Fig. [\[fig:flowstructure\]](#fig:flowstructure){reference-type="ref" reference="fig:flowstructure"}d, as suggested by the theory of Grossmann and Lohse. Here, \(T\) is the temperature field. It is seen in Fig. [\[fig:flowstructure\]](#fig:flowstructure){reference-type="ref" reference="fig:flowstructure"}d that the BL contribution of \(\epsilon_T\) is much larger than the bulk one, suggesting the BL-dominant thermal dissipation. This reveals that the dynamics of boundary layers plays a crucial role on the underlying mechanism of heat transport in vibroconvective turbulence.
## Heat transport scaling
Given that vibroconvective flow structure is different from that in gravity-induced convection, we next address the question of how the global heat transport depends on the control parameters of vibroconvection. First, we examine the dependence of heat transport on the vibration frequency. Figures [\[fig:nuscaling\]](#fig:nuscaling){reference-type="ref" reference="fig:nuscaling"}a,d show the measured \(\Nun\) as functions of vibration frequency \(\omega\) in a log-log plot for different amplitudes \(\Am\) in three-dimensional (3D) and two-dimensional (2D) cases. Here, the \(\Nun\) number, as the nondimensional ratio of the measured heat flux to the conductive one, is calculated by \(\Nun = \langle w T-\kappa \partial_z T\rangle/(\kappa \Delta/H)\), where \(w\) is the vertical velocity and \(\langle \cdot \rangle\) denotes the time and space averaging. It is observed that the \(\Nun \sim \omega^{\beta_a}\) scaling relation is not unique for different amplitudes, namely, the value of exponent \(\beta_a\) depends on the vibration amplitude. With increasing \(\Am\), the scaling exponent decreases from \(\beta_a = 1\) at \(\Am = 10^{-3}\) to \(\beta_a = 2/3\) at \(\Am = 10^{-1}\) for 3D (at \(\Am = 3\times10^{-1}\) for 2D), as shown by the dashed lines. It should be noted that the precise values of \(\beta_a\) are obtained from the physical model in subsection [2.3](#subsec:unifiedlaw){reference-type="ref" reference="subsec:unifiedlaw"}, not adjusted from the fitting with the numerical data. Further, we examine the dependency of heat transport on the two important analogous Rayleigh numbers in vibroconvective turbulence, which are the vibrational Rayleigh number \(\Ra_\mathrm{vib}\) and oscillational Rayleigh number \(\Ra_\mathrm{os}\). The first one is the vibrational Rayleigh number \(\Ra_\mathrm{vib} = (\alpha A \Omega \Delta H)^2/(2\nu\kappa)\), which is obtained from applying the averaged approach on vibroconvective equations in the limit of small amplitudes and high frequencies, and quantifies the intensity of the external vibrational source. Figures [\[fig:nuscaling\]](#fig:nuscaling){reference-type="ref" reference="fig:nuscaling"}b,e depict respectively the measured \(\Nun\) as functions of \(\Ra_\mathrm{vib}\) in a log-log plot at different amplitudes for 3D and 2D cases. We find that at small \(\Ra_\mathrm{vib}\), numerical data almost collapse together on the same scaling law, i.e., \(\Nun \sim \Ra_\mathrm{vib}^{1/2}\), as shown by the dashed lines. However, at large \(\Ra_\mathrm{vib}\), a significant departure from this scaling behavior is observed for large amplitudes. The other is the oscillational Rayleigh number \(\Ra_\mathrm{os} = \alpha A \Omega^2 \Delta H^3/(\nu\kappa)\), which is analogous to Rayleigh number in RB convection but replacing the gravitation by the vibration-induced acceleration. Figures [\[fig:nuscaling\]](#fig:nuscaling){reference-type="ref" reference="fig:nuscaling"}c,f show the variation of \(\Nun\) as functions of \(\Ra_\mathrm{os}\) for various amplitudes in 3D and 2D cases. We find that at large \(\Ra_\mathrm{os}\), numerical data almost collapse onto the same scaling relation \(\Nun \sim \Ra_\mathrm{os}^{1/3}\) as shown by the dashed line, but at small \(\Ra_\mathrm{os}\), numerical data points deviate a lot from this scaling for small amplitudes. Both \(\Nun \sim \Ra_\mathrm{vib}^{1/2}\) and \(\Nun \sim \Ra_\mathrm{os}^{1/3}\) show the independence of the cell height \(H\), but exhibit different scaling behaviors with the intensity of vibration and the temperature difference \(\Delta\) between hot and cold plates. From above, using solely the common control parameters like \(\omega\), \(\Ra_\mathrm{vib}\) or \(\Ra_\mathrm{os}\), unifying the heat transport scaling in vibroconvective turbulence can not be achieved.
## Unified constitutive law {#subsec:unifiedlaw}
There are two important questions remaining to be answered in vibroconvective turbulence: one is why there exists two different heat transport scaling laws, i.e., \(\Nun \sim \Ra_\mathrm{vib}^{1/2}\) and \(\Nun \sim \Ra_\mathrm{os}^{1/3}\); the other is whether a unified constitutive law emerges in vibroconvective turbulence. Hereafter, we propose a physical model to address both questions and uncover underlying mechanism of heat transport. As illustrated in subsection [2.1](#subsec:flowstructure){reference-type="ref" reference="subsec:flowstructure"}, we know that the BL-contribution to the global thermal dissipation rate is dominant, implying that the BL dynamics plays a crucial role in heat transport mechanism. In vibroconvective turbulence, there are two types of BL: the thermal boundary layer (TBL) with the thickness of \(\delta_\mathrm{th}\), which is estimated by \(\delta_\mathrm{th}\approx H/(2\Nun)\); the other is the oscillating boundary layer (OBL) induced by the external vibration. The modulation depth of OBL referring to \(\delta_\mathrm{os}\) is defined as the depth, at which the delaying rate of the intensity of vibration-induced shear effect equals to \(99\%\). Considering the intensity of vibrational modulation falling off exponentially from the surface, one easily obtains \(\delta_\mathrm{os} =-\ln(1-0.99) \delta_S \approx 4.605 \delta_S\) where \(\delta_S = \sqrt{2\nu/\Omega}\) is the Stokes layer thickness. First, when \(\delta_\mathrm{th} > \delta_\mathrm{os}\) as sketched in Fig. [\[fig:Unifiednuscaling\]](#fig:Unifiednuscaling){reference-type="ref" reference="fig:Unifiednuscaling"}a above, by taking into account the balance between the convective and conductive transports within TBL, the dimensional analysis of the governing equation of temperature field gives rise to \[\label{eq:temp_simeq} w \frac{\Delta}{\delta_\mathrm{th}} \sim \kappa \frac{\Delta}{\delta^2_\mathrm{th}}.\] And in the momentum equation, the balance between the vibration-induced buoyancy and the viscous dissipation leads to \[\label{eq:mom_simeq_deltaos} \alpha A\Omega^2 \Delta \sim \nu \frac{u}{\delta^2_\mathrm{os}},\] where \(u\) is the horizontal velocity. From equation [\[eq:temp_simeq\]](#eq:temp_simeq){reference-type="eqref" reference="eq:temp_simeq"}, the thickness of TBL is found to scale as \(\delta_\mathrm{th} \sim \kappa w^{-1}\). From equation [\[eq:mom_simeq_deltaos\]](#eq:mom_simeq_deltaos){reference-type="eqref" reference="eq:mom_simeq_deltaos"}, one can obtain \(u \sim \alpha A \Omega^2 \Delta \delta^2_\mathrm{os} \nu^{-1}\). Assuming that the magnitude of velocity components \(u\) and \(w\) follows a similar scaling behavior, i.e., \(u\sim w\), together with \(\delta_\mathrm{os} \sim \sqrt{\nu/\Omega}\) and \(\delta_\mathrm{th} \sim H/\Nun\), one obtains the scaling relation between \(\Nun\) and \(\Ra_\mathrm{vib}\), \[\label{eq:nu_ravib} \Nun \sim \Ra_\mathrm{vib}^{1/2} \Prn^{1/2}.\] The scaling relation in equation [\[eq:nu_ravib\]](#eq:nu_ravib){reference-type="eqref" reference="eq:nu_ravib"} shows that vibroconvective heat transport is independent of viscosity \(\nu\), but depends on thermal diffusion coefficient \(\kappa\). This implies that the dynamics of TBL is dominant to heat transport in cases of \(\delta_\mathrm{th} > \delta_\mathrm{os}\). When \(\delta_\mathrm{th} < \delta_\mathrm{os}\) as sketched in the lower panel of Fig.[\[fig:Unifiednuscaling\]](#fig:Unifiednuscaling){reference-type="ref" reference="fig:Unifiednuscaling"}a, the balance between the vibration-induced buoyancy and the viscous dissipation within TBL allows one to rewrite the momentum equation as below using dimensional analysis \[\label{eq:mom_simeq_deltath} \alpha A\Omega^2 \Delta \sim \nu \frac{u}{\delta^2_\mathrm{th}}.\] Combining equations [\[eq:temp_simeq\]](#eq:temp_simeq){reference-type="eqref" reference="eq:temp_simeq"} and [\[eq:mom_simeq_deltath\]](#eq:mom_simeq_deltath){reference-type="eqref" reference="eq:mom_simeq_deltath"}, together with \(u\sim w\), \(\delta_\mathrm{os} \sim \sqrt{\nu/\Omega}\) and \(\delta_\mathrm{th} \sim H/\Nun\), one deduces the scaling relation between \(\Nun\) and \(\Ra_\mathrm{os}\) \[\label{eq:nu_raos} \Nun \sim \Ra_\mathrm{os}^{1/3}.\] The heat transport scaling in equation [\[eq:nu_raos\]](#eq:nu_raos){reference-type="eqref" reference="eq:nu_raos"} is similar to that of RB convection in the classical regime through replacing the gravitation by vibration-induced acceleration. Both heat transport scalings predicted in equations [\[eq:nu_ravib\]](#eq:nu_ravib){reference-type="eqref" reference="eq:nu_ravib"} and [\[eq:nu_raos\]](#eq:nu_raos){reference-type="eqref" reference="eq:nu_raos"} agree well with numerical results shown in Fig. [\[fig:nuscaling\]](#fig:nuscaling){reference-type="ref" reference="fig:nuscaling"}. The competition between TBL and OBL results in the two different heat transport scaling relations, namely, \(\Nun \sim \Ra_\mathrm{vib}^{1/2}\) and \(\Nun \sim \Ra_\mathrm{os}^{1/3}\). Furthermore, we address the second question of whether the universal constitutive law of vibroconvective turbulence emerges. First, to quantify the dynamics of OBL, we define the oscillational Reynolds number \(\Ren_\mathrm{os} = \alpha \Delta A \Omega \delta_\mathrm{os}/\nu\), which is related to the vibrational velocity with the Boussinesq parameter \(\alpha \Delta A \Omega\) and the modulation depth \(\delta_\mathrm{os}\). Second, we study the dependency of \(\Nun\) on \(\Ren_\mathrm{os}\). It is intriguing to find that both \(\Nun \sim \Ra_\mathrm{vib}^{1/2}\) and \(\Nun \sim \Ra_\mathrm{os}^{1/3}\) scaling laws can be rewritten as \(\Nun \sim \Am^{-1} \Ren_\mathrm{os}^\beta\) with \(\beta=2\) for the TBL-dominant heat transport regime (\(\delta_\mathrm{th}>\delta_\mathrm{os}\)), and \(\beta=4/3\) for the OBL-dominant heat transport regime (\(\delta_\mathrm{th}<\delta_\mathrm{os}\)). Therefore, we conclude that due to the competition between the dynamics TBL and OBL on heat transport, the underlying mechanism of heat transport in vibroconvective turbulence can be categorized into two following regimes:
1. TBL-dominant regime (\(\delta_\mathrm{th} > \delta_\mathrm{os}\)): the OBL is submerged into TBL. Thermal plumes facilitated by vibration-induced strong shear detach from OBL and move into TBL. The plume dynamics is then mainly dominant by the molecular diffusion between OBL and TBL. Those plumes thermally diffuse and then self-organize into columnar structures in bulk zones, which transport heat from the bottom hot plate to the top cold one. The heat transport scaling exhibits the scaling \(\Nun \sim \Am^{-1} \Ren_\mathrm{os}^{2}\).
2. OBL-dominant regime (\(\delta_\mathrm{os} > \delta_\mathrm{th}\)): the TBL is nested into OBL. The OBL dominates the dynamics of thermal plumes ejected from TBL by vibration-induced strong shear. Between OBL and TBL, the shear effect mixes those plumes and sweep away some of them. The remaining plumes then move into bulk zones and self-organize into columnar structures. In this regime, due to the plume-sweeping mechanism between OBL and TBL, the heat transport is depleted and obeys the scaling with a smaller scaling relation exponent \(\Nun \sim \Am^{-1} \Ren_\mathrm{os}^{4/3}\).
Finally, we use the simulated data to confirm the theoretically deduced unified constitutive law. First, we plot in Fig. [\[fig:Unifiednuscaling\]](#fig:Unifiednuscaling){reference-type="ref" reference="fig:Unifiednuscaling"}b,d the variation of both \(\Am \delta_\mathrm{th}\) and \(\Am \delta_\mathrm{os}\) as functions of \(\Ren_\mathrm{os}\). It is shown that for all fixed amplitudes, the value of both \(\delta_\mathrm{th}\) and \(\delta_\mathrm{os}\) monotonically decreases as increasing \(\Ren_\mathrm{os}\), and \(\delta_\mathrm{th}\) decreases faster than \(\delta_\mathrm{os}\). The insection point between the curves of \(\Am \delta_\mathrm{th}\) and \(\Am \delta_\mathrm{os}\) divides the the plane into two regions, which corresponds to the TBL-dominant regime in the left side (\(\delta_\mathrm{th} > \delta_\mathrm{os}\)) and OBL-dominant regime in the right side (\(\delta_\mathrm{th} > \delta_\mathrm{os}\)). As depicted in the inset of Fig. [\[fig:Unifiednuscaling\]](#fig:Unifiednuscaling){reference-type="ref" reference="fig:Unifiednuscaling"}b,d, the dividing line between TBL-dominant and OBL-dominant regimes is nearly at the position of \(\delta_\mathrm{os}/\delta_\mathrm{th} = 1\). This confirms that the underlying mechanism of vibroconvective heat transport is attributed to the competition between the dynamics of TBL and OBL. Second, we plot the calculated \(\Am \Nun\) as functions of \(\Ren_\mathrm{os}\) as shown in Fig. [\[fig:Unifiednuscaling\]](#fig:Unifiednuscaling){reference-type="ref" reference="fig:Unifiednuscaling"}c,e. It is expected that all numerical data collapse together onto the derived universal constitutive law. Evidently, the numerical data and theoretical model show an excellent agreement. This confirms the emergence of universal constitutive law of vibroconvective turbulence in microgravity.
# Discussion and conclusions {#Discussion}
In summary, the innovative point of our study is the reveal of the universal constitutive law in vibroconvective turbulence, which is a new class of physical mechanism to transport heat and matter in microgravity. In the absence of gravitational acceleration, vibration creates an "artificial gravity" in microgravity to destabilize thermal boundary layer (TBL) and trigger massive eruption of thermal plumes. We find that those plumes are finally self-organized into columnar structures in bulk zones to transport heat from the bottom hot plate to the top cold one. This is different from the gravity-induced convection, like Rayleigh-Bénard convection, in which large-scale circulation is formed in bulk and dominates heat transport. By analyzing the basic properties of heat transport, we find at small vibration amplitudes, the heat transport exhibits a power-low relation with the vibrational Rayleigh number \(\Ra_\mathrm{vib}\), i.e., \(\Nun \sim \Ra_\mathrm{vib}^{1/2}\); at large amplitudes, the heat transport scaling can be well described by the oscillational Rayleigh number \(\Ra_\mathrm{os}\), i.e., \(\Nun \sim \Ra_\mathrm{os}^{1/3}\). Both \(\Nun\)-relations shows that the global heat flux is independent of the cell height. However, vibroconvective heat transport exhibits different scaling trends with the intensity of vibration and the temperature difference between bottom and top plates, indicating that the underlying mechanism of the two \(\Nun\)-relations are completely different. We also find that the BL-contribution is dominant to the global thermal dissipation rate, implying that the dynamics of boundary layer plays an essential role in vibroconvective heat transport. We then propose a physical model to theoretically deduce both \(\Nun\)-scaling-relations, and explain the distinct properties of viboconvective heat transport, based on the competition between the thermal boundary layer (TBL) and oscillating boundary layer (OBL) induced by the external vibration. To look for the universal features, we define the oscillational Reynolds number \(\Ren_\mathrm{os}\) quantifying the dynamics of OBL, and study the dependency of heat transport on \(\Ren_\mathrm{os}\). Both theoretical results and numerical data shows the emergence of universal constitutive law in vibroconvective turbulence, i.e., \(\Nun \sim \Am^{-1} \Ren_\mathrm{os}^{\beta}\), where \(\beta\) is the universal scaling exponent. We also find that the exponent \(\beta\) is determined by the relative importance between the dynamics of TBL and of OBL to heat transport, and identify \(\beta = 2\) in TBL-dominant regime and \(\beta = 4/3\) in OBL-dominant regime. It is concluded that the type of vibroconvective turbulence in microgravity owns a universal constitutive law with its underlying heat transport mechanism different from that in gravity-induced convective turbulence. The emergence of universal constitutive laws in vibroconvective turbulence provides a powerful basis on generating a controllable heat transport under microgravity conditions. In addition, in microfluidic environment where the effect of gravity is nearly absent, vibroconvective turbulence can provide alternative to effective heat and mass transport in a controllable manner, which surpasses the pure diffusive transport. Looking forward, we expect that our findings can be tested by experimental realizations in space station.
# Methods
We consider the coupled equations of motion for the velocity field \(\boldsymbol{u}\) and the temperature field \(T\) in turbulent vibroconvection under the Boussinesq approximation under microgravity conditions. Horizontal harmonic vibration \(A \cos(\Omega t)\) is applied on the vibroconvection system to generate heat transport. In the non-inertial frame associated to the imposed vibration, an inertial acceleration of \(A\Omega^2 \cos(\Omega t) \boldsymbol{e}_x\) is added to the system, where \(\boldsymbol{e}_x\) is the unit vector in \(y\)-direction. The governing equations for vibroconvective turbulence is then can be writen as \[\begin{aligned}
& \nabla \cdot \boldsymbol{u}=0, \\ &\partial_t \boldsymbol{u} + (\boldsymbol{u} \cdot \nabla) \boldsymbol{u} =-{\nabla}{p} + \nu \nabla^2 \boldsymbol{u}-\alpha A \Omega^2 \cos(\Omega t) T \boldsymbol{e}_{x}, \\ &\partial_t T + (\boldsymbol{u} \cdot \nabla) T =\kappa \nabla^2 T, \end{aligned}\] where \(p\) is the kinematic pressure field. All quantities studied above have been made dimensionless with respect to the cell's height \(H\), the temperature difference across the fluid layer \(\Delta\), and the viscous diffusion velocity \(\nu/H\). Based on these choices, the relevant control parameters for the vibroconvection system are the dimensionless vibration amplitdue \(\Am\), the dimensionless vibration frequency \(\omega\), and the Prandtl number \(\Prn\). The governing equations are numerically solved by a second-order finite difference code, which has been validated many times in the literature. At all solid boundaries, no-slip boundary conditions are applied for the velocity. At top and bottom plates, constant temperatures \(\theta_\mathrm{top}=0\) and \(\theta_\mathrm{bot}=1\) are given; and at all side walls, the adiabatic conditions are adopted. We performed a series of direct numerical simulations of microgravity vibroconvective turbulence over the vibration amplitude range \(0.001 \le \Am \le 0.1\) and the frequency range \(10^5 \le \omega \le 10^7\) for 3D cases, and over the vibration amplitude range \(0.001 \le \Am \le 0.3\) and the frequency range \(10^3 \le \omega \le 10^7\) for 2D cases at fixed Prandtl number \(\Prn = 4.38\). For all simulations, the computational mesh size is chosen to adequately resolve the dynamics both the TBL and OBL, and the time step is chosen to not only fulfil the Courant-Friedrichs-Lewy (CFL) conditions, but also resolve the time scale of one percent of the vibration period. Further details of numerical parameters are given in the Supplementary Materials.
# Author contributions {#author-contributions .unnumbered}
J.-Z. Wu performed the 3D simulations and wrote the paper. X.-L. Guo performed the 2D simulations. J.-Z. Wu, C.-B. Zhao, B.-F. Wang, K. L. Chong, and Q. Zhou analyzed and interpreted the data. B.-F. Wang, K.L. Chong, and Q. Zhou supervised the project revised the paper. | {'timestamp': '2022-12-19T02:11:28', 'yymm': '2212', 'arxiv_id': '2212.08461', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08461'} |
null | null |
null | null |
# Introduction
Extended reality (XR) is changing the way that humans interact with the virtual world and bringing the revolution to immersive gaming, videoconferencing, and remote shopping. XR technologies can be categorized into augmented reality (AR), mixed reality (MR), and virtual reality (VR), dependent on the portions of virtual contents to be rendered and displayed . The rendered virtual contents are usually displayed by XR head-mounted devices (HMDs). As the users have to wear these devices sometimes for a long time, XR HMDs have strict restrictions on their weight, energy consumption, and heat dissipation. To meet the restrictions, HMDs have to offload most of the storage and computing tasks to remote servers, which requires ultra low latency connections between HMDs and remote servers. Recently, high-quality wireless ultimate XR has been developed . As HMDs are connected to remote servers through wireless networks, users can move freely with better quality-of-experience (QoE). However, wireless ultimate XR still faces some critical challenges. First, ultimate XR requires the motion-to-photo latency to be less than \(10\) ms , which means the corresponding changes caused by users' controlling commands or the eye/head/foot movements should be presented to the users within \(10\) ms. However, latency in nearly all components in a wireless XR system, including sensing, rendering, and communication parts makes the ultra-low latency difficult to achieve when generating and transmitting virtual contents with high resolutions. Secondly, the data transmission rate is expected to exceed tens of Gbps in ultimate XR , which is far above the achievable capacity of the existing wireless networks with peak from 0.1 Gbps to 2.0 Gbps . Last but not least, the high power consumption caused by data processing and transmission also limits the lifetime of HMDs. For example, in ultimate VR, uncompressed bit rates can reach up to 1.9 Tbps for full-view videos and 1.3 Tbps for single-view videos, respectively . Even if a 300:1 video lossy compression technique is adopted, the reduced data rates are still very high (6.37 Gbps for full-view and 4.43 for single-view), which are beyond the achievable capacity of existing wireless networks, where peak data rates have only reached around 0.1 to 2.0 Gbps from recent network measurements . Therefore, a new communication network architecture for supporting wireless ultimate XR is more than desired. To overcome the ultra-high data rate challenges in wireless ultimate XR, semantic communication is considered as a promising solution. Different from conventional communications, semantic communications only transmit semantic information relevant to the task at the receiver, significantly reducing the volume of data traffic. Recent studies have shown that semantic communication systems can greatly improve the transmission efficiency in serving various applications, such as wireless image retrieval , video conferencing , and object detection . In wireless ultimate XR, large amounts of sensed data and rendered results will be transmitted between XR HMDs and the remote server for computing tasks and view synthesis tasks. By exploring and extracting the task-relevant semantic information from the sensed data and rendered results, the size of data to be transmitted can be greatly reduced Despite the attractive performance of semantic communications, they are insufficient for satisfying the latency and power consumption restrictions in wireless ultimate XR systems. In ultimate XR systems, high-resolution images/videos are frequently captured or rendered. Generating and processing these high-resolution sensed or rendered images/videos introduce significant latency and heavy power consumption to the system. For example, 360\(^{\circ}\) stereo images are required for precise object insertion in AR/MR applications . Color and depth images in \(2\)K resolution are captured for 3D reconstruction . Full-view \(24\)K video need to be rendered to provide VR users immersive perceptual experience . Therefore, it is imperative to design fast and energy-efficient sensing and rendering techniques to put wireless ultimate XR into practice. In this article, we propose a framework, including semantic sensing, semantic rendering, and semantic communications, to support wireless ultimate XR. The proposed framework can alleviate the sensing costs at HMDs, fasten the rendering process at the server, and reduce the data traffic between HMDs and the server. The rest of this article is organized as follows. Section [2](#section II){reference-type="ref" reference="section II"} discusses the general architecture and challenges of wireless ultimate XR. In Section [3](#section III){reference-type="ref" reference="section III"}, we demonstrate the proposed framework with semantic sensing, semantic rendering, and semantic communications, followed by user case studies in Section [4](#section IV){reference-type="ref" reference="section IV"}. Section [5](#IIII){reference-type="ref" reference="IIII"} concludes this article with potential research directions.
# Wireless Ultimate XR Architectures {#section II}
The procedures for wireless XR applications usually consists of sensing, transmission, computing, rendering, and displaying, as shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. To facilitate the introduction of our semantic based framework, we will provide a brief description on the wireless ultimate XR.
## Sensor Data Acquisition
To map objects and scenes in user's environments into the virtual display, some sensors are required at the user side, including microphones, cameras, and depth sensors. The amount of sensed data generated by all types of sensors can be quite large. For example, in ultimate VR video conferencing, high-resolution images of the users should be captured, helping the remote servers to add body, facial expressions, and dressing details to the users' virtual images. In ultimate AR shopping, high-resolution color and depth images of indoor environment are required by the servers, so that the servers can estimate the three-dimensional (3D) geometric information of the indoor environment, and display the virtual products correctly with reasonable shadows and occlusion . Generating and processing a large amount of sensor data is energy-and time-consuming for HMDs, especially when these sensory data needs to be further compressed and transmitted. Therefore, it is imperative to design a faster and power-efficient sensing system.
## Sensed Data Processing and Transmission
After collecting the sensed raw data, data aggregation, coding, and compression will be performed at HMDs, before they are uploaded to the remote servers for post-processing. However, the stringent latency restriction of ultimate XR applications poses an extremely high requirement on the wireless transmission rates. For example, if the sensed data is an image with \(2\)K resolution and is encoded by Better Portable Graphics (BPG) with \(0.8\) bit per pixel (bpp), the compressed sensed data size will be of \(1.6\) Mb. Considering the sensed image will affect the virtual contents, this transmission follows \(10\) ms motion-to-photo requirement. If we further consider the time spent on rendering, downlink transmission, and displaying, the time budget for uplink transmission will be less than \(1.25\) ms, which requires the data rate of about \(1.3\) Gbps. In the ultimate XR, 360\(^{\circ}\) stereo images and depth images will be used, which results in a large volume of sensed data and requires higher data transmission rate.
## Computing
Once the sensed data is received by the remote servers, a series of computing tasks will be conducted to extract useful information from these sensor data, which are used to construct or update various 3D models of objects and scenes used in virtual worlds. For example, hand pose estimation and tracking are widely used in VR gaming and conferencing. In this case, hand-related information will be extracted from the sensed data, and then used to construct a 3D hand model. As only task-related information in sensory data will be utilized at the server, it inspires us to reduce the transmitted data size in uplink connections through semantic communications. For example, after generating coarse 3D human meshes, more textures or details will be added to these 3D meshes to make them look real, such as appearances, facial expressions, and clothes. These edited 3D human meshes, alongside with the pre-stored 3D scenes, will be projected to a 2D image using ray-tracing techniques, where the camera rays are generated according to the viewpoint of users. During this process, blurry and shadow effects will also be applied to simulate the focus of cameras and light sources. To generate high-quality rendered results, rendering process will task a long time and consume a great amount of server resources. More efficient rendering techniques should be proposed.
## Rendered Results Generation and Transmission
The rendered images can be either perspective images or 360\(^{\circ}\) images. Perspective images, which only contain the visual information equal to or slightly larger than the FoV of users, are widely rendered and transmitted in VR applications due to their small data volume. However, this transmission method is not suitable for VR applications where users will have large eye/head movements. By contrast, transmitting 360\(^{\circ}\) images can deal with the FoV change problems caused by eye/head movements at the cost of \(3\times\) data rate growth. While, even if 360\(^{\circ}\) images are rendered/transmitted, a new image will still need to be re-rendered if the standing points of users in digital worlds change. In AR/MR applications, instead of transmitting 2D images, rendered 3D models of virtual objects will be transmitted, which can be displayed on AR/MR glasses. After obtaining the 3D models or scenes through computation, rendering will generate photo-realistic viewpoint-dependent perspective images or 360\(^{\circ}\) panoramic images for users to perceive 3D scenes. Viewpoint-dependent images only contain the visual contents at a specific viewing direction, which, however, are not suitable for VR applications where users' heads move frequently. By contrast, rendering 360\(^{\circ}\) images can deal with this problem with the cost of three times data rate. In ultimate AR/MR applications, rendering means adding light or shadow effects to 3D models to make it natural in real-world environment based on the computing results. Similar to the sensed data transmission, transmitting high-resolution rendered images or 3D models of objects and scenes under a stringent latency requirement is challenging. As stated in , full-view \(24\)K images need to be transmitted within about \(2\)-\(4\) ms for downlink transmission in ultimate VR applications, which requires the data rate of more than \(50\) Gbps. A medium-size 3D model stored in Polygon File Format will exceed \(110\) Mb, resulting in the \(30\) Gbps data rate requirement in AR applications.
## Displaying
The last step is to display these rendered virtual contents on HMDs for users. Sometimes, local processing is also required to ensure the smooth displaying. For instance, 3D world locking is required when displaying 3D models of virtual objects in AR/MR applications, so that the placement of these virtual objects would not change with users' movements. Meanwhile, view synthesis techniques are also required when users' fields of view (FoV) are different from what are expected in the rendering process. Therefore, displaying will also cost certain time and computational resources.
# Semantic sensing, rendering, and communications {#section III}
From section [3](#section III){reference-type="ref" reference="section III"}, it is necessary to introduce semantic framework to address various challenges on implementing wireless ultimate XR. In this section, we will introduce the proposed semantic sensing, rendering, and communication architecture for wireless ultimate XR. An overall architecture is shown in Fig [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}.
## Semantic Sensing
Processing a large amount of sensed data is challenging when HMDs have limited storage resources, energy capacity, and computing performances. Therefore, reducing the amount of data acquired at the sensor level should be considered. Recently, image or video compressed sensing has drawn considerable attention as an effective framework to alleviate the computational and sampling costs of sensors . Different from the traditional sensing-then-compression pipeline, these works propose to conduct image or video sensing and compression simultaneously using compressed sensing, leading to a smaller sensory data size. The existing studies motivate us to rethink the sensing-then-task-execution framework, where plenty of task-irrelevant information will be also sensed, but discarded during the task-execution process.
### Design Principles
To improve the sampling efficiency, we propose the idea of semantic sensing, where only task-related signals will be extracted. Semantic sensing can be divided into two categories: prior-based sensing and feedback-based sensing. In prior-based sensing, the statistical distribution of task-related information is summarised from a task-related dataset, and then used to design a spatial-temporal 3D binary mask. Subsequently, the 3D binary mask is compiled into a programmable sensor and used to selectively switch off some sensing elements in spatial-temporal dimensions, where the samples are semantically-redundant for a specific task, as shown in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. We now illustrate the idea of semantic sensing in more details by taking hand pose estimation and tracking as an example. Usually, the actions of user hands follow some fixed patterns in VR driving or shooting games. Studying these patterns can help us summarize the activity area and moving speed of user hands. Based on these prior information, we can selectively turn off some camera sensor elements in the spatial-temporal dimensions to reduce the sampling rates without affecting task accuracy. For the feedback-based sensing, the historical sensed results are used to infer the distribution of task-related samples, and guide the sensing process at the current timeslot. Specifically, when the locations of the interesting objects are unknown in captured images/videos, we can first collect all samples. After applying an object detection algorithm at the server side, the bounding box of the interested objects will be fed back to the sensors, which can then be used to reduce the number of samples to be taken.
### Architectural Details
To implement semantic sensing, a mask generation network will be deployed at the HMDs, which is used to generate task-specific 3D binary masks by taking task tokens as inputs. For feedback-based sensing, this neural network will also take the estimated areas of interesting objects and historical task performance as inputs. The generated mask will dynamically adjust the sensing policy for each task, and then the sensed data will be fed into a task execution network. The task accuracy will be used to guide the training process of the mask generation network with the goal of maximizing the sparsity level of the binary mask under a task accuracy constraint. The semantic sensing module should also be jointly designed with the semantic communication modules, so that the latter can handle the changes of sensing policy.
## Semantic Rendering
Rendering is a time-consuming task. A fast rendering process should be designed to help the system satisfy the latency requirements. Inspired by images/videos compressed sensing , we find that existing rendering-then-compression pipeline is also sub-optimal, where each pixel in the rendered images/videos is computed with great efforts, but some pixels will later be discarded in the compression process because of inter-pixel semantic redundancy. Therefore, it will be more efficient if we can jointly consider the rendering process and the compression process. Different from the sensing process, where the real environment is unknown before sensed, the rendering process can be guided by pre-stored 3D models of objects or scenes.
### Design Principles
Semantic rendering is proposed to dynamically adjust the rendering costs for each pixel in rendered images and 3D models of objects/scenes according to their semantic importance. To make it clear, we will explain the semantic rendering by the widely used volumetric neural render, NeRF . In NeRF, each 3D scene is first represented by a fully-connected deep network. Then, for a given viewing direction, a certain number of camera-rays will emit from a virtual camera. If the expected resolution of rendered images is \(H\times W\), then there will be \(H\times W\) rays in total. Each ray represents one pixel here. To get the color value of each pixel, NeRF will divide the corresponding ray into \(N\) evenly spaced bins, and draw a sample uniformly from each bin. Then the deep network will generate an emited color value and a density value for each sample. The final color is calculated from the color/density values of the \(N\) samples. After processing each ray, a rendered image is generated. Therefore, \(H\times W\times N\) samples will be passed through the deep network during the rendering process. Considering \(N=192\) used in and the spatial sizes of high-resolution images, this rendering process will cause significant latency. In the semantic rendering, the important pixels at the semantic level will use more samples to estimate the corresponding color values. To achieve this goal, we propose to learn a 3D binary mask for the \(H\times W\times N\) samples, and only the samples with +1 values in the mask will be actually used for color value calculation, as shown in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. This setting will fasten the rendering speed at the cost of a spatially-varying rendering accuracy. At the same time, a image restoration network will be jointly leaned with the 3D binary mask, which will recover the discarded information during the rendering process by exploring the inter-pixel semantic redundancy. The image quality after the restoration network will be used to guide the design of the 3D binary mask with the goal of maximizing the sparsity level under an image quality constraint.
### Architectural Details
To implement semantic rendering, a mask generation network will be deployed at the server, which is usually designed for each 3D model of scene/object. The best 3D binary masks for the rendering process is dependent on the viewing directions. To get the 3D binary masks for different viewing directions, the mask generation network takes the viewing directions as inputs, and outputs the mask details. As mentioned above, this mask-generation network will be trained with the goal of finding the best trade-off between reconstructed images' or 3D objects' qualities and the sparsity level of the binary masks. Similarly, the semantic rendering module should also be jointly designed with the semantic communication modules, so that the latter can deal with different rendering polices.
## Semantic Communications
In wireless ultimate XR, sensor data transmission and rendered results transmission pose new challenges to the existing wireless communication systems, where a great amount of data should be transmitted in a few milliseconds. To address this issue, we can use semantic communication for the transmission of the sensed data and rendered results. **Motivations:** To increase the transmission efficiency, one possible way is to change the design goals of communication systems from transmitting the source data as accurately as possible to transmitting semantically-important information only, so that the data to be transmitted can be reduced without deteriorating task execution accuracy or human perception experience. Semantic communications have been widely used in various applications but the studies in ultimate XR are still underdeveloped. **Design principles:**
### Design Principles
In the uplinks of wireless ultimate XR, sensed data is usually transmitted for executing some computing tasks at the server side. Since only parts of information inside the sensed data are actually used during the task-execution process, semantic communications can be used to reduce the data traffic by extracting task-related information from the sensed data. For example, in the VR conferencing, transmitting information that represents users' body shapes, gesture, facial expressions, and dressings are enough for remote servers to update users' images in virtual conferencing room, which consumes much less wireless resource than transmitting the original camera-captured images. This idea can be extended to other computing tasks, such as hand pose estimation, light source estimation, and object detection. For example, considering the VR conferencing application, where an external camera at the user side keeps capturing and sending the real-time videos of one attendee, and a remote server keeps using these videos to update the attendee's virtual image in a virtual conferencing room, the semantic encoder deployed at the user side can be chosen as a detection network that can detect human's body shape, gesture, and facial expression. At the same time, a semantic decoder that can use these shape/gesture/facial expression information to reconstruct a coarse 3D mesh of that person will be placed at the server side. In this way, instead of transmitting the captured videos, only the information related to the human mesh reconstruction task is transmitted through the uplink connections of wireless ultimate XR, significantly reducing the data rates requirements. Semantic communications can also be used to improve the data transmission efficiency of the downlink connections in wireless ultimate XR. Generally speaking, the rendered images or 3D models of objects/scenes at adjacent time stamps are quite similar with minor changes caused by users' movements or virtual world changes. Therefore, instead of transmitting a new image or 3D model of object/scene, transmitting the differences between time slots is more efficient. Note that different from the video transmission methods, the differences here are the changes of semantic features, which is much less than the original representations. Besides, in XR applications, multiple users may perceive the same virtual world from different viewing angles. In this case, the virtual contents delivered to different users are semantically-similar. By sharing the semantically-similar information on the broadcast channel and delivering semantically-different information to each user on per-user specific channel, we can reduce the overall downlink data size.
### Architectural Details
A semantic-aware communication system is composed of a semantic encoding module, which is used to extract semantic information from the source data to transmit, and a semantic decoding module that uses the reconstructed semantic information for task execution or human perception, as shown in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. The architectural details of the semantic encoding/decoding modules usually depend on the type of source data and specific task. In particular, the semantic coding modules for image-type sensed data/rendered results transmission are usually designed based on convolutional neural networks (CNNs) or visual transformers (VTs), which have shown excellent performance in extracting semantic features from images. For the transmission of 3D models of objects/scenes, which has not attracted wide attentions, graph neural networks (GNNs) are good choices, as the points clouds of 3D objects or scenes can be represented by graphs. In addition to using GNNs, implicitly representing the 3D models of objects/scenes by a neural network , and semantically encoding and decoding the network parameters is another promising direction. Besides semantic coding modules, source coding and channel coding modules are still required to ensure efficient delivery of the semantic information over noisy wireless channels, as shown in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. The source coding module aims to find a more compact representation of the semantic information statistically. This is usually achieved through a variable-length coding mechanism, based on the criterion that a shorter code shall be assigned to semantic information more likely to appear statistically. A practical source coding method for semantic representations is the deep compression methods proposed in recent years . The channel coding module is designed to add some redundancy to the coded symbols to help them conquer the channel noises. A widely used channel coding method is low-density parity-check (LDPC) code. In some works , a capacity-achieving channel code will be considered rather than the LDPC code, which can be used to represent the performance limit of any achievable digital channel coding method. As the interfaces between deep compression methods (served as source codes) and channel codes are digital, the channel coding module has to be designed separately from the semantic coding modules and deep compression module, resulting in a separation-based semantic communication system (SSC). Recently, deep joint semantic-channel coding (DeepSC) has been widely adopted in the recent semantic communication works, instead of using separated source/channel coding modules. These works use deep neural networks to represent the source coding and channel coding together without explicitly differentiating the source coding and channel coding. After choosing the network architectures, a training loss will be used to guide the training process of the semantic-source coding networks in separation-based methods and the deep joint semantic and channel coding networks. In semantic communications, the choices of the training loss are task-dependent. For the sensed data transmission, task accuracy will be used as the training loss. As for the rendered results transmission, objective perceptual losses or subjective evaluations are exploited. Specifically, a learned perceptual image patch similarity (LPIPS) loss combined with a generative adversarial network (GAN) loss can be used as a perceptual loss of images . GAN losses are also used in human mesh reconstruction for better visual quality .
## Joint designs
To get the optimal performance for the system models, the semantic sensing and semantic communications used in the uplinks should be jointly designed. Similary, the semantic rendering and semantic communications should be jointly considered for the downlinks. Different from the natural images/videos/3D models used as the source data in semantic communications, the source data generated after semantic sensing and rendering always have spatially-varying data distributions, which should be considered in the semantic encoding processes. Also, the semantic decoding module should consider the influence of semantic sensing/rendering on task executions/human perceptions, as some task-related information may be discarded due to the 3D binary masks. The working mechanisms of semantic coding modules can also in turn guide the generation process of binary masks in semantic sensing/rendering. Usually, the semantic coding modules depict how the task-related information will be extracted from the source data, based on which the 3D binary masks can learn to discard semantically-unimportant information and keep the others.
# Case Studies {#section IV}
In this section, we will demonstrate how the proposed semantic framework can be used to overcome the data rate challenges in wireless ultimate XR. In particular, we focus on the sensed data transmission in the uplink for 3D human mesh recovery task, and rendered image transmission in the downlink.
## Uplink 3D Human Mesh Transmission
3D human mesh reconstruction is an important task in ultimate VR conferencing and gaming, where the 3D virtual images of users should be consistent with their poses and body shapes in reality. To realize the synchronous update between real images and virtual images, a camera sensor shall be deployed at the user side, keep capturing the real images of users, and transmit the camera-captured images to remote servers. The server will then estimate the user's body shape and pose from the received images, and reconstruct the 3D human mesh. In this case study, we simulate the wireless ultimate XR with the proposed framework, and compare its performance with the case with conventional communications under different channel conditions. In the traditional communications (TC) shown in Fig. [\[subfig:tradition\]](#subfig:tradition){reference-type="ref" reference="subfig:tradition"}, the color images are first bicubically-downsampled, and then compressed using image compression method BPG. The compressed data file is then transmitted through a capacity-achieving channel code. At the server side, the original RGB image is reconstructed in a reverse way. After that, the server constructs the 3D human mesh using a 3D human mesh reconstruction network, HMR . In semantic communications, the 3D human body shape and pose information are extracted and transmitted, rather than the original color image. As shown in Fig. [\[subfig:semantic\]](#subfig:semantic){reference-type="ref" reference="subfig:semantic"}, the semantic encoder consists of a ResNet50 and a regressor . The ResNet50 is used to extract the high-level convolutional features from the RGB input. The regressor then maps the features to a smaller feature vector of size \(85\). These \(85\) features are composed of camera angle information (\(3\) features), body shape information (\(10\) features), and body pose information (\(72\) features). After obtaining the \(85\) semantic features, we consider both the SSC and DeepSC methods for transmitting these features. In the SSC method, deep compression is used as the source coding module . Specifically, a FCN is used to map the semantic features to a compact latent space with \(10\) dimensions, and then entropy coding methods are used to encode features in the latent space. After the source coding modules, capacity-achieving channel codes are considered. At the receiver side, the \(85\) semantic features are first reconstructed, which is then applied to predict the 3D human mesh through a SMPL model . In the DeepSC method, we use a 5-layer FCN with \(1024\) hidden neurons to map the semantic features to modulated symbols, whose number is set as two times the channel bandwidth . After receiving the symbols, a 5-layer FCN is used to recover the semantic features, which are then passed into a SMPL model. We use the MPI-INF-3DHP dataset and the losses developed in to train the neural networks used in the above methods. We set the bandwidth as \(1\) kHz, and change the signal-to-noise-ratio (SNR) of transmitted symbols over additive white Gaussian noise (AWGN) channels from-20dB to 30dB. We use mean per joint position error (MPJPE) as the evaluation metric, where a smaller MPJPE value represents a better performance. We compare their performance in Fig [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}, where HMR represents the performance of human mesh reconstruction network developped in when there is no noise. From Fig [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}, semantic communication systems work better than the traditional method, denoting the efficiency of the proposed framework.
## Downlink VR Content Delivery: Broadcast
In this case study, we use broadcasting transmission method alongside with the 3D-aware view synthesizing techniques to reduce the transmission data rates. Specifically, multi-users may watch the same VR contents from different places in the virtual world. For example, many attendees at different locations may watch someone presenting the works in VR conferencing. The 3D models of the room and the attendees are fixed during this presentation period. Therefore, instead of rendering and transmitting a specific view for each attendee, we semantically encode the 3D scene, broadcast the semantic representations of the scene, and allow users to synthesize their views locally using the broadcast information. In this way, multiple users can be served simultaneously with the shared wireless resource. We consider the situations where multiple users sit/stand closely in the virtual world and have very close viewing directions. In this case, it is possible to synthesize multiple views by transmitting a single view of the scene. As shown in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}, the VR content provider broadcasts a single-view image to a group of users. The developed 3D-aware synthesizing model enables the group to synthesize the required content based on their own viewpoints. Considering the increasing privacy concern and local device computing resources, a federated learning algorithm is proposed to enable the practical training of the view-synthesizing model. We also divide the models into two parts and each client is only responsible for the model update of one part. To be more exact, the volumetric rendering and color prediction parts are updated by some clients. The viewpoint analysis and camera control parts of the model are updated by the others. With this task division, the computing cost of each user can be significantly reduced. The delivered VR content to each user is rendered with required viewports of an observed object, which can be modeled as single-view images. A virtual camera shoots rays to the samples, with the pitch and yaw sampled uniformly at random over the sphere of radius 3.5 and field-of-view of \(\pi/4\). The horizontal datasets and vertical datasets are utilized with human face images from different countries and human face images from specific countries. The position and viewing direction are with 10 Fourier feature dimensions. The models are trained progressively and 32 images in each communication round. The optimization is performed by Adam with a learning rate of 0.002, and discriminators are trained with R1 regularization with 0.05 weights. In this case, the users requesting a specific range of FoV can be served simultaneously, which reduces the bandwidth cost and transmission latency. The multi-view synthesizing model also benefits the required data volume by sharing the single-view image rather than sending meshes of whole \(360^{\circ}\) videos at once. It also reduces the dependency on the viewport prediction accuracy compared to the tile-based schemes. Furthermore, training a VR-related model distributed is a critical issue for its huge amount of data. The proposed federated learning scheme with horizontal and vertical datasets can reduce the latency in the federated learning process. In each communication round, the overall data volume is the same as the traditional federated learning. The reduced communication overhead for individual clients motivates more users to contribute to the model. With increased diversity from participation, federated learning performance can be improved. The users with horizontal datasets contribute the majority of the layers for less bias than users with vertical datasets.
## A complete VR contents delivery
In this case study, we consider a complete VR content delivery process under the proposed semantic sensing, rendering and communication architecture. We consider the situation that a user is visiting a digital world by walking and looking around. In the proposed architecture, a semantic sensing module will activate some sensing elements in the inertial or external programmable sensors deployed at the user side. The sensed results will then be fed into a semantic communication module, which will extract and send the semantic features representing the user's facial expression, body pose, and viewpoint. After receiving these features, the server will use the user's facial expression and body shape information to update her virtual image. At the same time, the viewpoint features will be sent to a semantic render for view synthesis. The semantic render will take as input current and past viewpoint features of the users, and output a binary mask denoting the features required by the users for current view. These features will then be acquired from a global 3D model of the world by the render and be sent to the user through a semantic communication module. At last, the VR device use these new features and previous received features to render the current view for user.
# Conclusions and Future Directions {#IIII}
In this article, we developed the framework with semantic sensing, rendering, and communication to support wireless ultimate XR. To bring the semantic-aware wireless ultimate XR from theory to practice, the following challenges should be addressed: **1) Trade-off between performance and computing costs:** Existing semantic coding modules have shown great success in data size reduction. However, running semantic coding modules on HMDs usually takes a long time due to the limited on-board computing resources. To reduce the running time, one potential way is to consider model distillation to find the best trade-off between the model performance and computing costs. **2) Task-specific network generation:** Most of the existing semantic sensing or communication modules are designed in a task-specific way. This will result in a great amount of semantic models to be stored at HMDs, which usually have limited storage space. One potential way to reduce the memory costs is to use meta-learning to train a network-generation network, which dynamically generates the optimal network architectures and parameters for each task in wireless ultimate XR. **3) Semantic-aware resource allocation:** In a practical communication systems, users of a particular XR application need to share wireless resource with others. The data generated by each XR user also has various size and different semantic-importance level. Therefore, it is challenging to find the optimal resource allocation policy in this complicated situation. To maximize the resource utilization efficiency, deep reinforcement learning could be considered. | {'timestamp': '2022-12-19T02:13:45', 'yymm': '2212', 'arxiv_id': '2212.08533', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08533'} |
# Abstract {#abstract .unnumbered}
**Compact stellar objects are promising cosmic laboratories to test the nature of dark matter (DM). DM captured by the strong gravitational field of these stellar remnants transfers kinetic energy to the star during the collision. This can have various effects such as anomalous heating of old compact stars. The proper calculation of the DM capture rate is key to derive bounds on DM interactions in any scenario involving DM accretion in a star. We improve former calculations, which rely on approximations, for both white dwarfs (WDs) and neutron stars (NSs). We account for the stellar structure, gravitational focusing, relativistic kinematics, Pauli blocking, realistic form factors, and strong interactions (NSs). Considering DM capture by scattering off either ions or degenerate electrons in WDs, we show that old WDs in DM-rich environments could probe the elusive sub-GeV mass regime for both DM-nucleon and DM-electron scattering. In NSs, DM can be captured via collisions with strongly interacting baryons or relativistic leptons. We project the NS sensitivity to DM-nucleon and DM-lepton scattering cross sections which greatly exceeds that of direct detection experiments, especially for low mass DM.**
# Introduction {#sec:intro}
Direct detection (DD) experiments lead the quest to unveil the particle nature of dark matter (DM). In recent years, they have seen an impressive increase in sensitivity, especially to spin-independent (SI) interactions. However, their reach is limited by the achievable mass of the target material and the recoil energy threshold. In addition, DD experiments are less sensitive to spin-dependent (SD) and DM-electron cross sections. It is then natural to look for alternative systems in which DM interactions lead to observable consequences. In this sense, DM capture in the Sun has long been used as an indirect detection technique. If DM couples to visible matter, it will scatter with the Sun constituents. Provided that DM loses enough energy in the collision, it becomes gravitationally bound to the star. Accreted DM can be detected via its annihilation to neutrinos that escape the Sun . Because of their high density that will result in more efficient DM capture, compact stars were identified long ago as potential DM probes. It was recently pointed out that DM capture could transfer enough kinetic energy to heat old, isolated neutron stars (NSs) in the solar neighbourhood to infrared temperatures. In light of this, in a series of papers we improved former calculations of the DM capture rate, which rely on simplifying assumptions, in both white dwarfs (WDs) and NSs . We accounted for the stellar structure, gravitational focusing, a fully relativistic treatment of the scattering process, the star opacity, Pauli blocking (for degenerate targets), nuclear (WDs) and nucleon (NSs) form factors, and strong interactions (for baryonic targets in NSs). Using observations of old WDs in the globular cluster Messier 4 (M4) , which we assumed to be formed in a DM subhalo, we derive bounds on DM-nucleon and DM-electron scattering cross sections. For NSs, we provide sensitivity projections to DM-nucleon and DM-lepton interactions, which surpass that of DD experiments especially for light DM. This paper is structured as follows. In section [2](#sec:compactstars){reference-type="ref" reference="sec:compactstars"}, we briefly summarise the internal structure of compact stars. In section [3](#sec:capture){reference-type="ref" reference="sec:capture"}, we outline the capture rate calculation in both WDs and NSs. Our results are presented in section [4](#sec:results){reference-type="ref" reference="sec:results"} and concluding remarks in section [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}.
# Compact Stars {#sec:compactstars}
The fate of a star is determined by its mass when it enters the main sequence. Main sequence stars with masses below \(\sim 8-10 \, M_\odot\) end up their life cycles as WDs. More massive stars have a more spectacular end, a core-collapse supernova explosion that leaves behind a proto NS.
## White Dwarfs {#sec:WDs}
WD progenitors are low and intermediate mass stars, therefore WDs are the most abundant stellar remnants. Moreover, WD physics is far more constrained than that of NSs. E.g., there is much less uncertainty in their equation of state (EoS), and their luminosity-age relation is better understood. WDs are supported against gravitational collapse by electron degeneracy pressure. Most of them are composed mainly of carbon and oxygen. To solve the WD structure equations, we coupled the relativistic Feynman-Metropolis-Teller EoS with the Tolman-Oppenheimer-Volkoff (TOV) equations (hydrostatic equilibrium in general relativity), and obtained the WD mass \(M_\star\), radius \(R_\star\), as well as radial profiles of the ion \(n_T(r)\) and electron \(n_e(r)\) number densities, electron Fermi energy and escape velocity \(v_{esc}(r)\) .
## Neutron Stars {#sec:NSs}
NSs are the most compact stars known in the Universe. Neutron degeneracy pressure supports them against collapse. Despite recent breakthroughs in NS physics, their exact composition remains still unknown and the EoS of neutron-rich matter an open problem in nuclear astrophysics. NSs are mainly composed of degenerate neutrons, but inverse beta equilibrium allows the presence of protons and electrons. Muons appear in the NS core when the electron chemical potential reaches the muon mass. We model the NS interior and related microphysics by assuming a relativistic EoS that satisfies current observational constraints and enables the presence of hyperons in the NS inner core, the quark-meson coupling (QMC) model , and solving the TOV equations. Radial profiles of the relevant quantities can be found in ref. .
# Capture of Dark Matter in Compact Stars {#sec:capture}
## Capture by scattering off ions {#sec:capions}
First, we consider DM scattering off the ionic targets in WDs. Since ions are non-relativistic and the WD gravitational potential is sufficiently weak so that Newtonian gravity holds, we use an approach similar to that of the Sun to compute the capture rate \[C = \frac{16\pi\mu_+^2\rho_\chi}{\mu m_\chi}\int_0^{R_\star} dr \, n_T(r) \eta(r) r^2 \int_0^\infty du_\chi \frac{f_{\rm MB}(u_\chi)}{u_\chi} \int_{w(r)\frac{|\mu_-|}{\mu_+}}^{v_{esc}(r)}dv v \frac{d\sigma_{T\chi}}{d\cos\theta_{cm}} (w,q^2),\label{eq:capions}\] where \(\mu = m_\chi/m_T\), \(\mu_\pm = (\mu\pm 1)/2\), \(\rho_\chi\) is the DM density, \(m_\chi\) the DM mass and \(m_T\) the target mass, respectively; \(\eta(r)\) is the optical factor that accounts for the star opacity, defined in refs. ; \(w^2(r)=u_\chi^2+v_{esc}^2(r)\) and \(v\) are the DM velocity before and after the collision, respectively; \(q\) is the momentum transfer. We assumed a Maxwell Boltzmann distribution \(f_{\rm MB}(u_\chi)\) for the DM velocity far away from the star \(u_\chi\). Note that the differential DM-target cross section is written in the basis of non-relativistic operators and includes the nuclear response function (form factors) as calculated in ref. (see ref. for further details).
## Capture by scattering off a free Fermi gas of degenerate leptons {#sec:caplept}
Degenerate leptons in both, white dwarfs and neutron stars, are relativistic and subject to Pauli blocking. Therefore, Eq. [\[eq:capions\]](#eq:capions){reference-type="ref" reference="eq:capions"} cannot be applied to this case. We re-derived this expression using the TOV equations, the Schwarzschild metric and relativistic kinematics, and found \[\begin{aligned}
C &= 4\pi \frac{\rho_\chi}{m_\chi} \int_0^\infty du_\chi \frac{f_{\rm MB}(u_\chi)}{u_\chi} \int_0^{R_\star} dr \eta(r) r^2 \frac{\sqrt{1-B(r)}}{B(r)} \Omega^{-}(r) , \label{eq:caprate}\\ \Omega^{-}(r) &= \frac{\zeta(r)}{32\pi^3}\int dt dE_i ds \frac{ |\overline{M}(s,t,m_i)|^2}{s^2-[m_i^2-m_\chi^2]^2}\frac{ s E_i}{m_\chi}\sqrt{\frac{B(r)}{1-B(r)}}\frac{ f_{\rm FD}(E_i,r)(1-f_{\rm FD}(E_i^{'},r))}{\sqrt{(s-m_i^2-m_\chi^2)^2-4m_i^2m_\chi^2}}, \label{eq:intrate} \end{aligned}\] where \(f_{\rm FD}\) is the Fermi Dirac distribution, terms containing this function deal with Pauli suppression of the target initial and final states, \(B(r)\) is the coefficient of the time part of the Schwarzschild metric and encodes general relativity corrections (very relevant for NSs), \(|\overline{M}|^2\), is the squared matrix element, \(m_i\) is the mass of the target \(i\), \(s\) and \(t\) are the Mandelstam variables, \(E_i\) and \(E_i^{'}\) are the target initial and final energies, respectively. The integration range for \(s\), \(t\) and \(E_i\) can be found in refs. . \(\zeta(r)=n_i(r)/n_{free}(r)\) is a correction factor that accounts for the fact that we are using realistic number density \(n_i(r)\) and Fermi energy profiles while assuming a free Fermi gas. The expression for \(n_{free}\) is given in ref. .
## Capture by scattering off a Fermi sea of interacting baryons {#sec:capbar}
At the extreme densities found in NSs, nucleons, and in general baryons undergo strong interactions. Strong many body forces are described in terms of relativistic scalar and vector mean fields in the QMC EoS. Under the former field, baryons develop an effective mass \(m_i^{\rm eff}\), which decreases with increasing density. Thus, \(m_i^{\rm eff}\), where \(i\) denotes the specific baryon, is lower than the rest mass in vacuum \(m_i\) towards the NS centre, and can be as low as \(\sim0.5m_i\) for nucleons . This entails that the ideal Fermi gas is not a good approximation to calculate the DM-baryon interaction rate Eq. [\[eq:intrate\]](#eq:intrate){reference-type="ref" reference="eq:intrate"}. Properly incorporating the effect of strong interactions in Eq. [\[eq:intrate\]](#eq:intrate){reference-type="ref" reference="eq:intrate"} implies not only replacing \(m_i\) with \(m_i^{\rm eff}\), but also calculating the Fermi energy of a single baryon as a function of its number density and \(m_i^{\rm eff}\), and thereby \(\zeta(r)=1\) . In addition, since DM is accelerated to quasi-relativistic speeds upon infall to a NS, the momentum transfer in the DM-baryon scattering process is sufficiently large that baryon targets cannot be treated as point-like particles. We take this into account by incorporating the momentum dependence of the hadronic matrix elements. Thus, the squared couplings of the baryon \(i\) are \(c_i(t) = c_i(0)/(1-t/Q_0^2)^4\), where \(Q_0\simeq1{\rm \,GeV}\) is a scale that depends on the specific interaction and target, and \(c_i(0)\) are the squared coefficients at zero momentum transfer which depend on the hadronic matrix elements of the specific interaction and baryon as in DD . Note that the \(t\)-dependent baryon couplings are embedded in the squared matrix element \(|\overline{M}(s,t,m_i^{\rm eff})|^2\).
# Results {#sec:results}
We consider fermionic DM that scatters off either electron or ion targets in WDs, these interactions are described by the dimension-6 effective field theory (EFT) scattering operators . We compute the capture rate for carbon WDs using Eq. [\[eq:capions\]](#eq:capions){reference-type="ref" reference="eq:capions"} for ions and Eqs. [\[eq:caprate\]](#eq:caprate){reference-type="ref" reference="eq:caprate"} and [\[eq:intrate\]](#eq:intrate){reference-type="ref" reference="eq:intrate"} for electron targets, and the radial profiles obtained in section [2.1](#sec:WDs){reference-type="ref" reference="sec:WDs"}. Next, we derive limits on the cutoff scale of these operators by comparing the DM contribution to the WD luminosity due to capture and further annihilation with the observed luminosity of old WDs in the globular cluster M4 . The most constraining WD being the heaviest \(M_\star=1.38M_\odot\) and faintest. Note that we have assumed the existence of DM in M4, which is yet to be proved, and \(\rho_\chi\simeq531.5-798{\rm \,GeV}{\rm \,cm}^{-3}\). In Fig. [\[fig:Wdbounds\]](#fig:Wdbounds){reference-type="ref" reference="fig:Wdbounds"}, we recast these bounds in terms of the DM-proton (left panel) and DM-electron (right panel) cross sections (light blue band) for the scalar and vector operators, respectively. For DM-nucleon scattering, we find that WDs can probe the sub-GeV mass range, with its reach limited by evaporation . For DM-electron scattering, the WD bound outperforms electron recoil experiments in the full mass range, with its low mass endpoint limited by DM annihilation to neutrinos that escape the WD . To project the NS sensitivity to DM-nucleon and DM-lepton scattering cross sections, we calculate the capture rate in the optically thin limit, \(\eta(r)=1\), for the EFT operators, as outlined in sections [3.2](#sec:caplept){reference-type="ref" reference="sec:caplept"} and [3.3](#sec:capbar){reference-type="ref" reference="sec:capbar"}, and the radial profiles from section [2.2](#sec:NSs){reference-type="ref" reference="sec:NSs"} for NSs of mass in the \(1-1.9M_\odot\) range. To determine the maximum cross section that can be probed with NSs, the threshold cross section \(\sigma_{th}\), we equate \(C (m_\chi,\sigma_{i\chi})\) with the expression for the geometric limit given in refs. . Note that for \(\sigma_{i\chi}>\sigma_{th}\), the capture rate saturates the geometric limit. In Fig. [\[fig:sigmath\]](#fig:sigmath){reference-type="ref" reference="fig:sigmath"}, we show \(\sigma_{th}\) for the vector operator and leptonic targets (top panel), as well as nucleon targets for the scalar (bottom left panel) and axialvector (bottom right panel) operators. The decrease in sensitivity below \(m_\chi\sim0.2{\rm \,GeV}\) is due to Pauli blocking and that above \(m_\chi\simeq4\times10^5{\rm \,GeV}\) (nucleons) and \(m_\chi\sim[1,2]\times10^5{\rm \,GeV}\) (leptons) to the fact that multiple collisions are required to capture heavy DM. As we can see, the NS sensitivity greatly surpasses that of DD in the whole DM mass range considered for DM-neutron SD and DM-lepton interactions. The leading SI DD bounds are more stringent in the \(\sim10-10^4{\rm \,GeV}\) mass range, below which the NS sensitivity outperforms present and future DD experiments.
# Conclusion {#sec:conclusion}
The extreme conditions found in compact stars made them promising dark matter (DM) probes. DM that accumulates and annihilates in the interior of old isolated white dwarfs (WDs), may transfer enough energy to these stars that can prevent them from cooling, provided that they are located in DM-rich environments. Thus, the null detection of anomalously warm old WDs could constrain DM interactions with ordinary matter. In neutron stars (NSs), on the other hand, due to their stronger gravitational field that accelerates DM to quasi-relativistic speeds, only the energy transferred in the capture process would be enough to heat local NSs up to infrared temperatures for maximal capture efficiency. We have shown that the NS sensitivity excels that of direct detection experiments for DM-nucleon spin-dependent and DM-lepton scattering in the full DM mass range, and for the spin-independent scattering of sub-GeV DM. | {'timestamp': '2022-12-19T02:12:04', 'yymm': '2212', 'arxiv_id': '2212.08478', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08478'} |
null | null |
null | null |
null | null |
# Introduction
In data protection, cybersecurity plays a key role in authenticating users and giving them access to personal and untransferable information. A failure in the user authentication system can lead to big economic, reputational, and social damage, and therefore, authentication systems are becoming more and more robust. In this scenario, biometry is playing an important role because it enables a universal, singular, permanent in time and measurable system of authenticating users. In the last years, user authentication systems based on biometry are being used in very diverse scenarios, such as in airport scanners, banking, military access control, smartphones, and forensics among others ,. These systems, usually based on Machine Learning techniques, extract feature measurements and decide whether they correspond to the characteristics of the user that is requesting access. Biometrics is divided into two subfields: physical biometrics and behavioral biometrics. However, this study is focused on behavioral biometrics. Despite presenting optimistic results in a wide range of applications, behavioral biometrics can be attacked in different ways. The most common attack on a user authentication system is the Presentation Attack (PA), which consists of an attack on the biometric sensor that captures the individual's measurements. Keystroke dynamics is a type of behavioral biometrics that refers to the way a user types on a keyboard taking into account the speed, the rhythm, the common mistakes and the times spent pressing and releasing each keycode. It has been demonstrated that users can be differentiated and verified depending on their typing pattern. However, it can be very difficult to imitate such keystroke patterns by another user or bot. In literature, all major achievements have been made with fixed text, and, to the best of our knowledge, free-text dynamics have not been studied at a large scale yet. Moreover, specifically in Presentation Attack with this type of data has not been studied yet. For this reason, in this project, we propose a new approach that uses free-text dynamics for learning and generalizing user keystroke behavior. Furthermore, we study how to use such information to implement a keystroke presentation attack on a biometric authenticator. In particular, we propose to use conditional Generative Adversarial Networks (cGAN) for learning keystroke dynamics and generate synthetic patterns that can deceive keystroke authentication systems. The rest of the paper is organized as follows. In Section [2](#sec:preliminaries){reference-type="ref" reference="sec:preliminaries"}, concepts regarding keystroke dynamics, generative models, and related work are introduced. In Section [3](#sec:methodology){reference-type="ref" reference="sec:methodology"}, the proposed methodology is detailed and the data used, the metrics, the experiments, and the validation process carried out in this study are presented. In Section [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}, results are reported and last, Section [\[sec:conclusions\]](#sec:conclusions){reference-type="ref" reference="sec:conclusions"}, provides conclusions and guidelines for future work.
# Preliminaries {#sec:preliminaries}
In this section, some background concepts for understanding this study are explained. In section [2.1](#subsec:keystroke){reference-type="ref" reference="subsec:keystroke"}, keystroke dynamics are presented, and in section [2.2](#subsec:gan){reference-type="ref" reference="subsec:gan"} GANs and an extension of these are presented (Conditional Generative Adversarial Networks, cGAN), and their benefits and drawbacks are explained. Last, the dataset used in this study is presented in section [2.3](#subsec:data){reference-type="ref" reference="subsec:data"}.
## Keystroke dynamics {#subsec:keystroke}
Keystroke dynamics are a type of behavioral biometric that measure the speed, the rhythm, the common errors, and the typing times for pressing and releasing each keycode to identify patterns of typing IDs to identify or verify a user's identity. It has been proved that each person has their typing ID due to the habit of typing certain sequences always in a similar way, such as the user's name, password, etc. The main advantages of using keystroke dynamics in user authentication systems are that there is no need for any special hardware because a keyboard in a device is the only requirement and it enables continuous user authentication without disturbing the user experience while using such device. It is thought that each user has its own typing ID and it can be very difficult to be imitated by another user or bot In literature about keystroke dynamics, the robustness of keystroke dynamics against synthetic falsification attacks and the robustness against replay attacks (an attack that collects and re-sends data to try to spoof the system) have been studied and it has been shown that keystroke dynamics can be used to increase user authentication reliability using different modeling techniques such as SVM and NN. Acien et al. presented a Siamese Network capable of distinguishing whether two given keystroke sequences belonged to the same user or belonged to two different users. A Siamese Network consists of a type of NN composed of several subnetworks that are identical in architecture and weights but have got different inputs whose outputs are combined in a common loss function in charge of measuring the similarity between inputs. In this work, this model will be referenced as TypeNet.
## Generative Adversarial Network {#subsec:gan}
Generative Adversarial Networks (GAN) are a type of generative modeling from deep learning introduced by Ian Goodfellow. They make use of two Neural Networks that compete with each other during their training: the generator (\(G\)) and the discriminator (\(D\)). The idea is that, through this adversarial training, both networks improve their performances. These networks play a zero-sum or min-max game, that is, the two networks play an adversarial game, one loses when the other wins and the other way round. This min-max game may not always reach an equilibrium, such a problem is known as *non-convergence* problem. Apart from the *non-convergence problem*, training a GAN may also result in *mode collapse* and *vanishing gradient* which refer to a problem in the output of the generator and a problem with the value of the gradient during training respectively. GANs are widely used to create or generate new samples of data that are as realistic as the samples in the original dataset. Since this new framework was proposed, GANs have been used for very diverse applications, and due to the difficulty when training, many extensions of GANs have been developed. They have been used for image creation employing Deep Convolutional Generative Adversarial Networks (DCGAN), data augmentation using GANs, image-to-image translation utilizing Conditional Generative Adversarial Networks (cGAN) and CycleGANs, melody generation from lyrics with Conditional LSTM-GANs. What is more, in 2016, Zhang et al. proposed the StackGAN, another extension of GANs that transforms text descriptions into realistic images. In 2017, Karras et al. proposed a new methodology for GANs to improve training stability and speed it up through a Progressive Growing of GANs. In the last years, new GAN extensions have been also used to generate synthetic sequential data. In 2017, Esteban et al. proposed the Recurrent GAN and the Recurrent Conditional GAN to generate synthetic multi-dimensional time series. Among these GAN structures, in this study, we implement the Conditional GAN (or cGAN). **Conditional Generative Adversarial Networks (cGAN)** are an extension of the GANs presented by Mehdi Mirza in 2014,. The main difference is that cGANs can generate data from a previously specified category or class. Both the generator and the discriminator in cGANs have got an extra input, the extra information that will condition the generated sample. This condition can be a class label, a numerical value, or an embedding vector. Therefore, the generator will try to generate samples from the corresponding data distribution of that class and the discriminator will learn to decide whether the given pair (sample and condition) is a matching real sample or has been artificially generated by the generator. The cGAN designed, implemented, and validated in this study aims to generate synthetic data on keystroke dynamics starting with two inputs: a latent vector and a character embedding. An embedding is a numerical representation of some high dimensional vector or text data into a low dimensional vector. The use of embeddings in Machine Learning (ML) models is widespread due to their contribution to making model training easier and more efficient. In this study, a pre-trained character embedding known as *char2vec* has been used.
## Dataset {#subsec:data}
This study has been carried out using a dataset provided by the Aalto University in Finland, in which they collected data from \(168,000\) volunteers in an online study. This dataset is a large-scale dataset that contains information on the keystroke patterns of each of the volunteers. It contains keystroke times of \(15\) sentences for each volunteer. For each keystroke pressed to type the sentence, the pressing and releasing times were registered for more than \(136\) million of keystrokes. However, in this study, as the aim is to impersonate a single user and validate it with an external model, a subset of the whole dataset has been used. On the one hand, to train the external model, TypeNet, a dataset composed of data about \(25\) volunteers have been used. In this study, it will be referenced as TypeNet dataset. On the other hand, to train the cGAN, a dataset of a single user has been used, which will be referenced as single-user dataset. It is important to outline that despite having the records of the pressing and releasing timestamps of the keycodes in the dataset, in line with the literature in the area, four variables were extracted to build the ML models: Hold Latency (HL), Press Latency (PL), Release Latency (RL) and Inter-Key Latency (IL).
# Methodology {#sec:methodology}
In this section, first, a global view of the methodology is explained, next, in section [\[subsec:model\]](#subsec:model){reference-type="ref" reference="subsec:model"}, the architecture of the cGAN is detailed. In section [\[subsec:framework\]](#subsec:framework){reference-type="ref" reference="subsec:framework"}, some important aspects of the experimental framework are presented. In section [\[subsec:metrics\]](#subsec:metrics){reference-type="ref" reference="subsec:metrics"}, the metrics used to validate the results are presented, and finally, in section [\[subsec:validation\]](#subsec:validation){reference-type="ref" reference="subsec:validation"}, the validation procedure is described.
## Proposal {#subsec:proposal}
This study aims to generate synthetic data on keystroke dynamics to impersonate a user and develop a presentation attack. To that end, we propose a cGAN able to learn the keystroke dynamics of a specific user and generate realistic data samples. These generated data samples will be able to fool an external model, TypeNet, used as a biometric authenticator to access a determined service. Let us assume that Alice is a registered user of concrete service. To access this service, a double verification is carried out. the user that wants to access the service, not only has to type its password, but the system also checks the keystroke dynamics of the user typing it and verifies the user's typing ID to enable access or deny, see step \(A\) in Fig. [\[fig:stepa\]](#fig:stepa){reference-type="ref" reference="fig:stepa"}. However, a non-registered user, Bob, wants to access the service with Alice's credentials. To do so, let us assume Bob already knows Alice's password, see step \(B\) in Fig. [\[fig:stepa\]](#fig:stepa){reference-type="ref" reference="fig:stepa"}. To access the service, he needs to imitate Alice's typing behavior. In other words, he needs to impersonate her by generating synthetic behavioral biometric data to fool the biometric authenticator, see step \(C\) in Fig. [\[fig:stepa\]](#fig:stepa){reference-type="ref" reference="fig:stepa"}. | {'timestamp': '2022-12-19T02:11:09', 'yymm': '2212', 'arxiv_id': '2212.08445', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08445'} |
# Introduction
Around 20 years ago, Stephen Anco and George Bluman introduced a comprehensive practical method for determining conservation laws of partial differential equations (PDEs) in Kovalevskaya form. The method is based on finding adjoint symmetries and applying Helmholtz conditions[^1]. A key part of the calculation is the inversion of the total divergence operator \(\mathrm{Div}\) to obtain the components of the conservation law. Usually, this can be done by using a homotopy operator, but the following three problems may occur with the standard homotopy formula (which is given by Olver in ).
1. The homotopy formula uses definite integrals. This works well if the divergence is a differential polynomial; by contrast, rational polynomials commonly have a singularity at one limit. Hickman and Poole & Hereman suggest avoiding these by working in terms of indefinite integrals, an approach that we use throughout this paper. Alternatively, one can move the singularity by modifying the dependent variable (see Anco & Bluman and Poole & Hereman ).
2. Scaling is fundamental to the homotopy approach to inversion. For instance, varying the scaling parameter in the standard homotopy formula moves contributions from variables along a ray to the origin. However, scaling does not change rational polynomials that are homogeneous of degree zero, so the standard inversion process does not work for such terms. Deconinck & Nivala discussed this problem in some detail (for one independent variable only) and suggested possible workarounds, but commented, 'We are unaware of a homotopy method that algorithmically avoids all problems like the ones demonstrated\...'. Poole & Hereman proposed an approach that works well for problems with one independent variable, but noted the difficulties of extending this to multiple independent variables (in a way that can be programmed).
3. The standard homotopy operator applies to a starshaped domain and integrates along rays to the origin, changing all Cartesian coordinates at once. This gives an inefficient inversion, in the sense that the number of terms is generally very much greater than necessary; the homotopy formula creates superfluous elements of \(\mathrm{ker}(\mathrm{Div})\). For polynomial divergences, Poole & Hereman removed curls by parametrizing all terms given by the homotopy formula and optimizing the resulting linear system. This approach is very effective, because (except where there are cancellations), the homotopy formula tends to include every possible term that can appear in an inversion. However, it is unclear whether this approach can be generalized to non-polynomial divergences. Moreover, the removal of curl terms takes extra processing time and does not allow for the possibility that new terms might appear in the most concise inversion of the divergence. If inversion could be done with respect to one independent variable at a time, this might prevent the occurrence of superfluous curls from the outset.
The current paper extends the efficient one-variable-at-a-time approach to total derivatives. Indefinite integration is used, as advocated by Hickman for standard homotopy operators; it avoids the complications resulting from singularities. From the computational viewpoint, the biggest advantage of integration with respect to just one independent variable is that the major computer algebra systems have efficient procedures for computing antiderivatives. The keys to inverting a total divergence one variable at a time are 'partial Euler operators' (see Section [3](#pEul){reference-type="ref" reference="pEul"}). These enable the inversion of a total derivative \(D_x\) to be written as an indefinite line integral. Section [4](#DivInversion){reference-type="ref" reference="DivInversion"} introduces a new iterative method for inverting a given total divergence; typically, this does not produce superfluous terms and very few iterations are needed. Furthermore, it can cope with components that are unchanged by the relevant scaling operator. The methods in this paper are systematic, leading to procedures that are intended to be suitable for implementation in computer algebra systems.
# Standard differential and homotopy operators {#Basics}
Here is a brief summary of the standard operators that are relevant to total divergences; for further details, see Olver. The independent variables \(\mathbf{x} = (x^1,\dots, x^p)\) are local Cartesian coordinates and the dependent variables \(\mathbf{u}=(u^1,\dots,u^q)\) may be real-or complex-valued. The Einstein summation convention is used to explain the main ideas and state general results. In examples, commonly-used notation is used where this aids clarity. All functions are assumed to be locally smooth, to allow the key ideas to be presented simply. Derivatives of each \(u^\alpha\) are written as \(u^\alpha_{\mathbf{J}}\), where \(\mathbf{J}=(j^1,\dots,j^p)\) is a multi-index; each \(j^i\) denotes the number of derivatives with respect to \(x^i\), so \(u^\alpha_{\mathbf{0}}= u^\alpha\). The variables \(x^i\) and \(u^\alpha_{\mathbf{J}}\) can be regarded as jet space coordinates. The *total derivative* with respect to \(x^i\), \[D_i=\frac{\partial}{\partial x^i}+u^\alpha_{\mathbf{J} +\boldsymbol{1}_i}\,\frac{\partial}{\partial u^\alpha_{\mathbf{J}}}\,,\quad\text{where}\quad \mathbf{J} +\boldsymbol{1}_i=(j^1,\dots,j^{i-1},j^i+1,j^{i+1},\dots, j^p),\] treats each \(u^\alpha_{\mathbf{J}}\) as a function of \(\mathbf{x}\). To keep the notation concise, write \[D_\mathbf{J}=D_1^{j^1}D_2^{j^2}\cdots D_p^{j^p};\] note that \(u^\alpha_{\mathbf{J}}=D_\mathbf{J}(u^\alpha)\). Let \([\mathbf{u}]\) represent \(\mathbf{u}\) and finitely many of its derivatives; more generally, square brackets around an expression denote the expression and as many of its total derivatives as are needed. A *total divergence* is an expression of the form \[\mathcal{C}=\mathrm{Div}(\mathbf{F}):= D_iF^i(\mathbf{x},[\mathbf{u}]).\] (If all \(F^i\) depend on \(\mathbf{x}\) only, \(\mathcal{C}\) is an ordinary divergence.) A conservation law of a given system of partial differential equations (PDEs), \(\mathcal{A}_\ell(\mathbf{x},[\mathbf{u}])=0,\ \ell=1,\dots,L\), is a total divergence that is zero on all solutions of the system; each \(F^i\) is a finite sum of terms. By using elementary algebraic operations (in particular, expanding logarithms of products and products of sums), the number of linearly independent terms may be maximized. When the number of linearly independent terms is maximal for each \(i\), we call the result the *fully-expanded* form of \(\mathbf{F}\). When \(p>1\), the \(p\)-tuple of components, \(\mathbf{F}\), is determined by \(\mathcal{C}\) up to a transformation of the form \[\label{trans} F^i\longmapsto F^i+D_j\left\{f^{ij}(\mathbf{x},[\mathbf{u}])\right\},\qquad\text{where}\ f^{ji}=-f^{ij}.\] (If \(p=3\), such a transformation adds a total curl to \(\mathbf{F}\).) The total number of terms in \(\mathbf{F}\) is the sum of the number of terms in all of the fully-expanded components \(F^i\). If this cannot be lowered by any transformation [\[trans\]](#trans){reference-type="eqref" reference="trans"}, we call \(\mathbf{F}\) *minimal*. Commonly, there is more than one minimal \(\mathbf{F}\), any of which puts the inversion of \(\mathrm{Div}\) in as concise a form as possible. If \(p=1\), the sole component \(F^1\) (also denoted \(F\)) is determined up to an arbitrary constant, so the number of non-constant terms is fixed. The formal adjoint of a differential operator (with total derivatives), \(\mathcal{D}\), is the unique differential operator \(\mathcal{D}^{\boldsymbol{\dag}}\) such that \[f\,\mathcal{D} g-\left(\mathcal{D}^{\boldsymbol{\dag}} f\right)g\] is a total divergence for all functions \(f(\mathbf{x},[\mathbf{u}])\) and \(g(\mathbf{x},[\mathbf{u}])\). In particular, \[\label{modJ} (D_\mathbf{J})^{\boldsymbol{\dag}}=(-D)_\mathbf{J}:=(-1)^{|\mathbf{J}|}D_\mathbf{J},\quad\text{where}\quad |\mathbf{J}|=j^1+\cdots+j^p.\] Thus the (standard) Euler--Lagrange operator corresponding to variations in \(u^\alpha\) is \[\mbE_{u^\alpha}=(-D)_\mathbf{J}\frac{\partial}{\partial u^\alpha_{\mathbf{J}}}\.\] Total divergences satisfy a useful identity: a function \(\mathcal{C}(\mathbf{x},[\mathbf{u}])\) is a total divergence if and only if \[\label{divcrit} \mbE_{u^\alpha}(\mathcal{C})=0,\qquad \alpha=1,\dots,q.\] Given a Lagrangian function \(L(\mathbf{x},[\mathbf{u}])\), the Euler--Lagrange equations are \(\mbE_{u^\alpha}(L)=0\). Given a set of Euler--Lagrange equations that are polynomial in the variables \((\mathbf{x},[\mathbf{u}])\), the function \(\overline{L}\) given by the homotopy formula \[\overline{L}(\mathbf{x},[\mathbf{u}])=\int_{0}^1u^\alpha\left\{\mbE_{u^\alpha}(L)\right\}\!\big|_{[\mathbf{u}\mapsto\lambda\mathbf{u}]}\,\mathrm{d}\lambda\] differs from \(L\) by a total divergence. (The same applies to many, but not all, non-polynomial Euler--Lagrange equations.) When \(p=1\), the equation \(P(x,[\mathbf{u}])=D_xF\) is invertible (at least, for polynomial \(P\)) by the following standard homotopy formula: \[\label{homDxF} F(x,[\mathbf{u}])=\!\int_{0}^1\sum_{i=1}^\infty D_x^{i-1}\!\left(u^\alpha\!\left\{\sum_{k\geq i}\binom{k}{i}(-D_x)^{k-i}\frac{\partial P(x,[\mathbf{u}])}{\partial (D_x^ku^\alpha)}\right\}\Bigg|_{[\mathbf{u}\mapsto\lambda\mathbf{u}]}\right)\!\,\mathrm{d}\lambda +\int_{0}^1\!xP(\lambda x,[0])\,\mathrm{d}\lambda.\] The operator acting on \(P\) in the braces above is the *higher Euler operator* of order \(i\) for \(p=1\). When \(p\geq 2\), the standard homotopy formula is similar, but somewhat more complex (see Olver for details); it is based on higher Euler operators and integration along rays in a totally starshaped domain. The following example illustrates that even for quite simple divergences, this formula commonly yields inversions with many superfluous terms.
The homotopy formulae above can be applied or adapted to some, but not all, classes of non-polynomial Lagrangians and divergences.
# Partial Euler operators and partial scalings {#pEul}
This section introduces some ideas and results that underpin integration with respect to one independent variable at a time. The independent variable over which one integrates is distinguished; this is denoted by \(x\). For instance, if \(x=x^1\), replace the derivative index \(\mathbf{J}\) by \((\mathbf{I},j)\), where \(j=j^1\) and \(\mathbf{I}=(j^2,\dots,j^{p})\). So the dependent variables and their derivatives are denoted \[u^\alpha_{\mathbf{I},\,j}=D_x^{\,j}u^\alpha_{\mathbf{I}},\qquad\text{where}\ u^\alpha_{\mathbf{I}}=u^\alpha_{\mathbf{I},0}.\] In examples, however, we write each \(u^\alpha_{\mathbf{I}}\) more simply (as \(u\), \(v_y\), \(u_{yt}\), and so on), using \(\phantom{}_{,j}\) for \(D_x^{\,j}\).
## Partial Euler operators
The *partial Euler operator* with respect to \(x\) and \(u^\alpha_{\mathbf{I}}\) is obtained by varying each \(u^\alpha_{\mathbf{I}}\) independently, treating \(x\) as the sole independent variable: \[\mbE_{u^\alpha_\mathbf{I}}^{x}=(-D_x)^j\frac{\partial}{\partial u^\alpha_{\mathbf{I},j}}\.\] Consequently, the standard Euler operator with respect to \(u^\alpha\) amounts to \[\label{Eulp} \mbE_{u^\alpha}=(-D)_\mathbf{I}\mbE_{u^\alpha_\mathbf{I}}^{x}\.\] Similarly, the partial Euler operator with respect to \(x\) and \(u^\alpha_{\mathbf{I},k}\) is \[\mbE_{u^\alpha_{\mathbf{I},k}}^{x}=(-D_x)^j\frac{\partial}{\partial u^\alpha_{\mathbf{I},j+k}}\.\] Note that \[\label{uppE} \mbE_{u^\alpha_{\mathbf{I},k}}^{x}=\frac{\partial}{\partial u^\alpha_{\mathbf{I},k}}-D_x\mathbf{E}^x_{u^\alpha_{\mathbf{I},k+1}}.\] The following identities are easily verified; here, \(f(\mathbf{x},[\mathbf{u}])\) is an arbitrary function. \[\begin{aligned}
&\mbE_{u^\alpha_\mathbf{I}}^{x}(D_x f)=0,\label{id1}\\ &\mbE_{u^\alpha_{\mathbf{I},k}}^{x}(D_x f)=\frac{\partial f}{\partial u^\alpha_{\mathbf{I},k-1}}\,,\qquad k\geq 1,\label{id2}\\ &\mbE_{u^\alpha_{\mathbf{I},k}}^{x}(D_i f)=D_i\left(\mbE_{u^\alpha_{\mathbf{I},k}}^{x}(f)\right)+\mathbf{E}^x_{u^\alpha_{\mathbf{I}-\boldsymbol{1}_i,k}}(f),\qquad x^i\neq x,\label{id3} \end{aligned}\] where the last term in [\[id3\]](#id3){reference-type="eqref" reference="id3"} is zero if \(j^i=0\).
## Inversion of \(D_x\)
The identities [\[id1\]](#id1){reference-type="eqref" reference="id1"}, [\[id2\]](#id2){reference-type="eqref" reference="id2"} and [\[id3\]](#id3){reference-type="eqref" reference="id3"} are central to the inversion of total divergences, including the following inversion of \(P=D_xF\) as an indefinite line integral.
## Integration by parts
From here on, we will restrict attention to total divergences \(\mathcal{C}\) whose fully-expanded form has no terms that depend on \(\mathbf{x}\) only. Such terms can be inverted easily by evaluating an indefinite integral, as explained in the Introduction. Henceforth, all indefinite integrals denote antiderivatives with the minimal number of terms in their fully-expanded form. Any arbitrary constants and functions that would increase the number of terms are set to zero. This restriction facilitates the search for minimal inversions. The indefinite line integral formula [\[line\]](#line){reference-type="eqref" reference="line"} is closely related to integration by parts. To see this, we introduce a *positive ranking* on the variables \(u^\alpha_{\mathbf{J}}\,\); this is a total order \(\preceq\) that is subject to two conditions: \[(i)\,\ u^\alpha\prec u^\alpha_{\mathbf{J}}\,,\quad\mathbf{J}\neq\mathbf{0},\qquad (ii)\,\ u^\beta_{\mathbf{I}}\prec u^\alpha_{\mathbf{J}}\Longrightarrow D_{\mathbf{K} }u^\beta_{\mathbf{I}}\prec D_{\mathbf{K} }u^\alpha_{\mathbf{J}}\.\] The *leading part* of a differential function is the sum of terms in the function that depend on the highest-ranked \(u^\alpha_{\mathbf{J}}\), and the *rank* of the function is the rank of its leading part (see Rust *et al.* for details and references). Let \(f(\mathbf{x},[\mathbf{u}])\) denote the leading part of the fully-expanded form of \(F\) and let \(\mathrm{U}_{,k}\) denote the highest-ranked \(u^\alpha_{\mathbf{I},k}\); then the highest-ranked part of \(P=D_xF\) is \(\mathrm{U}_{,k+1}\partial f/\partial \mathrm{U}_{,k}\). Then [\[line\]](#line){reference-type="eqref" reference="line"} includes the contribution \[\int \mathbf{E}^x_{\mathrm{U}_{,k+1}}(P)\,\mathrm{d} \mathrm{U}_{,k}=\int \frac{\partial f}{\partial \mathrm{U}_{,k}}\,\,\mathrm{d} \mathrm{U}_{,k}= f +\ \text{lower-ranked terms}.\] Integration by parts gives the same result. Subtracting \(f\) from \(F\) and iterating shows that evaluating the line integral [\[line\]](#line){reference-type="eqref" reference="line"} is equivalent to integrating by parts from the highest-ranked terms downwards. Integration by parts is useful for splitting a differential expression \(P(\mathbf{x},[\mathbf{u}])\), with \(P(\mathbf{x},[0])=0\), into \(D_xF\) and a remainder, \(R\), whose \(x\)-derivatives are of the lowest-possible order. The splitting is achieved by the following procedure.
The reason for choosing an \(x\)-dominant ranking is to ensure that the derivative order with respect to \(x\) outweighs all other ranking criteria. Consequently, the minimally-ranked remainder cannot contain \(x\)-derivatives of unnecessarily high order.
## Partial scalings
To investigate partial Euler operators further, it is helpful to use a variant of the homotopy approach. The *partial scaling* (by a positive real parameter, \(\lambda\)) of a function \(f(\mathbf{x},[\mathbf{u}])\) with respect to \(x\) and \(u^\alpha_{\mathbf{I}}\) is the mapping \[\sigma_{u^\alpha_\mathbf{I}}^{x}:(f;\lambda)\mapsto f\big|_{\{u^\alpha_{\mathbf{I},j}\mapsto \lambda u^\alpha_{\mathbf{I},j},\ j\geq 0\}}.\] Again, each \(u^\alpha_{\mathbf{I}}\) is treated as a distinct dependent variable. Note the identity \[\label{scalder} \sigma_{u^\alpha_\mathbf{I}}^{x} D_x= D_x \sigma_{u^\alpha_\mathbf{I}}^{x}\.\]
By definition, the partial scaling \(\sigma_{u^\alpha_\mathbf{I}}^{x}\) fails to be a good scaling for \(f\) if and only if there are terms that are independent of \(\lambda\) in the fully-expanded form of \(\sigma_{u^\alpha_\mathbf{I}}^{x}(f;\lambda)\). The simplest cause of this is that the fully-expanded form of \(f\) has terms that are independent of \(u^\alpha_{\mathbf{I}}\) and its \(x\)-derivatives. However, this is not the only cause, as the following example illustrates.
Given a partial scaling \(\sigma_{u^\alpha_\mathbf{I}}^{x}\) and a differential function \(f(\mathbf{x},[\mathbf{u}])\), let \[\label{proj} \pi_{u^\alpha_\mathbf{I}}^{x}(f)=\lim_{\lambda\rightarrow 1}\int \frac{\,\mathrm{d}}{\,\mathrm{d} \lambda}\!\left(\sigma_{u^\alpha_\mathbf{I}}^{x}(f;\lambda)\right)\,\mathrm{d} \lambda.\] For \(\mu\) sufficiently close to \(1\) (using \(\widehat{f}\) as shorthand for \(\pi_{u^\alpha_\mathbf{I}}^{x}(f)\)), \[\sigma_{u^\alpha_\mathbf{I}}^{x}(\widehat{f};\mu)=\!\int \frac{\,\mathrm{d}}{\,\mathrm{d} \mu}\!\left(\sigma_{u^\alpha_\mathbf{I}}^{x}(f;\mu)\right)\!\,\mathrm{d} \mu=\!\int\frac{\,\mathrm{d}}{\,\mathrm{d} \mu}\!\left\{ \int \frac{\,\mathrm{d}}{\,\mathrm{d} \mu}\!\left(\sigma_{u^\alpha_\mathbf{I}}^{x}(f;\mu)\right)\,\mathrm{d} \mu\right\}\!\,\mathrm{d} \mu=\!\int \frac{\,\mathrm{d}}{\,\mathrm{d} \mu}\!\left(\sigma_{u^\alpha_\mathbf{I}}^{x}(\widehat{f};\mu)\right)\!\,\mathrm{d} \mu;\] the first equality comes from the proof of Lemma [\[goodslem\]](#goodslem){reference-type="ref" reference="goodslem"}. Therefore, \(\sigma_{u^\alpha_\mathbf{I}}^{x}\) is a good scaling for \(\pi_{u^\alpha_\mathbf{I}}^{x}(f)\). Moreover, there are no terms in the fully-expanded form of the remainder, \(f-\pi_{u^\alpha_\mathbf{I}}^{x}(f)\), for which \(\sigma_{u^\alpha_\mathbf{I}}^{x}\) is a good scaling, because \[\frac{\,\mathrm{d}}{\,\mathrm{d} \mu}\!\left(\sigma_{u^\alpha_\mathbf{I}}^{x}(f;\mu)\right)-\frac{\,\mathrm{d}}{\,\mathrm{d} \mu}\!\left(\sigma_{u^\alpha_\mathbf{I}}^{x}(\widehat{f};\mu)\right)=0.\] So \(\pi_{u^\alpha_\mathbf{I}}^{x}\) is the projection that maps a given function onto the component which has \(\sigma_{u^\alpha_\mathbf{I}}^{x}\) as a good scaling.
For instance, both \(\sigma^y_{u}\) and \(\sigma^y_v\) are poor scalings for [\[awkDiv\]](#awkDiv){reference-type="eqref" reference="awkDiv"}, as explained in Example [\[kerpi\]](#kerpi){reference-type="ref" reference="kerpi"}. Section [4.3](#poors){reference-type="ref" reference="poors"} addresses the problem of inverting divergences such as [\[awkDiv\]](#awkDiv){reference-type="eqref" reference="awkDiv"} that have poor scalings. First, we develop the inversion process for general divergences. The following results are fundamental.
Note that [\[invDF\]](#invDF){reference-type="eqref" reference="invDF"} is a homotopy formula for (at least partially) inverting \(D_xF\), giving a third way to do this. The line integral formula [\[line\]](#line){reference-type="eqref" reference="line"} carries out the full inversion in one step, but may take longer to compute.
# The inversion method for Div {#DivInversion}
This section introduces a procedure to invert \(\mathrm{Div}\), with a ranking heuristic (informed by experience) that is intended to keep the calculation short and efficient. To motivate the procedure, it is helpful to examine a simple example.
## A single iteration {#iteration}
The basic method for inverting a given total divergence one independent variable at a time works similarly to the example above. Suppose that after \(n\) iterations the inversion process has yielded components \(F^i_n\) and that an expression of the form \(\mathcal{C}=D_if_n^i\) remains to be inverted. For the next iteration, let \(\mbE_{u}^{x}\) be the partial Euler operator that is applied to \(\mathcal{C}\). Here \(u\) is one of the variables \(u^\alpha_{\mathbf{I}}\), which is chosen to ensure that for each \(i\) such that \(x^i\neq x\), \[\label{EDcom} \mbE_{u}^{x} D_if_n^i =D_i\mbE_{u}^{x} f_n^i.\] This requires care, in view of the identity [\[id3\]](#id3){reference-type="eqref" reference="id3"}. However, it is achievable by using the variables \(u^\alpha_{\mathbf{I}}\) in the order given by a ranking that is discussed in Section [4.2](#rankvars){reference-type="ref" reference="rankvars"}. This ranking is entirely determined by user-defined rankings of the variables \(x^j\) and \(u^\alpha\). Taking [\[id1\]](#id1){reference-type="eqref" reference="id1"} into account leads to the identity \[\label{Eulp0} \mbE_{u}^{x}(\mathcal{C})=\sum_{x^i\neq x}D_i(\mbE_{u}^{x}(f_n^i)),\] which, with together with Theorem [\[fundEul\]](#fundEul){reference-type="ref" reference="fundEul"}, is the basis of the inversion method. The method works without modification provided that:
- there are no poor scalings for any terms in \(\mathcal{C}\);
- the fully-expanded form of \(\mathcal{C}\) has no terms that are linear in \([\mathbf{u}]\).
We begin by restricting attention to divergences for which these conditions hold, so that \[\label{Eulp1} \mbE_{u}^{x}(\mathcal{C})=\sum_{x^i\neq x}D_i(\mbE_{u}^{x}(\pi_u^x f_n^i)),\] where every term in \(\mbE_{u}^{x}(\pi_u^x f_n^i)\) depends on \([\mathbf{u}]\). The modifications needed if either condition does not hold are given in Sections [4.3](#poors){reference-type="ref" reference="poors"} and [4.4](#lindiv){reference-type="ref" reference="lindiv"}. The iteration of the inversion process runs as follows. Calculate \(\mbE_{u}^{x}(\mathcal{C})\), which is a divergence \(D_iP^i\) with no \(D_x\) term, by [\[Eulp1\]](#Eulp1){reference-type="eqref" reference="Eulp1"}; it involves at most \(p-1\) (but commonly, very few) nonzero functions \(P^i\). Invert this divergence, treating \(x\) as a parameter. *If it is possible to invert in more than one way, always invert into the \(P^i\) for which \(x^i\) is ranked as low as possible*; the reason for this is given in the next paragraph. If \(\mbE_{u}^{x}(\mathcal{C})\) has nonlinear terms that involve derivatives with respect to more than one \(D_i\) (excluding \(D_x\)), this is accomplished by iterating the inversion process with as few independent variables as are needed. Otherwise, \(P^i\) can be determined more quickly by using integration by parts (Procedure [\[ibp\]](#ibp){reference-type="ref" reference="ibp"}, with \(x\) replaced by the appropriate \(x^i\)), and/or the method for linear terms (see Procedure [\[lininv\]](#lininv){reference-type="ref" reference="lininv"} in Section [4.4](#lindiv){reference-type="ref" reference="lindiv"}). Note that this shortcut can be used whenever there are only two independent variables. *At this stage, check that the fully-expanded form of each \(P^i\) has no terms that are ranked lower than \(u\).* If any term is ranked lower than \(u\), stop the calculation and try a different ranking of the variables \(x^j\) and/or \(u^\alpha\). This is essential because, to satisfy [\[EDcom\]](#EDcom){reference-type="eqref" reference="EDcom"} and avoid infinite loops, the variables \(u^\alpha_{\mathbf{I}}\) that are chosen to be \(u\) in successive iterations must progress upwards through the ranking. Where there is a choice of inversion, the rank of each term in \(P^i\) is maximized by using the \(x^i\) of minimum order; this avoids unnecessary re-ranking. Having found and checked \(P^i\), use [\[exp2\]](#exp2){reference-type="eqref" reference="exp2"} to obtain \[\label{invEul1} \pi_u^x(f_n^i)=\left\{\lim_{\lambda\rightarrow 1}\int u\,\sigma_u^x\!\left(P^i\,;\lambda\right)\,\mathrm{d} \lambda\right\} +D_xh^i,\] for arbitrary functions \(h^i(\mathbf{x},[\mathbf{u}])\). Apply Procedure [\[ibp\]](#ibp){reference-type="ref" reference="ibp"} to the function in braces and choose \(h^i\) to make the right-hand side of [\[invEul1\]](#invEul1){reference-type="eqref" reference="invEul1"} equal the remainder from this procedure. This yields the representation of \(\pi_u^x(f_n^i)\) that has the lowest-order derivatives (with respect to \(x\)) consistent with the inversion of \(D_iP^i\); call this representation \(f^i\). Commonly, such a lowest-order representation is needed to obtain a minimal inversion. By Theorem [\[fundEul\]](#fundEul){reference-type="ref" reference="fundEul"}, there exists \(\phi\) such that \[\label{Fdef} \pi_u^x\left(\mathcal{C}-\sum_{x^i\neq x}D_if^i\right)=D_x\phi,\] because (by construction) the expression in parentheses belongs to \(\mathrm{ker}(\mbE_{u}^{x})\). Use the line integral formula [\[line\]](#line){reference-type="eqref" reference="line"} or Procedure [\[ibp\]](#ibp){reference-type="ref" reference="ibp"} to obtain \(\phi\), then set \(f^i:=\phi\) for \(x^i=x\). Now update: set \[\mathcal{C}:=\mathcal{C}-D_if^i,\qquad F^i_{n+1}:=F^i_n+f^i.\]
## Ranking and using the variables {#rankvars}
Having described a single iteration, we now turn to the question of how to choose \(x\) and \(u\) effectively. The starting-point is to construct a *derivative-dominant* ranking of the variables \(u^\alpha_{\mathbf{J}}\). This is a positive ranking that is determined by:
- a ranking of the independent variables, \(x^1\prec x^2\prec \cdots \prec x^p\);
- a ranking of the dependent variables, \(u^1\prec u^2\prec \cdots \prec u^q\).
(Later in this section, we give a heuristic for ranking the dependent and independent variables effectively.) The derivative-dominant ranking (denoted \(\mathbf{u}_p\)) is constructed iteratively, as follows. \[\begin{aligned}
\mathbf{u}_0&=u^1\prec\cdots\prec u^q,\\ \mathbf{u}_1&=\mathbf{u}_0\prec D_1\mathbf{u}_0 \prec D_1^2\mathbf{u}_0 \prec\cdots ,\\ \mathbf{u}_2&=\mathbf{u}_1\prec D_2\mathbf{u}_1 \prec D_2^2\mathbf{u}_1 \prec\cdots ,\\ \vdots&\qquad\vdots\\ \mathbf{u}_p&=\mathbf{u}_{p-1}\prec D_p\mathbf{u}_{p-1} \prec D_p^2\mathbf{u}_{p-1} \prec\cdots. \end{aligned}\] In practice, very few \(u^\alpha_{\mathbf{J}}\) are needed to carry out many inversions of interest, but it is essential that these are used in the order given by their ranking, subject to a constraint on \(|\mathbf{I}|\) that is explained below. Given an independent variable, \(x\), we call \(u^\alpha_{\mathbf{I}}\) *relevant* if the updated \(\mathcal{C}\) depends on \(u^\alpha_{\mathbf{I},k}\) for some \(k\geq 0\). The first set of iterations uses \(x=x^1\). For the initial iteration, \(u\) is the lowest-ranked relevant \(u^\alpha\). In the following iteration, \(u\) is the next-lowest-ranked relevant \(u^\alpha\) and so on, up to and including \(u^q\). (From [\[id3\]](#id3){reference-type="eqref" reference="id3"}, the condition [\[EDcom\]](#EDcom){reference-type="eqref" reference="EDcom"} holds whenever \(u=u^\alpha,\ \alpha = 1,\dots, q\).) After these iterations, the updated \(\mathcal{C}\) is independent of \(\mathbf{u}\) and its unmixed \(x\)-derivatives. If the updated \(\mathcal{C}\) has any remaining \(x\)-derivatives, these are mixed. Thus, as \(\mathcal{C}\) has no linear terms, a necessary condition for the inversion to be minimal is that every \(f_n^i\) is independent of \(\mathbf{u}\) and its unmixed \(x\)-derivatives. Consequently, [\[EDcom\]](#EDcom){reference-type="eqref" reference="EDcom"} holds for \(u=u^\alpha_{\mathbf{I}}\) whenever \(|\mathbf{I}|=1\), because \[\mathbf{E}^x_{u^\alpha_{\mathbf{I}-\boldsymbol{1}_i}}(f_n^i)=0,\qquad x^i\neq x.\] Therefore, the process can be continued using each relevant \(u=u^\alpha_{\mathbf{I}}\) with \(|\mathbf{I}|=1\) in the ranked order. Iterating, the same argument is used with \(|\mathbf{I}|=2,3,\dots\), until \(\mathcal{C}\) is independent of \(x\)-derivatives. Now set \(x=x^2\) and iterate, treating \(x^1\) as a parameter. In principle, this can be continued up to \(x=x^p\); in practice, only a very few iterations are usually needed to complete the inversion. The best rankings invert many terms at each iteration. On the basis of some experience with conservation laws, the following heuristic for ranking the variables \(x^j\) and \(u^\alpha\) is recommended.
## How to deal with poor scalings {#poors}
To remove an earlier restriction on the inversion process, we now address the problem of poor scalings, namely, that \(u^\alpha_{\mathbf{I}}\) and its \(x\)-derivatives (denoted \([u^\alpha_{\mathbf{I}}]_x\)) may occur in terms that belong to \(\mathcal{C}-\pi_{u^\alpha_\mathbf{I}}^{x}(\mathcal{C})\). Such terms (when fully expanded) are products of homogeneous rational polynomials in \([u^\alpha_{\mathbf{I}}]_x\) of degree zero and logarithms of a single element of \([u^\alpha_{\mathbf{I}}]_x\). We refer to these collectively as *zero-degree* terms. To overcome this difficulty, we modify the approach used by Anco & Bluman to treat singularities. In our context, \([u^\alpha_{\mathbf{I}}]_x\) is replaced by \([u^\alpha_{\mathbf{I}}+U^\alpha_\mathbf{I}]_x\), where \(U^\alpha_\mathbf{I}\) is regarded as a new dependent variable that is ranked higher than \(u^\alpha_{\mathbf{I}}\). This approach works equally well for logarithms, ensuring that \(\pi_{u^\alpha_\mathbf{I}}^{x}\) is a good scaling for all terms that depend on \([u^\alpha_{\mathbf{I}}]_x\), so that its kernel consists only of terms that are independent of these variables. At the end of the calculation, all members of \([U^\alpha_\mathbf{I}]_x\) are set to zero. Note that there is no need to replace \(u^\alpha_{\mathbf{I}}\) in terms that are not zero-degree in \([u^\alpha_{\mathbf{I}}]_x\), as total differentiation preserves the degree of homogeneity[^2].
## Linear divergences {#lindiv}
The inversion process runs into a difficulty when a given divergence has terms that are linear in \([\mathbf{u}]\), with mixed derivatives. Then it is possible to invert in more than one way, some of which may not produce a minimal result. To address this, it is helpful to invert using a different process. Suppose that \(\mathcal{C}\) is a linear divergence (in fully-expanded form). Instead of using the derivative-dominant ranking, integrate by parts, working down the total order \(|\mathbf{J}|\) of the derivatives \(u^\alpha_{\mathbf{J}}\). For a given total order, we will invert the mixed derivatives first, though this is not essential. Starting with the highest-order derivatives, one could integrate each term \(f(\mathbf{x})u^\alpha_{\mathbf{J}}\) in \(\mathcal{C}\) by parts with respect to any \(x^i\) such that \(j^i\geq 1\), yielding the remainder \(-D_i(f)u^\alpha_{\mathbf{J}-\boldsymbol{1}_i}\). If \(D_\mathbf{J}\) is a mixed derivative, we seek to choose \(D_i\) in a way that keeps the result concise. Here are some simple criteria that are commonly effective, listed in order of precedence.
1. \(f\) is independent of \(x^i\).
2. \(\mathcal{C}\) includes the term \(D_i(f)u^\alpha_{\mathbf{J}-\boldsymbol{1}_i}\).
3. \(f\) is linear in \(x^i\).
These criteria can be used as an initial pass to invert \(\mathcal{C}\) at least partially, leaving a remainder to be inverted that may have far fewer terms than \(\mathcal{C}\) does. Integrating the remainder by parts is straightforward if each \(u^\alpha_{\mathbf{J}}\) is an unmixed derivative. If \(\mathbf{J}\) denotes a mixed derivative, integrate with respect to each \(x^i\) such that \(j^i\geq 1\) in turn, multiplying each result by a parameter (with the parameters summing to \(1\)). Iterate until either the remainder has a factor that is zero for some choice of parameters or there is no remainder. The final stage is to choose the parameters so as to minimize the number of terms in the final expression. Although this produces a minimal inversion, it comes at the cost of extra computational time spent doing all possible inversions followed by the parameter optimization.
To summarize, one can invert a divergence that is linear in \([\mathbf{u}]\) by applying the following procedure.
## Summary: a procedure for inverting Div
Having addressed potential modifications, we are now in a position to summarize the inversion process for any total divergence \(\mathcal{C}\) whose fully-expanded form has no terms depending on \(\mathbf{x}\) only.
The Appendix lists inversions of various other divergences; in each case, the inversion produced by Procedure [\[inversion\]](#inversion){reference-type="ref" reference="inversion"} and the ranking heuristic is minimal and takes very few iterations to complete.
## Splitting a divergence using discrete symmetries
A given divergence may have discrete symmetries between various terms in its fully-expanded form. If the divergence has very many terms that are connected by a particular discrete symmetry group, it can be worth splitting these into disjoint divergences that are mapped to one another by the group elements. Then it is only necessary to invert one of these divergences, using the symmetries to create inversions of the others without the need for much extra computation. However, to use Procedure [\[inversion\]](#inversion){reference-type="ref" reference="inversion"}, it is necessary to check that all split terms are grouped into divergences; this check is done by using the criterion [\[divcrit\]](#divcrit){reference-type="eqref" reference="divcrit"}. Polynomial divergences can first be split by degree, yielding divergences that are homogeneous in \([\mathbf{u}]\). Such splitting does not add significantly to the computation time, nor does it need to be checked using [\[divcrit\]](#divcrit){reference-type="eqref" reference="divcrit"}, which holds automatically. Splitting by degree can make it easy to identify terms that are linked by discrete symmetries, as illustrated by the following example.
# Concluding remarks
Partial Euler operators and partial scalings make it possible to invert divergences with respect to one independent variable at a time, a by-product being that some contributions to other components are determined at each iteration step. Although each iteration involves a fair amount of computation, very few iterations are needed for many systems of interest. Given the potential complexity of functions, it is unlikely that every divergence can be inverted, even in principle, by Procedure [\[inversion\]](#inversion){reference-type="ref" reference="inversion"}. The question of how to prove or disprove this is open. In practice, products of mixed derivatives present the greatest challenge to concise inversion, although the option of re-ranking part-way through the procedure enables a divide-and-conquer approach to be taken. The focus of this work has been on inverting the total divergence operator Div. However, this immediately applies to expressions that can be recast as total divergences. For instance, for \(p=3\), the total curl \(\mathbf{F}=\mathrm{Curl}(\mathbf{G})\) can be inverted by writing \[\label{Curlinv} F^i=\mathrm{Div}(H^{ij}\mathbf{e}_j)=D_jH^{ij},\qquad H^{ij}:=\epsilon^{ijk}G_k=-H^{ji},\] where \(\epsilon^{ijk}\) is the Levi--Civita symbol, then inverting one component at a time and using the results at each stage to simplify the remainder of the calculation. Once \(H^{ij}\) is known, the identity \(G_l=\frac{1}{2}\epsilon_{ijl}H^{ij}\) recovers \(\mathbf{G}\). Typically, a minimal inversion is achieved by using a different ranking for each \(F^i\), in accordance with the ranking heuristic. Here is a simple illustration of the general approach. | {'timestamp': '2022-12-19T02:11:19', 'yymm': '2212', 'arxiv_id': '2212.08455', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08455'} |
# Introduction
Understanding the formation and earliest evolution of star clusters involves the physics of the interstellar medium, stellar dynamics, stellar evolution, and the interplay between all three. Stars form over a period of a few million years, in close proximity of each other inside giant molecular clouds. There is a phase during which the evolving stars are exerting an influence on the gas through their radiation, winds, and eventually supernovae, all of which act to remove the gas from the cluster. Molecular clouds are turbulent, and the resulting initial stellar distribution is usually clumpy and quite sub-structured. Any stars that make up a bound structure will relax into a spherical system fairly quickly. The crowded environment in which stars form and spend their early lives could affect stellar properties such as the distribution of binary parameters, triples and higher order multiples, and the properties of planetary systems. We want to understand the processes that convert a cloud of gas into a spherical, bound, gas-free star cluster or a field of unbound stars, or both. The timescales for these processes, and the conditions which determine the fraction of the resultant stars in clusters, are important for understanding stellar populations in galaxies. One approach to answering these questions is observational. Detailed photometric studies of young local star-forming regions in the optical, infrared, and X-ray have been used to constrain the spatial distributions of young stellar objects, and the addition of Gaia data has allowed for wider explorations in the velocity realm to look at the expansion or contraction of the clusters. Timescales are inferred by comparing different regions and trying to use pre-main sequence isochrone fitting to date the systems. One limitation of using only the stars to understand early cluster properties is that at the earliest, deeply embedded stages, the gas is the dominant component of the region, and the stars are difficult to observe, especially in the optical and infrared. Even after much of the gas has been removed from the region, extinction can still be large, and therefore the (optical) Gaia coverage of such regions is poor, which reduces the number of stars for which we have reasonable proper motion measurements. Radial velocities are possible to obtain from infrared spectra, but obtaining spectra for many stars in a cluster is an expensive prospect with current instruments. Age, mass, and distance measurements of the stars are also difficult, because of extinction and the necessity of working in the infrared, but also because pre-main sequence stars are variable which affects their position in the HR diagram. Stellar observations also ignore the other major component in these forming clusters, namely the gas. Surveys of local star-forming regions in the far infrared, such as those from *Herschel*, use continuum dust emission to map out the column density distribution of material. This provides an estimate of the mass in the cloud. Molecular line observations in the (sub-)millimeter provide velocity information along the line of sight, temperature information, and column density. And because different lines trace different density ranges of gas, a combination of lines can nicely map out the gas distribution in the region. Unfortunately, there is almost no possibility of obtaining information about gas proper motions, and the depth of gas clouds can only be inferred in specific cases using density estimates. A complementary approach to answering questions about the earliest stages of star cluster formation is to use computational approaches. Simulations allow us to probe in great detail all the physical processes that are going on, and we can also follow the systems in three dimensions and in time to understand how they will evolve. However, the initial conditions for these kinds of simulations are almost always idealistic. Gas and star distributions are often spherical, usually in virial equilibrium, and with a density distribution that is straightforward to set up (e.g. constant density for gas although see for an exploration of more realistic distributions, and a Plummer or King model for stars). Simulations have traditionally been optimized for only one of the components--hydrodynamics for gas; N-body dynamics for stars--and an approximation has been included for the other. Hydrodynamics simulations often use sink particles to represent stars or regions which will form stars; N-body codes can include an analytic background potential to mimic the presence of gas. An increasing number of groups are using different techniques to improve the realism of simulations, such as including both stellar dynamics and hydrodynamics simultaneously, or using the results of hydrodynamics simulations as initial conditions for N-body runs. In order to be most useful, simulations must include all the appropriate physics for the question at hand, and of course they must start with appropriate initial conditions. Our hybrid approach is to bridge the divide between observations and simulations, and investigate the impact of more realistic, observationally-motivated initial conditions on our understanding of the formation and evolution of star clusters. In a previous paper we used observed positions of stars in embedded young clusters as a starting point for stellar dynamics plus hydrodynamics simulations. We showed that the initially clumpy sub-structure of the clusters evolved towards a more spherical distribution over a relatively short time. In those simulations, we included the presence of gas in a somewhat idealized way (following the clumpy distribution of the stars) and we showed that the total amount of gas present could have a significant impact on the subsequent evolution of the cluster. In this paper, we take the next step to improve the realism of this kind of simulation. In addition to the observed properties of stars, we include observed properties of the dense molecular gas in simulations of young embedded star clusters. Our goal is to explore realistic conditions for cloud evolution, as we expect that the idealized simulation environment that has been used in previous work may not correctly capture the impact of a range of gas densities and morphologies on the star formation rate and dynamical evolution of the forming star cluster. Specifically we look at the Orion region, near the Orion Nebula Cluster and its Integral Shaped Filament. We describe the methods by which we use the observational data to dictate the properties of our simulated gas, and highlight some tests we used to determine which choices of non-observed parameters were the most reasonable. Finally, we predict some likely outcomes of the star formation and dynamical evolution of the Orion region.
# Methods
## Computational Framework
For the simulations in this paper, we use the `Ekster` code. This code combines the SPH code `Phantom`, the \(N\)-body code `PeTar` and the stellar evolution code `SeBa` within the `AMUSE` framework. The main benefit of `Ekster` is that it allows for the formation of individual stars in a simulation that would otherwise not have the required resolution to form these directly from the gas. We restrict ourselves to an isothermal EOS in this study, and feedback effects (like stellar winds and ionising radiation) are not taken into account. Star formation in `Ekster` works similarly to the method described in. Once gas reaches a density \(\rho_{\rm crit}\) (here: \(10^{-18}\) g cm\(^{-3}\) or \(~ 1.5\times10^4\) M\(_{\odot}\) pc\(^{-3}\)), it will form a sink particle if the gas is collapsing. This sink will then accrete gas within a radius of 0.1 pc. A sink is considered as a source of star formation, and will form individual stars by probing a Initial Mass Function (IMF) until the next star is more massive than the remaining mass in the sink. We use the modified 'grouping' method described in, in which groups of sinks act as a combined star-forming region. This has the benefit over the original 'single sink' method of not requiring a high mass for each sink in order to probe the IMF and still form massive stars.
## Initial Conditions
In order to run a simulation, we need (at a minimum) 7 parameters for stars and 8 parameters for gas. For stars, we must specify their mass and 3-D positions and velocities; for gas particles, we need those same parameters and also a temperature or internal energy. However, the observations are incomplete in both cases--typically we can determine quantities only in the plane of the sky (e.g. positions from imaging, or velocities from proper motions), or only along the line of sight (e.g. velocities from spectra, or distances from parallax). In this section we describe which quantities we take directly from observations of the Orion region, and the choices we make for the other quantities. We note that the gas particles could be given additional properties, such as a chemical composition, dust-to-gas ratio, or a magnetic field. When such quantities are available from observations, they should also be included wherever possible. In this suite of simulations, we assume a constant mean molecular weight of 2.4 for our gas particles, no dust, and we neglect magnetic fields.
### Stars
The initial properties of the stars are chosen as described in, with some additions to extend the stellar distribution to larger distances. Here, we briefly summarize our method. We use stars from the Massive Young Stars in the Infrared and Xray (MYStIX) survey, and their clustering properties as described in. We use the observed positions of these stars (right ascension and declination) and take the distance to Orion to be 414 pc. This determines the stellar position in the x and y plane of our simulation. Most of these objects are not detected by Gaia, so the line of sight (z) position of each star is drawn randomly between-0.5 and 0.5 pc. The masses of the observed stars are drawn randomly from a IMF between 0.8 and 10 \(M_{\odot}\), corresponding to the observational limits of the survey. An appropriate number of lower mass stars (down to 0.2 \(M_{\odot}\)) were added to complete the IMF, and were distributed randomly within Plummer ellipsoids whose sizes on the sky were determined by the size and orientation of the MYStIX subclusters. We also included a sphere of lower mass stars to complete the IMF down to 0.2 \(M_{\odot}\) corresponding to the 'unclustered/unknown' component as identified in MYStIX. The MYStIX region only covers the central portion (r \(\approx\) 1 pc) of the area covered by the gas observations. We used stars identified by Spitzer in an annulus between 1 and 5 pc from the centre of the region to complete this region. We treated these stars in the same way as the MYStIX stars: identified Spitzer sources were assigned masses between 0.8 and 10 \(M_{\odot}\), and their line-of-sight position was randomly selected in the same range as the MYStIX stars. We then populate the IMF with the correct number (according to the adopted IMF) of lower mass stars, and randomly distributed them in this spherical shell. Finally, the Orion Cluster has 6 stars with masses greater than 10 \(M_{\odot}\). We used the observed positions and masses of those stars from, and assigned them z positions randomly distributed within 0.025 pc of z=0.2 (i.e. slightly in front of the gas, as suggested by ). We note that we are explicitly neglecting the binarity/multiplicity properties of stars in these simulations and assuming that all stars are single. The presence of binaries will modify the stellar dynamics in these simulations, but we do not expect them to have a large effect on the gas. All stars were then given a random velocity such that the overall velocity dispersion was 1 km s\(^{-1}\).
### Gas
To describe the initial gas conditions we use molecular line observations along the Integral Shape Filament in Orion. We use N\(_2\)H\(^+\) (1-0) observations tracing gas with densities higher than n(H\(_2\)) = \(10^{4.5}\) cm\(^{-3}\) together with ancillary NH\(_3\)-based gas temperatures estimates in this cloud taken from and, respectively, both Nyquist sampled at 30 arcsec resolution. There is also lower-density gas in the Orion region which we will include, but with properties not taken directly from observations. This combined dataset provides the column density, mass, temperature, line of sight velocity, and line of sight velocity dispersion for dense gas in pixels distributed along the filament. Each pixel is 0.03 pc (equivalent to 15 arcsec sampling) on a side at the distance of Orion. We wish to distribute gas particles in our simulation volume so that they have the same properties as the observations. We choose a mass for our gas particles (0.02 \(M_{\odot}\) in these simulations) and then determine how many such particles are needed to match the mass in each pixel. If a pixel contains less mass than our particle mass, we do not place a particle in the simulation. If the pixel contains enough mass for 1 particle then it is placed at the centre of the pixel in the z=0 plane. If more than one gas particle is required, they are placed randomly within the pixel (uniformly sampling in x and y), and the particles are given a random position in z between 0.15 pc and-0.15 pc. The total mass in dense gas is 1225 M\(_{\odot}\). The temperature of each particle is determined by the measured temperature in that pixel, which range between 10 K in the densest parts of the cloud away from the Trapezium and 137 K in the vicinity of the Orion BN/KL region. This temperature is used to determine the internal energy of the particle, assuming a mean molecular weight of 2.4. The velocities of these parcels of gas along the line of sight are determined by the observations. We remove a systematic mean motion of the observed gas (9.74 km s\(^{-1}\) away from us) and assign each particle the resultant velocity within its pixel along the z direction of our simulation. If there is more than one particle within a pixel, we modify their velocities so that they are drawn from a Gaussian distribution with a mean given by the gas velocity relative to the local standard of rest of the pixel, and a dispersion given by the velocity dispersion as measured in that pixel. We have no observed information about the velocity of the gas in the plane of the sky. Therefore, we need to make some choices in order to have a reasonable representation of the gas properties. Those choices, and the tests we performed to assess which ones are most reasonable, are described in more detail in the following section. The dense gas observations do not capture the entire interstellar medium in this region. The whole area is filled with relatively cool, low density gas as well. We add a sphere of gas, with a temperature of 30 K, a density of 500 particles per cm\(^3\), and a radius of 5 pc to our simulation. This results in a total of \(1.57 \times 10^4\) M\(_{\odot}\) of low-density gas. Specifically, we used the `molecular_cloud` routine within the [AMUSE]{.smallcaps} framework which includes a turbulent power spectrum of the form \(k^{-4}\) for the gas velocities. Our virial ratio, defined as the ratio of kinetic to potential energy, was set to 1.0, leading to a 3D velocity dispersion of about 4 km s\({^-1}\). We also varied many of these parameters of this gas sphere, and we will discuss the effect of other choices on our simulations below. Figure [\[fig:xy\]](#fig:xy){reference-type="ref" reference="fig:xy"} shows an image of the xy plane of our default choice of initial conditions (model vx_vy_0) at t=0. The colour scheme gives the column density of gas and the stars are plotted as white circles, with the circle size proportional to stellar mass. In all our simulations, the placement and properties of the stars are the same. The 1225 \(M_{\odot}\) of observed dense gas is modelled with almost 60 000 particles. In this case, the background sphere of gas is at a constant density of 30 \(M_{\odot} pc^{-3}\) and has a radius of 5 pc. As expected, the distribution of gas in the plane of the sky looks like a realistic filament of gas near a star cluster. Figure [\[fig:velocity\]](#fig:velocity){reference-type="ref" reference="fig:velocity"} shows the distribution of line-of-sight (z) velocities for particles in our dense gas filament (blue) and sphere of low-density gas (orange) compared to the observed line-of-sight velocities in the filament (green). The particles in the filament follow the observations closely, confirming that our initialization of those particles is correct to the extent constrained by the data. The particles in the sphere have velocities set by a turbulent power spectrum and, in this particular case, have a virial ratio of 1.0. The observed velocity structure of the filament is well-reproduced by our method, and shows more structure than the artificially-created low-density background cloud. That cloud does have a reasonable velocity dispersion, centred at zero by construction. In Figure [\[fig:PDF\]](#fig:PDF){reference-type="ref" reference="fig:PDF"} we show the probability density function (PDF) of the column density of the gas at t=0. This is a common quantity calculated for observed molecular clouds and star-forming regions. In observations, the column density PDF is often seen to take a log-normal shape, sometimes with a power law tail to high column densities. The blue histogram is the column density of the filament alone, and the orange histogram shows the column densities of the background cloud. Our simulated gas is the sum of these two. The green histogram shows the observed column density distribution, and at 100 \(M_{\odot} pc^{-2}\) and higher, the simulation correctly represents the observations. The excess of power in the simulations compared to the observations at lower densities is a product of our choice of a gas particle mass of 0.02 \(M_{\odot}\).
# Results
## Evolution of system with the default parameters
In Figure [\[fig:evolution_vxy0\]](#fig:evolution_vxy0){reference-type="ref" reference="fig:evolution_vxy0"} we show 1 Myr of evolution of our simulation with the initial conditions described above and shown in the previous figures. Very quickly, the dense gas in the filament collapses towards the centre ridge line of the filament, and stars begin to form. The filament gas, and also the recently formed stars, fall into the central star cluster. At the same time, the larger sphere of gas develops a density structure driven by the initial turbulent velocity spectrum. As some of that gas reaches high enough density, stars begin to form further away from the filament. Finally, the bulk of the low-density gas also falls into the central region of the simulation, and the stellar system becomes more spherical. Figure [\[fig:masses\]](#fig:masses){reference-type="ref" reference="fig:masses"} shows the mass of various simulation components as a function of time. Over the 1 Myr of this simulation, the gas is converted to stars to the point where the two contributions are approximately equal. The dense gas (all gas above \(10^3 M_{\odot}pc^{-3}\)) decreases slightly faster than the total gas mass. There is always a background of low-density gas that does not participate in star formation, which becomes a larger fraction of the gas mass as more and more stars are formed. We can see that the densest gas (above \(10^4 M_{\odot}pc^{-3}\), originally just in the filament) is quickly depleted (within the first 0.2 Myr) as the filament collapses, and the star formation rate peaks at a very high rate of 4000 \(M_{\odot}\) Myr\(^{-1}\). After a short pause, the background cloud's collapse replenishes this reservoir of dense gas, and in fact increases the amount available for star formation. By about 0.5 Myr, the amount of densest gas begins to level off, but the large amount available means that the star formation rate at this point in the cluster's evolution is high and increasing, becoming larger than 10 000 \(M_{\odot}\) Myr\(^{-1}\). Estimates of the average star formation rate in the solar neighbourhood are around 2500 \(M_{\odot}\) Myr\(^{-1}\) based on studies of star clusters, and less than 1000 \(M_{\odot}\) Myr\(^{-1}\) in local star-forming clouds. Our star formation rates are far too high, and the timescale of collapse of the filament (about 0.2 Myr or so) is extremely quick. In this particular model, the filament has no velocity in the plane of the sky so this collapse is not unexpected. We will explore other choices for that velocity in the next section.
## Free parameters in gas initial conditions
In this section, we describe in more detail the choices we made to fill in the missing information for the observed dense gas and for the low-density gas background. Each of the model parameters are summarized in Table [\[table:models\]](#table:models){reference-type="ref" reference="table:models"}. We changed two general properties--the velocity of the dense gas in the plane of the sky, and the parameters of the background sphere of gas. We have velocity information for the dense gas along the line of sight, but no information in the plane of the sky (the xy plane of our simulation). Therefore, we have tested three different possibilities, as described below.
1. In model vx_vy_0 we set the bulk x and y velocities of each particle to 0 km s\(^{-1}\). This is clearly not physical but provides a lower limit to the total amount of kinetic energy from the bulk velocity of the dense gas.
2. In model vx_vy_vz we set the net x and y velocities to the same value as the net z velocity in each pixel. This assumes that the filament is symmetric along all three axes.
3. In model vx_vy_grad, we use the gradient in line-of-sight velocities as a proxy for the velocities in the plane of the sky. For each pixel, the net line of sight velocity was compared to the line of sight velocity in the pixel directly above, and directly below (in declination, or the y direction), the pixel in question. The net y velocity for that pixel was taken to be the difference in line of sight velocities between those two pixels (i.e. the local gradient), and we preserved the direction of the gradient as well. If there was only information in one of the two pixels (either above, or below, but not both) then the gradient was calculated using the velocity of the pixel that we are considering and the one pixel above/below. If there was no information either above or below the pixel, the net y velocity was taken to be 0 km s\(^{-1}\). Similarly, the x velocity was taken using the gradient of the line-of-sight velocity in the pixels to the left and right of the pixel in question (i.e. in the right ascension direction).
Our default simulation sets the velocity dispersion in the background gas assuming it is in virial equilibrium (with itself only--not including the stars and the dense gas), i.e. with a ratio of kinetic to potential energy of 1.0 as in. We also ran a super-virial simulation (vx_vy_grad_2.0), with a ratio of kinetic to potential energy of 2.0. Finally, we also increased the size of the sphere from 5 pc to 10 pc. In the first case (vx_vy_grad_10pc) we kept the density of the gas the same but added more to the cloud (i.e. increased the mass by a factor of 8) and in a second case (vx_vy_grad_10pc_3.75) we kept the mass the same (i.e. reduced the density by a factor of 8). For comparison purposes, we have also set up a simulation in which the gas is initialized with a distribution more like the traditionally, idealized initial conditions for gas simulations. We removed the filament, and added the total amount of mass in the dense gas observations to a spherical cloud. We set the radius of this cloud to 10 pc, and gave it a virial ratio of 1.0. In Figure [\[fig:compare_sfr\]](#fig:compare_sfr){reference-type="ref" reference="fig:compare_sfr"} we plot the star formation rate as a function of time for all the simulations described above. The three simulations with difference choices of the gas velocity in the plane of the sky (orange dotted, purple dashed, and pink dot-dashed lines) have almost exactly the same star formation history, and therefore we conclude that the evolution of the system is relatively insensitive to any reasonable choice of velocity for the dense gas, likely due to its short free-fall time. We can see that simply adding more gas of the same density (red line) also changes very little at early times, but allows star formation to continue for longer and at an increasing rate. However, if we remove the filament of dense gas, and instead just follow the evolution of the background material of approximately the same high density in the 10 pc cloud (the 'idealized' run), we do not see an initial peak of star formation found in all the other runs. This tells us that the first burst of star formation comes from the collapse of the filament. The only simulation which does include the filament and also produces a reasonable continuous star formation rate of about 400 \(M_{\odot}\) Myr\(^{-1}\) is the one in which the filament is embedded in a lower-density larger sphere of material (green dashed line). Snapshots of the evolution of this simulation are shown in Figure [\[fig:evolution_vxygrad_10pc_375\]](#fig:evolution_vxygrad_10pc_375){reference-type="ref" reference="fig:evolution_vxygrad_10pc_375"}, and its column density PDF is shown in Figure [\[fig:PDF_best\]](#fig:PDF_best){reference-type="ref" reference="fig:PDF_best"}. The background cloud is centred at lower column density, and so much less of the gas can reach the star-forming densities even as the cloud collapses. We conclude that the larger-scale environment of the star-forming region is critically important to its subsequent evolution. The presence of the dense filament is important to drive the first burst of star formation that we see, but the properties of the background gas have more of an effect on the star formation rate at later times. We point out that we have only tested a few parameters of a deliberately simple background of gas, and we expect that a more realistic gas distribution, preferably taken from observations but also those motivated by larger-scale simulations, would be important to include. We also note that improvements to the physics of the simulations, such as including magnetic fields, are likely to also change the evolution of the system.
# Evolution of stellar system and impact of gas properties
Ultimately, the goal of this work is to model the formation of stellar clusters as they emerge from their natal gas clouds. We know that open clusters are gas-free and more or less spherical, virialized systems. The structure of younger systems is more complicated and clumpy in both space and time. There is observational evidence for age spreads and gradients in embedded clusters, and simulations, including those described above, also predict that star formation can last at least a few Myr. In fact, multiple stellar populations have been identified inside the ONC region. That spatial and temporal clumpiness can have implications for how larger clusters are built up, for stellar mass segregation, for the binary and planetary systems, etc. The evolution of embedded clusters will also depend on the distribution of gas within or near the stellar systems. Therefore, we are interested in characterizing the stellar properties in our simulations. To this end, we use the package INDICATE[^1] (INdex to Define Inherent Clustering And TEndencies; @2019A&A...622A.184B) which is a 2D+ local statistical tool, to assess and quantify the degree of spatial clustering of stars in our simulation. INDICATE assigns a unit-less index \(I\) to each star \(j\), defined as the actual number of neighbours \(N_{\bar{r}}\), and the expected number of neighbours \(N\) if the star were not clustered, within fixed radius \(\bar{r}\). Following we employ a value of \(N=5\) for our analysis below, i.e. \[\\ I_{j,5}=\frac{N_{\bar{r}}}{5}.\] Higher values of \(I_{j,5}\) represent greater degree of spatial association, and the index is calibrated against random distributions to define the 'significance threshold' \(I_{sig}\)-that is-the minimum value which denotes a star is spatially clustered. Extensive statistical testing has shown there is no dependency between the index and cluster shape, size or stellar density, and it is robust against edge effects, outliers and sample incompleteness up to \(83.3\%\) (@2019A&A...622A.184B, @2022A&A...659A..72B). As a local measure of spatial association, INDICATE can also reliably identify and quantify signatures of mass segregation in clusters, defined as either the concentration of (I) high mass stars together, typically at a cluster's centre or (II) low/intermediate mass stars around high mass stars. Potentially signatures of both Type I and II can be present in a cluster but the assessment is independent i.e. if low/intermediate mass stars are in concentrations around high mass stars, this does not necessarily mean that high mass stars are also concentrated together within a cluster, and vice versa. We applied INDICATE to the simulation shown in figure [\[fig:evolution_vxygrad_10pc_375\]](#fig:evolution_vxygrad_10pc_375){reference-type="ref" reference="fig:evolution_vxygrad_10pc_375"} in time step intervals of 0.1 Myr to follow the spatial evolution of stars. As expected, there is a change in the behaviour of stars as the system evolves. Initially stars are loosely concentrated about the centre of the cluster with a median index value of 411.8 (\(I_{sig}=44.5\)) for the entire population of which \(69.9\%\) are categorized as spatially clustered (\(I_{j,5}>I_{sig}\)). A steady increase then occurs in both the number of stars that are clustered, and the degree of association, such that by 1 Myr the median index for the population has increased to 1206.8 (\(I_{sig}=44.6\)) of which \(84.3\%\) are categorized as spatially clustered. This suggests that as the region evolves, not only is a greater proportion of its population likely to become clustered, but those stars which are clustered will become more spatially concentrated than at present after 1 Myr. This is the expected behaviour of a region which is undergoing the global collapse shown in figure [\[fig:evolution_vxygrad_10pc_375\]](#fig:evolution_vxygrad_10pc_375){reference-type="ref" reference="fig:evolution_vxygrad_10pc_375"}. We employed a two sample Kolmogorov-Smirnov Test (2sKST) with a strict significance boundary of \(p < 0.01\) to assess the significance of the differences the found disparities in the spatial behaviours at 0.1 Myr and 1 Myr. We conclude that while there is an observed change in the spatial behaviours of stars in the simulation between 0.1 Myr and 1 Myr, the nature of the behaviour remains similar. Examination of the index distributions (Figure [\[Fig_indicate_results\]](#Fig_indicate_results){reference-type="ref" reference="Fig_indicate_results"}) clarifies this result. Throughout the cluster's evolution the index distribution is bimodal with the first peak occurring at increasingly high values corresponding to the most central stars. In contrast, the second peak occurs at a lower, approximately constant value. Values in this second peak which are greater than the significance threshold correspond to central stars at relatively larger radii than in the first peak, and below the threshold is the non-core stellar population. This suggests that this region is contracting into a more centrally condensed configuration and will retain three spatially distinct populations: (i) a central core of stars that are very tightly clustered, encased in a (ii) 'halo' of stars that are more loosely clustered, and (iii) a dispersed population at larger radii. To determine if the stellar component of the simulation is contracting we express their velocities \(v_{*}\) in terms of their 3D outward component with respect to the centre of the system follow the prescription of: \[\\ v^{*}_{out}=\vec{v}_{*}\cdot \hat{r},\] where \(\hat{r}\) is the 3D unit vector of velocity which is directed outward from the system centre. The median of these velocities \(V_{out}\) indicates whether the cluster is expanding (\(>0\) km s\(^{-1}\)) or contracting (\(<0\) km s\(^{-1}\)). Figure [\[Fig_vout\]](#Fig_vout){reference-type="ref" reference="Fig_vout"} shows \(V_{out}\) and the mean 1D directional velocities as a function of cluster age in the top and bottom panels respectively. Clearly the simulation is undergoing overall contraction at all time steps, but the 1D velocities reveal that this is occurring non-uniformly at different rates along each axis and periodically expanding in one direction. We can also characterize the spatial distribution in the cluster by calculating the 3-dimensional Q parameter at each time step (Figure [\[Fig_Qval\]](#Fig_Qval){reference-type="ref" reference="Fig_Qval"}) (where a value of \(Q<0.7\) indicates a fractal distribution, \(Q>0.7\) radial, and \(Q\approx0.7\) randomised). We confirm that the initial distribution of stars in the simulation is a fractal distribution ( \(D=2.0\)), but by 0.5 Myrs a radial distribution has been achieved that progressively strengthens through to the end point of the simulation(at which time the fractal dimension has reduced to \(D=1.6\)). In particular, the high mass population (HM) is spatially evolving into an increasingly strong central concentration distribution. We find signatures of Type I mass segregation in every snapshot with the majority of high-mass members concentrated together above spatial randomness at the simulation centre. The degree of the segregation increases as the cluster evolves, both in fraction of stars, from \(56.1\%\) at 0.1 Myr, increasing to \(79.8\%\) by 1 Myr; and clustering strength of segregated stars, from \(\tilde{I}_{5} = 84.8\) (\(I_{sig}=44.3\)) at 0.1 Myr, increasing to \(124.8\) (\(I_{sig}=44.1\)) by 1 Myr. There is also a steady decrease observed in \(\frac{r_{max}^{HM}}{r_{max}}\) from 0.31 (0.1 Myrs) to 0.29 (1.0 Myrs). High mass stars are not found to be in stronger stellar concentrations in the general population than their lower mass counterparts at any point in the evolution (no signatures of Type II mass segregation are present). We consider the typical spatial conditions under which new stars are formed in the cluster. Are stars forming in regions of progressively tighter stellar concentrations as the cluster evolves or are they forming in lower density regions and migrating to ones of higher stellar concentration? For each time step we compute the median index values of stars that formed since the last timestep with the pre-existing population in the cluster at that time. Prior to \(\sim\)`<!-- -->`{=html}0.6 Myr stars are forming in regions of lower stellar concentrations than typical of the older cluster population, and after that time the opposite is true. Again we run a 2sKSTs to compare the indexes of newly formed and pre-existing stars at each timestep with strict significance boundaries of \(p < 0.01\). In all snapshots, stars that formed since the previous timestep have a distinctly different distribution to that of the pre-existing (older) population. We calculate the distance percentile at which new stars are forming-that is, the mean radius from the cluster centre where star formation is occurring expressed as a function of the radial distribution of the pre-existing population (Figure [\[Fig_indicate_newwhere\]](#Fig_indicate_newwhere){reference-type="ref" reference="Fig_indicate_newwhere"}). Initially stars form toward the outer boundary of the cluster in less populous regions, then later form with increasing centrality and in more populous regions. As stars form in the simulation when the critical gas density criteria for star formation is met, this observed behavioural change is directly attributable to the infall of gas in the system: initially stars form less centrally where the gas has a high enough density for star formation to occur, but stellar density is low. Then as the system evolves, gas in falls towards the cluster's centre so a critical gas density is achieved at relatively smaller radii to the simulation centre, where stellar density is higher. Similarly, early on in the simulation a band of stars in the X-Y plane has unusually high clustering tendencies (w.r.t. the rest population in the same region), which corresponds to the location the dense gas filament shown in Figure [\[fig:evolution_vxygrad_10pc_375\]](#fig:evolution_vxygrad_10pc_375){reference-type="ref" reference="fig:evolution_vxygrad_10pc_375"}.
# Summary and Discussion
We use observations of dense star-forming gas to guide our initial conditions of star cluster formation simulations. This work represents an improvement on our previous efforts using observed stellar properties with only assumed gas properties. We outlined a method to convert the observed gas information suitable for a particle-based hydrodynamics approach. We explored options for the quantities that are not available from the observations, particularly the velocity structure of the dense gas in the plane of the sky, and some assumptions for a wider background of gas. We find that most reasonable assumptions for the velocity of the filament result in similar overall start formation rates and evolution of the region. We note that the presence of the filament, in any form, does have a significant effect on the star formation rate and overall structure of the forming star cluster. The wider gas environment, however, which we modelled as a spherical configuration of less dense gas, is quite important. In our models, only the largest cloud with the lowest density produces star formation rates which are in line with observations of nearby star-forming regions. We suggest that observations should be used to constrain all the gas in a simulation (for example, using *Herschel* observations of intermediate-density gas instead of an idealized cloud). Based on our simulations we argue that the Orion region will most likely continue to form stars, at first along the dense gas filament and later as the background gas falls into the central cluster, and at the same time, the entire region will contract into a stronger radial concentration. The INDICATE results suggest that the degree of association in Orion will increase, but that the currently observed three spatially distinct populations (a tightly clustered core, surrounded by a loosely clustered halo, and an outer dispersed population) would be retained. Our prediction from this work is that it is reasonable to expect Orion will become a spherical, likely gas-free cluster, within a relatively short period of time (\(\sim\) 1 Myr) from the present. In our simulations, we restricted ourselves to single choices of some physical parameters, such as the equation of state and the prescription of star formation. We have also neglected any feedback from the stars onto the gas including stellar winds, ionizing radiation, and protostellar jets which are expected to impact the star formation rate of low-mass clouds significantly before the first supernova can explode. Given the sensitivity of our results to the structure of the background cloud of gas, combined with the relative insensitivity of our results to the treatment of the dense gas, we suggest that the physics that is included in star cluster formation simulations may be at least as important, and possibly more so, as the initial conditions. Nevertheless, we have also shown that the presence of the dense filament of gas affects the early star formation rate. It is incumbent on simulators to use as much direct information from observations as possible to improve the realism of our work. In the particular context of the formation and early evolution of star clusters, the distribution and velocity structure of the natal gas at all densities are the key properties that should be determined or guided by observations. Simulations of astrophysical objects and processes necessarily begin, in their earliest incarnations, with simple initial conditions. A turbulent sphere of gas with a constant density has only a few free parameters and it is straightforward to interpret the results from such a simulation. However, at some point we have learned all we can from the simple set-ups and we must turn to more complicated initial conditions. That opens up a wealth of possibilities and choices of parameters. There is a very significant need for those parameters to be constrained by observations, as otherwise the parameter space is simply too large to explore, and much of it will simply not be relevant. We have shown here that it is not very difficult to go one step even further, and to directly use the observations of gas and stars to dictate initial conditions for simulations. This approach may not be applicable to all simulations or all astrophysical objects, but it provides a useful additional method of investigating the universe. | {'timestamp': '2022-12-19T02:11:03', 'yymm': '2212', 'arxiv_id': '2212.08440', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08440'} |
# Introduction
Compared with the existing massive multiple-input-multiple-output (MIMO) with dozens or hundreds of antennas for 5G, extremely large-scale MIMO (XL-MIMO) embraces a 10-fold increase in the scale of antenna array, which enables the considerable improvement of the spatial multiplexing gain . Moreover, benefiting from the high beamforming gain achieved by extremely narrow pencil-like beams, XL-MIMO is essential to alleviate the serious path loss of millimeter wave (mmWave) and terahertz (THz) signals with abundant spectrum resources . Consequently, the natural integration of high-frequency wideband communications and XL-MIMO is regarded as a promising technology in future 6G networks. To unleash the performance superiority of wideband XL-MIMO, it is essential to acquire accurate channel state information (CSI). Unfortunately, due to the enormous number of antennas, XL-MIMO channel estimation requires an unacceptable pilot overhead, particularly when the base station (BS) is deploying the hybrid analog and digital precoding architecture with only a small number of radio-frequency (RF) chains . Thereby, how to exploit the channel structure to reduce the pilot overhead for channel estimation has always been a hot topic from 5G massive MIMO to 6G XL-MIMO.
## Prior Works
There are extensive works to explore the sparse structure of mmWave and THz channels to design channel estimation algorithms with low pilot overhead. These works can be generally categorized into three types: far-field narrowband , far-field wideband , and near-field narrowband channel estimations. The first type of *far-field narrowband* channel estimation schemes mainly works for 5G massive MIMO, and they usually assume that the bandwidth is limited, e.g., around several hundreds of MHz. The array aperture of massive MIMO is not large enough, resulting in the dominance of far-field transmission environments, where a channel path can be modeled by the planar wave model. In this case, the multipath far-field channel is built up by the superposition of several planar waves . As each planar wavefront is dependent on a certain angle-of-arrival (AoA), and the number of channel paths is usually small, this far-field channel exhibits sparsity in the angle domain. Thus it can be recovered by compressed sensing (CS) based algorithms with low pilot overhead. For instance, utilized orthogonal matching pursuit (OMP) scheme to estimate the angle-domain sparse channel at a single frequency. Moreover, several works have considered to improve the OMP algorithm by using the channel sparse structure at multiple frequencies. Specifically, due to the limited bandwidth, the common sparse support assumption has been widely used to transform the multi-frequency channel estimation into a multiple measurement vector (MMV) problem, where the sparse support sets are the same at different frequencies. Then, various MMV-specific algorithms can be utilized to estimate the far-field channel, such as the simultaneously OMP (SOMP) and the group-sparse Bayesian CS approach . Compared with the *far-field narrowband* schemes above, the *far-field wideband* channel estimation algorithms assume a pretty wide bandwidth. In this condition, although communications are still operating in far-field environments, the widely-used common sparse support assumption does not hold anymore due to the beam split effect . The beam split effect indicates that owing to the wide bandwidth, the spatial channel directions are separated from each other at different frequencies, which results in a frequency-dependent angle-domain sparse structure and severely degrades the performance of classical MMV-specific solutions. Recently, a few advanced signal processing techniques have been proposed to cope with this beam split effect in far-field scenarios. Specifically, proved the beam split pattern in the angle domain, i.e., the sparse support set of a physical angle grows linearly with the subcarrier frequency. Then, a beam split pattern detection (BSPD) based algorithm was proposed in to construct the match between a certain physical angle and the beam split pattern for recovering the far-field wideband channel. Besides, in , the authors proposed to construct a series of dictionary matrices for each subcarrier frequency to match the beam split pattern, based on which a generalized simultaneous OMP (GSOMP) was applied to overcome the beam split effect in the far field. For the third category of the *near-field narrowband* algorithms, they place heavy emphasis on the fundamental change from 5G massive MIMO to 6G XL-MIMO by considering the near-field propagation. The deployment of XL-MIMO makes the near-field propagation highly possible to happen, particularly for high-frequency communications . To be specific, the radius of near-field areas, which is also called as Fraunhofer distance, is proportional to the square of array aperture and frequency . The Frauhofer distance may reach several dozens or hundreds of meters in high-frequency XL-MIMO systems. As opposed to the far-field channel model based on the planar wavefront assumption, the multi-path near-field channel model is built up by the superposition of multiple *spherical waves* . Although the number of channel paths is still limited, the angle-domain near-field channel is not sparse any more, since each spherical wave is dependent on a range of physical angles . Thus, existing angle-domain sparsity based techniques cannot work well in near-field environments. To address this problem, proposed to transform the near-field channel into the joint angle-distance domain (polar domain). As an angle-distance sample contains enough information to describe a spherical wave, the near-field channel can be then sparsely represented in the polar domain. Accordingly, the polar-domain SOMP algorithms were used in to recover the near-field channel accurately. Despite the extensive studies on XL-MIMO channel estimation, the above three categories of techniques neglect the crucial topic of *near-field wideband* channel estimation. In other words, they overlook the important integration of the near-field and beam split effect. For wideband XL-MIMO communications, both the array aperture and bandwidth are pretty large. Such a wideband XL-MIMO system poses a *near-field beam split* effect , where the sparse channel support set in the polar domain differs from each other at different frequencies. Furthermore, as all prior techniques (far-field narrowband, far-field wideband, and near-field narrowband approaches) mismatch this polar-domain frequency-dependent sparse structure, they will suffer from a serious performance loss in wideband XL-MIMO systems. To the best of our knowledge, the essential near-field wideband channel estimation has not been studied in the literature.
## Our Contributions
To fill in this gap, we propose a bilinear pattern detection (BPD) based algorithm to realize near-field wideband channel estimation by exploring the unique polar-domain sparse structure with near-field beam split. Our contributions are summarized as follows.
- First, we comprehensively analyze the polar-domain channel sparse structure over different subcarrier frequencies when the near-field beam split effect is considered. By analyzing the mutual coherence of near-field array response vectors at different frequencies, we reveal the phenomenon that both the angle-domain and distance-domain sparse support sets grow linearly with the subcarrier frequency, which is defined as the bilinear pattern (BP) of near-field beam split effect. This phenomenon indicates that the physical location of a scatter (or user) in a channel path has a one-to-one matching relationship with a certain pattern of the frequency-dependent sparse support sets in the polar domain. This relationship can be used to accurately estimate the physical location of each path.
- Then, a BPD based channel estimation algorithm is proposed according to the revealed phenomenon above. We first construct the polar-domain bilinear patterns for all angle-distance samples. Then, for each path component, we use these bilinear patterns to accumulate the power of each angle-distance sample from the entire bandwidth and capture the sample with the largest power as the estimated physical location. We carry out the above procedure several times to successively detect all channel paths. As the proposed algorithm can jointly exploit the polar-domain sparsity and the bilinear pattern, it is promising to achieve accurate channel estimation.
- Finally, we analyze the computational complexity of the BPD based algorithm and compare it with existing algorithms. Extensive simulation results are provided to demonstrate the superior performance of the proposed scheme. It is illustrated that the BPD based channel estimation technique outperforms existing schemes. More importantly, it is verified that the BPD based algorithm is capable of realizing accurate channel estimation in all far-field/near-field/narrowband/wideband conditions[^1].
*Organization*: The remainder of this paper is organized as follows. Section [2](#sec:2){reference-type="ref" reference="sec:2"} presents the system model, including the channel model and problem formulation. In Section [3](#sec:3){reference-type="ref" reference="sec:3"}, the bilinear pattern is proved, and the BPD based algorithm is provided. Simulations are carried out in Section [4](#sec:4){reference-type="ref" reference="sec:4"}, and finally conclusions are drawn in Section [5](#sec:5){reference-type="ref" reference="sec:5"}.
# System Model {#sec:2}
## Channel Model
In this paper, we investigate the uplink channel estimation for XL-MIMO system with orthogonal frequency division multiplexing (OFDM), where one base station (BS) uses an \(N\)-antenna uniform linear array to serve \(K\) single-antenna users. We adopt the Saleh-Valenzuela multipath channel to model the near-field channel . As presented in Fig. [\[fig:channel\]](#fig:channel){reference-type="ref" reference="fig:channel"}, the frequency-domain channel \(\mathbf{h}_{m} \in \mathbb{C}^{N \times 1}\) at the \(m\)-th subcarrier (\(m \in \{0, 1, \cdots, M-1\}\)) from a certain user is denoted as \[\begin{aligned}
\label{eq:Channel} \mathbf{h}_{m} = \sqrt{\frac{N}{L}}\sum_{l = 0}^{L-1} g_{l,m}e^{-j\frac{2\pi}{\lambda_m} r_l }\mathbf{a}(\vartheta_l, r_l, f_m). \end{aligned}\] Here, \(g_{l,m}\) denotes the complex path gain at \(m\)-th subcarrier of the \(l\)-th path, and \(L\) represents the number of channel paths. Besides, \(f_m = f_c + \frac{2m-M}{2M}B\) is the \(m\)-th subcarrier frequency with \(B\) and \(f_c\) corresponding to the bandwidth and the carrier frequency, and \(\lambda_m = \frac{c}{f_m}\) is the wavelength at \(f_m\) with \(c\) being the light speed. Moreover, as shown in Fig. [\[fig:channel\]](#fig:channel){reference-type="ref" reference="fig:channel"}, \(\vartheta_l\) represents the AOA of the \(l\)-th path, \(r_l\) is the distance from the last-hop scatter to the BS array's center, and \(\mathbf{a}(\vartheta_l, r_l, f_m)\) is the array response vector of \((\vartheta_l, r_l)\) describing the spherical wavefront between the last-hop scatter and BS. According to the geometrical structure of BS array, \(\mathbf{a}(\vartheta_l, r_l, f_m)\) is modeled as \[\begin{aligned}
\label{eq:ArrayResponseVector} \mathbf{a}(\vartheta_l, r_l, f_m) = \frac{1}{\sqrt{N}}[e^{j\phi_{l,m}^{(0)}}, e^{j\phi_{l, m}^{(1)}}, \cdots, e^{j\phi_{l, m}^{(N-1)}}]^T, \end{aligned}\] where \(\phi_{l, m}^{(n)} =-\frac{2\pi}{\lambda_m} (r_l^{(n)}-r_l)\) and \(r_l^{(n)}\) is the distance from the last-hop scatter to the \(n\)-th BS antenna. As indicated in , \(r_l^{(n)}\) can be written as \(r_l^{(n)} = \sqrt{r_l^2 + \delta_n^2d^2-2 r_l \delta_nd \sin\vartheta_l}\), where \(\delta_n = n-\frac{N-1}{2}\) with \(n \in \{0,1,\cdots, N-1\}\). In addition, \(d = \frac{\lambda_c}{2}\) is the antena spacing and \(\lambda_c = \frac{c}{f_c}\) is the carrier wavelengh. From ([\[eq:ArrayResponseVector\]](#eq:ArrayResponseVector){reference-type="ref" reference="eq:ArrayResponseVector"}), two crucial aspects are revealed: the near-field propagation and the beam split effect. First, the far-field scenario is always adopted in conventional massive MIMO to simplify \(\phi_{l,m}^{(n)}\) as \(\frac{2\pi}{\lambda_m}\delta_n d \sin \vartheta_l\), which is independent of distance \(r_l\). On the contrary, in XL-MIMO systems, the Fraunhofer distance \(\frac{2D^2}{\lambda_c}\) is comparable to \(r_l\), making it highly possible for signals to propagate in near-field environments . Here, \(D = Nd\) is the array aperture. For example, we consider a 0.3-meter-array working at 100 GHz, the Fraunhofer distance of which reaches 60 meters. In this case, the spherical wavefront model has to be investigated and the impact of \(r_l\) cannot be ignored . Second, in narrowband communications when \(f_m \approx f_c\), the array response vector is nearly frequency-independent, leading to the common sparse structure. This structure allows jointly estimating channels at different subcarriers by using MMV-specific CS algorithms, such as SOMP and the group-sparse Bayesian approach . However, in wideband communications when \(f_m \neq f_c\), the array response vector considerably varies over frequencies. More seriously, the support set of different subcarriers can be quite different from each other in wideband XL-MIMO systems, undermining the common sparse support structure of the MMV model. This phenomenon is defined as the beam split effect , which severely degrades the performance of conventional channel estimation schemes. To the best of our knowledge, existing channel estimation works fail to jointly exploit these two underlying characteristics of the wideband XL-MIMO channel, giving rise to poor estimation performance. To fill in this gap, this paper simultaneously considers the near-field propagation and beam split effect (or the near-field beam split effect) to realize high-accuracy channel estimation..
## Problem Formulation
As indicated in Fig. [\[fig:system\]](#fig:system){reference-type="ref" reference="fig:system"}, we adopt an uplink time-division-duplex (TDD) channel estimation scenario, where \(K\) users harness orthogonal time or frequency resources to transmit orthogonal pilots to BS. Thus we can consider channel estimation for an arbitrary user without loss of generality. We suppose the length of pilot sequence is \(P\). Denote \(s_{m, p}\) as the pilot signal at the \(m\)-th subcarrier in the \(p\)-th time slot. Then, the received signals \(\mathbf{y}_{m, p} \in \mathbb{C}^{N_{\text{RF}} \times 1}\) can be presented as \[\begin{aligned}
\mathbf{y}_{m, p} = \mathbf{A}_p \mathbf{h}_{m} s_{m, p} + \mathbf{A}_p \mathbf{n}_{m,p}, \end{aligned}\] where \(\mathbf{A}_p \in \mathbb{C}^{N_{\text{RF}} \times N}\) denotes the combining matrix and \(\mathbf{n}_{m,p} \in \mathbb{C}^{N \times 1}\) denotes the complex Gaussian noise following the distribution \(\mathcal{CN}(0, \sigma^2 \mathbf{I}_N)\). Assume \(s_{m, p} = 1\) for \(p = 1,2,\cdots, P\). Then, considering the total \(P\) time slots for pilot trasmission, we arrive at \[\begin{aligned}
\mathbf{y}_{m} = \mathbf{A} \mathbf{h}_{m} + \mathbf{n}_m, \end{aligned}\] where \(\mathbf{y}_m = [\mathbf{y}_{m, 1}^T, \mathbf{y}_{m, 2}^T, \cdots, \mathbf{y}_{m, P}^T]^T \in \mathbb{C}^{PN_{\text{RF}} \times 1}\) and \(\mathbf{n}_m = [\mathbf{n}_{m,1}^T \mathbf{A}_1^T, \mathbf{n}_{m,2}^T \mathbf{A}_2^T, \cdots, \mathbf{n}_{m,P}^T \mathbf{A}_P^T]^T \in \mathbb{C}^{PN_{\text{RF}} \times 1}\). Besides, \(\mathbf{A} = [\mathbf{A}_{1}^T, \mathbf{A}_{2}^T, \cdots, \mathbf{A}_{P}^T]^T \in \mathbb{C}^{PN_{\text{RF}} \times N}\) is the overall observation matrix, whose elements can be independently generated from the binomial distribution \(\frac{1}{\sqrt{N}}\{-1, 1\}\). To recover the high-dimensional near-field channel \(\mathbf{h}_{m}\) from the low-dimensional observation \(\mathbf{y}_m\), where \(PN_{\text{RF}} < N\), we can transform the antenna-domain channel to its angle-distance domain (polar domain) by a polar-domain representation matrix \(\mathbf{W} \in \mathbb{C}^{N \times N_aN_d}\) . As discussed in , \(\mathbf{W}\) is composed of \(N_d\) submatrices, where \[\begin{aligned}
\mathbf{W} = [\mathbf{W}_1, \mathbf{W}_2, \cdots, \mathbf{W}_{N_d}]. \end{aligned}\] Each submatrix \(\mathbf{W}_{n_d} \in \mathbb{C}^{N \times N_a}\) includes \(N_a\) array response vectors at the carrier frequency, where \[\begin{aligned}
\label{eq:Wn} \mathbf{W}_{n_d} = [\mathbf{a}(\bar{\vartheta}_1, \bar{r}_{n_d, 1}, f_c), \mathbf{a}(\bar{\vartheta}_2, \bar{r}_{n_d, 2}, f_c), \cdots, \mathbf{a}(\bar{\vartheta}_{n_a}, \bar{r}_{n_d, N_a}, f_c)]. \end{aligned}\] As discussed in , the sampled angles and distances, i.e., (\(\bar{\vartheta}_{n_a}, \bar{r}_{n_d, n_a}\)) with \(n_a \in [1, 2, \cdots, N_a]\) and \(n_d \in [1,2, \cdots, N_d]\), should satisfy \(\sin\bar{\vartheta}_{n_a} = \frac{2(n_a-1)-N_a}{N_a}\) and \(\bar{r}_{n_d, n_a} = \frac{D^2\cos^2\bar{\vartheta}_{n_a}}{2\beta^2 \lambda_c n_d}\), where \(\beta\) is a predefined parameter. Then, the polar-domain channel \(\bar{\mathbf{h}}_m\) can be obtained by solving the underdetermined equation \(\mathbf{h}_m = \mathbf{W} \bar{\mathbf{h}}_m\). As investigated in , matrix \(\mathbf{W}\) is capable of extracting the angle-distance information embedded in each array response vector of ([\[eq:Channel\]](#eq:Channel){reference-type="ref" reference="eq:Channel"}). In addition, the number of paths \(L\) is usually small, especially for mmWave and THz bands (e.g. \(L = 5\)). As a consequence, the polar-domain channel \(\bar{\mathbf{h}}_m\) is sparse and the near-field channel estimation problem is equivalent to a sparse signal recovery problem: \[\begin{aligned}
\label{eq:SparseRecover} \mathbf{y}_m = \mathbf{A}\mathbf{W}\bar{\mathbf{h}}_m + \mathbf{n}_m. \end{aligned}\] Finally, we attempt to jointly estimate channels for all subcarriers. Hence, we rewrite ([\[eq:SparseRecover\]](#eq:SparseRecover){reference-type="ref" reference="eq:SparseRecover"}) as \[\begin{aligned}
\label{eq:MMV} \mathbf{Y} = \mathbf{A}\mathbf{W}\bar{\mathbf{H}} + \mathbf{N}, \end{aligned}\] where \(\mathbf{Y} = [\mathbf{y}_1, \mathbf{y}_2, \cdots, \mathbf{y}_M] \in \mathbb{C}^{PN_{\text{RF}} \times M }\), \(\bar{\mathbf{H}} = [\bar{\mathbf{h}}_1, \bar{\mathbf{h}}_2, \cdots, \bar{\mathbf{h}}_M]\), and \(\mathbf{N} = [\mathbf{n}_1, \mathbf{n}_2, \cdots, \mathbf{n}_M]\). Notice that as the array response vector ([\[eq:ArrayResponseVector\]](#eq:ArrayResponseVector){reference-type="ref" reference="eq:ArrayResponseVector"}) is frequency-dependent, the sparse support set of \(\bar{\mathbf{h}}_m\) for different subcarriers \(f_m\) is quite different from each other. Therefore, existing wideband channel estimation algorithms tailored for MMV problems cannot perfectly match the model ([\[eq:MMV\]](#eq:MMV){reference-type="ref" reference="eq:MMV"}). Consequently, an accurate near-field channel estimation scheme for wideband XL-MIMO is essential.
# Bilinear Pattern Detection Based Channel Estimation {#sec:3}
In this section, we first reveal the bilinear pattern of near-field beam split effect. This pattern reveals that with proper preprocessing, the support set of near-field channels varies linearly over frequencies in both the angle and the distance domain. Then, we propose a bilinear pattern detection (BSD) based near-field wideband channel estimation scheme. Finally, we provide complexity analysis of our scheme.
## Bilinear Pattern of Near-Field Beam Split
To jointly recover the wideband channel for all subcarriers in ([\[eq:MMV\]](#eq:MMV){reference-type="ref" reference="eq:MMV"}), it is crucial to investigate the map between the support index and frequency for an arbitrary path, which is defined as the beam split pattern (BSP). This pattern has been extensively studied under *far-field* environments in the literature , while we are trying to discover the pattern of *near-field beam split*. Specifically, we investigate the \(l\)-th scatter located at \((\vartheta_l, r_l)\), which corresponds to a series of array response vectors \(\mathbf{a}(\vartheta_l, r_l, f_m)\) with different frequencies. According to the CS theory, the support index \((n_{a, l, m}^{\star}, n_{d, l, m}^{\star})\) of \(\mathbf{a}(\vartheta_l, r_l, f_m)\) refers to the index of the strongest element of \(\mathbf{W}^H\mathbf{a}(\vartheta_l, r_l, f_m)\), i.e., \((n_{a, l, m}^{\star}, n_{d, l, m}^{\star}) = \arg\max_{n_a, n_d} \|\mathbf{a}^H(\bar{\vartheta}_{n_a}, \bar{r}_{n_d, n_a}, f_c)\mathbf{a}(\vartheta_l, r_l, f_m)\|\). Unfortunately, as the phase \(\phi_{l, m}^{(n)} =-\frac{2\pi}{\lambda_m} (r_l^{(n)}-r_l)\) of the \(n\)-th element of \(\mathbf{a}(\vartheta_l, r_l, f_m)\) is a complicated radical function, it is intractable to obtain the close form of \((n_{a, l, m}^{\star}, n_{d, l, m}^{\star})\). To address this problem, we use the second-order Taylor expansion to approximate \(\phi_{l, m}^{(n)}\) as \[\begin{aligned}
\phi_{l, m}^{(n)} \approx \bar{\phi}_{l, m}^{(n)} = \frac{2\pi}{\lambda_m}(\delta_n d \sin\vartheta_l-\delta_n^2 d^2 \frac{\cos^2\vartheta_l}{2r_l}), \end{aligned}\] which is much more accurate than the conventional far-field approximation with first-order Taylor expansion. For expression simplicity, we use parameters \((\theta_l, \alpha_l) = (\sin\vartheta_l, \frac{\cos^2\vartheta_l}{2r_l})\) to replace parameters \((\vartheta_l, r_l)\), so we have \(\bar{\phi}_{l, m}^{(n)} = \frac{2\pi}{\lambda_m}(\delta_n d \theta_l-\delta_n^2 d^2 \alpha_l)\). As a result, we can use \(\bar{\mathbf{a}}(\theta_l, \alpha_l, f_m) = {\frac{1}{\sqrt{N}}}[e^{j\bar{\phi}_{l,m}^{(0)}}, \cdots, e^{j\bar{\phi}_{l, m}^{(N-1)}}]^T\) to simplify \(\mathbf{a}(\vartheta_l, r_l, f_m)\). Similarly, vector \(\mathbf{a}(\bar{\vartheta}_{n_a}, \bar{r}_{n_d, n_a}, f_c)\) can be approximated as \(\bar{\mathbf{a}}(\bar{\theta}_{n_a}, \bar{\alpha}_{n_d}, f_c)\). Notice that the sampled angles \(\bar{\vartheta}_{n_a}\) and distances \(\bar{r}_{n_d, n_a}\) are transformed to \(\bar{\theta}_{n_a} = \sin\bar{\vartheta}_{n_a} = \frac{2(n_a-1)-N_a}{N_a}\) and \(\alpha_{n_b} = \frac{\cos^2\bar{\vartheta}_{n_a}}{2\bar{r}_{n_d, n_a}} = \frac{\cos^2\bar{\vartheta}_{n_a}}{2 \frac{D^2\cos^2\bar{\vartheta}_{n_a}}{2\beta^2 \lambda_c n_d}} = \frac{\beta^2 \lambda_c n_d}{D^2}\), with \(n_a \in \{1, 2, \cdots, N_a\}\) and \(n_d \in \{1, 2, \cdots, N_d\}\). Based on the above discussion, the near-field beam split pattern can be acquired by solving the problem \((n_{a, l, m}^{\star}, n_{d, l, m}^{\star}) = \arg\max_{n_a, n_d} \|\bar{\mathbf{a}}^H(\bar{\theta}_{n_a}, \bar{\alpha}_{n_d}, f_c)\bar{\mathbf{a}}(\theta_l, \alpha_l, f_m)\|\). The following **Lemma 1** provides a more concise form of \(\|\bar{\mathbf{a}}^H(\bar{\theta}_{n_a}, \bar{\alpha}_{n_d}, f_c)\bar{\mathbf{a}}(\theta_l, \alpha_l, f_m)\|\). *Proof*: Accoding to the definition of \(\bar{\mathbf{a}}(\theta_1, \alpha_1, f_1)\) and \(\bar{\mathbf{a}}(\theta_2, \alpha_2, f_2)\), we have \[\begin{aligned}
\|\bar{\mathbf{a}}^H(\theta_1, \alpha_1, f_1)\bar{\mathbf{a}}(\theta_2, \alpha_2, f_2)\| &= \left|\frac{1}{N}\sum_{n =-(N-1)/2}^{(N-1)/2}e^{j\frac{2\pi nd}{\lambda_1}(\theta_1-\frac{f_2}{f_1}\theta_2)- j\frac{2\pi n^2d^2}{\lambda_1}(\alpha_1-\frac{f_2}{f_1}\alpha_2)} \right| \notag \\ & \overset{(a)}{=} \left|\sum_{x =-\frac{1}{2} + \frac{1}{2N}}^{\frac{1}{2}-\frac{1}{2N}}e^{j\frac{2\pi Dx}{\lambda_1}(\theta_1-\frac{f_2}{f_1}\theta_2)- j\frac{2\pi D^2x^2}{\lambda_1}(\alpha_1-\frac{f_2}{f_1}\alpha_2)} \frac{1}{N}\right|, \label{eq:Summation} \end{aligned}\] where (a) is achieved by letting \(x = \frac{n}{N}\) and \(D = Nd\). In addition, since XL-MIMO has a very large number of antennas \(N\), the summation ([\[eq:Summation\]](#eq:Summation){reference-type="ref" reference="eq:Summation"}) can be rewriten as the following Riemann integral \[\begin{aligned}
\label{eq:Integral} \|\bar{\mathbf{a}}^H(\theta_1, \alpha_1, f_1)\bar{\mathbf{a}}(\theta_2, \alpha_2, f_2)\| \overset{N \rightarrow +\infty}{=} \left| \int_{-\frac{1}{2}}^{\frac{1}{2}} e^{j\frac{2\pi Dx}{\lambda_1}(\theta_1-\frac{f_2}{f_1}\theta_2)- j\frac{2\pi D^2x^2}{\lambda_1}(\alpha_1-\frac{f_2}{f_1}\alpha_2)} \text{d}x \right|. \end{aligned}\] Eventually, plug \(\gamma = \frac{D}{\lambda_1}(\theta_1-\frac{f_2}{f_1}\theta_2)\) and \(\zeta = \frac{D^2}{\lambda_c}(\alpha_1-\frac{f_2}{f_1}\alpha_2)\) into ([\[eq:Integral\]](#eq:Integral){reference-type="ref" reference="eq:Integral"}), the right-hand side in ([\[eq:Lemma1\]](#eq:Lemma1){reference-type="ref" reference="eq:Lemma1"}) can be acquired, and the proof is completed. \(\hfill\blacksquare\) **Lemma 1** indicates that for XL-MIMO communications with a very large number of antennas \(N\), finding the support index \((n_{a, l, m}^{\star}, n_{d, l, m}^{\star})\) is equivalent to solving the following problem: \[\begin{aligned}
\label{eq:Optimization1} (n_{a, l, m}^{\star}, n_{d, l, m}^{\star}) = \arg\max_{n_a, n_d} \Xi\left(\frac{D}{\lambda_c}(\bar{\theta}_{n_a}-\frac{f_m}{f_c}\theta_l), \frac{D^2}{\lambda_c}(\bar{\alpha}_{n_d}-\frac{f_m}{f_c}\alpha_l)\right). \end{aligned}\] To address problem ([\[eq:Optimization1\]](#eq:Optimization1){reference-type="ref" reference="eq:Optimization1"}), it is of great importance to grasp the property of function \(\Xi(\gamma, \zeta)\). First of all, it is easy to prove that \(\Xi(\gamma, \zeta)\) is an even function with respect to (w.r.t) \(\gamma\) and \(\zeta\), i.e., \[\begin{aligned}
\Xi(\gamma, \zeta) = \Xi(-\gamma, \zeta), \quad \Xi(\gamma, \zeta) = \Xi(\gamma,-\zeta). \end{aligned}\] Therefore, we can rewrite \(\Xi(\gamma, \zeta)\) as \(\Xi(\gamma, \zeta) = \Xi(|\gamma|, |\zeta|)\), or we only need to discuss the domain \(\gamma > 0\) and \(\zeta > 0\). In addition, \(\Xi(\gamma, \zeta)\) is a non-parameter function, we can use numerical integration to acquire its numerical result, which is plotted in Fig. [\[fig:Xi\]](#fig:Xi){reference-type="ref" reference="fig:Xi"}. This numerical result indicates that within the main lobe of \(\Xi(\gamma, \zeta)\) (area surrounded by the black line in Fig. [\[fig:Xi\]](#fig:Xi){reference-type="ref" reference="fig:Xi"} (b)), the value of \(\Xi(\gamma, \zeta)\) declines with the increase of \(|\gamma|\) and \(|\zeta|\). That is to say, if \(|\gamma_1| \le |\gamma_2|\) and \(|\zeta_1| \le |\zeta_2|\), then we have \(\Xi(\gamma_1, \zeta_1) \ge \Xi(\gamma_2, \zeta_2)\). Making use of these properties, **Lemma 2** gives out the close form solution of ([\[eq:Optimization1\]](#eq:Optimization1){reference-type="ref" reference="eq:Optimization1"}). *Proof*: In the main lobe, since \(\Xi(|\gamma|, |\zeta|)\) is a decreasing function w.r.t \(|\gamma|\) and \(|\zeta|\), and \(\Xi(\gamma, \zeta) = \Xi(|\gamma|, |\zeta|)\), hence maximizing \(\Xi\left(\frac{D}{\lambda_c}(\bar{\theta}_{n_a}-\frac{f_m}{f_c}\theta_l), \frac{D^2}{\lambda_c}(\bar{\alpha}_{n_d}-\frac{f_m}{f_c}\alpha_l)\right)\) is equivalent to minimizing \(|\frac{D}{\lambda_c}(\bar{\theta}_{n_a}-\frac{f_m}{f_c}\theta_l)|\) and \(|\frac{D^2}{\lambda_c}(\bar{\alpha}_{n_d}-\frac{f_m}{f_c}\alpha_l)|\) separately, which gives rise to the solutions ([\[eq:s1\]](#eq:s1){reference-type="ref" reference="eq:s1"}) and ([\[eq:s2\]](#eq:s2){reference-type="ref" reference="eq:s2"}). \(\hfill\blacksquare\) **Lemma 2** explicitly displays the bilinear pattern over frequencies of the near-field beam split effect. As illustrated by the dotted lines in Fig. [\[fig:bilinear\]](#fig:bilinear){reference-type="ref" reference="fig:bilinear"}, both \(\frac{f_m}{f_c}\alpha\) and \(\frac{f_m}{f_c}\theta\) grow linearly to frequency \(f_m\). The support indices \(n_{a, l, m}^{\star}\) and \(n_{d, l, m}^{\star}\) correspond to the sampled \(\theta_{n_{a}}\) and \(\alpha_{n_d}\) that are nearest to \(\frac{f_m}{f_c}\theta\) and \(\frac{f_m}{f_c}\alpha\). Thus, this double linear growth pattern is defined as the bilinear pattern of near-field beam split. More importantly, the bilinear pattern can be utilized to capture the support set of the \(l\)-th path for all frequencies, so as to accurately estimate the location \((\vartheta_l, r_l)\). Its superiority comes from two aspects. At first, **Lemma 2** implies that the indices \(n_{a, l, m}^{\star}\) and \(n_{d, l, m}^{\star}\) can gather the largest power of the \(l\)-th path from the entire bandwidth. Therefore, compared to conventional common support set assumption in MMV model, the proposed bilinear pattern enjoys higher recovery accuracy for wideband channels. In addition, different from existing works that only explain the beam split pattern ([\[eq:s1\]](#eq:s1){reference-type="ref" reference="eq:s1"}) in the angle domain , the bilinear pattern reveals the linear-changing structure in the joint angle-distance domain. Thus, the proposed pattern is capable of estimating both far-field and near-field channels.
## Bilinear Pattern Detection Based Channel Estimation {#bilinear-pattern-detection-based-channel-estimation}
In this subsection, based on the bilinear pattern discussed before, a bilinear pattern detection based channel estimation scheme is proposed to solve problem \((\ref{eq:MMV})\). This algorithm is inspired by the classical CS-based polar-domain SOMP method. As illustrated in, the polar-domain SOMP method assumes that the polar-domain channel \(\bar{\mb{h}}_m\) is sparse and the sparse support sets of \(\bar{\mb{h}}_m, \forall m\) are the same. Based on this assumption, the polar-domain SOMP method computes the total power of each row of \(\bar{\mb{H}}\) so as to accumulate the information from the entire bandwidth. In this way, it can recover the location \((\vartheta_l, r_l)\). However, we have proved previously that the actual sparse support sets for different frequencies are not the same in wideband conditions. Indeed, they should satisfy the bilinear pattern presented in ([\[eq:s1\]](#eq:s1){reference-type="ref" reference="eq:s1"})-([\[eq:s2\]](#eq:s2){reference-type="ref" reference="eq:s2"}). Therefore, instead of computing the total power of each row of \(\bar{\mb{H}}\), our BPD based algorithm takes advantage of the bilinear pattern to collect powers from all frequencies, so as to improve the accuracy of the estimation for \((\vartheta_l, r_l)\).
In this subsection, the computational complexity of the BPD based channel estimation algorithm is analyzed, where we mainly count the number of complex multiplications. As indicated in **Algorithm 1**, the complexity mainly comes from the iteration procedure of steps 8, 9, 13, and 15. In step 8, the product of matrices \(\bm{\Psi}^H \in \mathbb{C}^{N_aN_d \times P N_{\text{RF}}}\) and \(\mb{R} \in \mathbb{C}^{P N_{\text{RF}} \times M}\) has a complexity in the order of \(\mathcal{O}(N_aN_d P N_{\text{RF}}M)\). Notice that the number of sampled angles \(N_a\) and distances \(N_d\) is usually proportional to the number of antennas \(N\). In step 9, we need to calculate the power \(\sum_{m = 1}^M \|u_m\left( (\Lambda(n_d, m)-1) N_a + \Gamma({n_a}, m) \right) \|^2\) for \(N_a N_d\) times, whose complexity is in the order of \(\mathcal{O}(N_aN_dM)\). The complexity of step 13 is dominated by the calculation of matrix \({\bm{\Psi}}^{\dagger}(:, \Upsilon_m) \in \mathbb{C}^{l \times PN_{\text{RF}}}\), which can be regarded as \({\bm{\Psi}}^{\dagger}(:, \Upsilon_m) = ({\bm{\Psi}}^H(:, \Upsilon_m){\bm{\Psi}}(:, \Upsilon_m))^{-1}{\bm{\Psi}}^H(:, \Upsilon_m)\). Hence, the complexity of calculating \({\bm{\Psi}}^{\dagger}(:, \Upsilon_m)\) consists of three computation steps: the product of \({\bm{\Psi}}^H(:, \Upsilon_m)\) and \({\bm{\Psi}}(:, \Upsilon_m)\), the inverse of \({\bm{\Psi}}^H(:, \Upsilon_m){\bm{\Psi}}(:, \Upsilon_m)\), and the product of \(({\bm{\Psi}}^H(:, \Upsilon_m){\bm{\Psi}}(:, \Upsilon_m))^{-1}\) and \({\bm{\Psi}}^H(:, \Upsilon_m)\). Therefore, step 13 has a complexity in the order of \(\mathcal{O}(l^2P N_{RF} + l^3 + l^2PN_{RF}) = \mathcal{O}(l^2PN_{RF} + l^3)\). As step 13 is carried out \(M\) times, its overall complexity is in the order of \(\mathcal{O}(M(l^2PN_{RF} + l^3))\). Next, in step 15, each column vector \(\hat{\bar{\mb{h}}}_m\) only has \(l\) non-zero elements. As a result, the left multiplying \([\hat{\bar{\mb{h}}}_1, \hat{\bar{\mb{h}}}_2, \cdots, \hat{\bar{\mb{h}}}_M]\) by matrix \(\bm{\Psi}\) has a complexity of \(\mathcal{O}(PN_{\text{RF}}lM)\). Eventually, steps 8, 9, 13, 15 are executed \(\hat{L}\) times with \(l = 1,2,\cdots, \hat{L}\). Therefore, the total complexity can be summarized as \[\begin{aligned}
\label{eq:comp} &\:\mathcal{O}(\hat{L}N_aN_d P N_{\text{RF}}M) + \mathcal{O}(\hat{L}N_aN_dM) + \mathcal{O}(\hat{L}^3MPN_{RF} + \hat{L}^4) + \mathcal{O}(\hat{L}^2PN_{\text{RF}}M) \notag \\ = &\:\mathcal{O}(\hat{L}N_aN_d P N_{\text{RF}}M) + \mathcal{O}(\hat{L}^3MPN_{RF} + \hat{L}^4) + \mathcal{O}(\hat{L}^2PN_{\text{RF}}M). \end{aligned}\] In Table [1](#tab1){reference-type="ref" reference="tab1"}, we have provided the computational complexity comparison for different channel estimation algorithms, including the angle-domain SOMP algorithm for far-field narrowband channel estimation, the BSPD algorithm for far-field wideband channel estimation, and the polar-domain SOMP algorithm for near-field narrowband channel estimation. It is clear from Table [1](#tab1){reference-type="ref" reference="tab1"} that the proposed BPD algorithm has the same complexity as the polar-domain SOMP scheme in. This is because the main difference between these two algorithms lies in step 9. Specifically, the polar-domain SOMP technique uses the common sparse support set assumption to accumulate the power while the proposed algorithm exploits the bilinear pattern to accumulate the power, both of which have the same complexity \(\mathcal{O}(N_aN_dM)\). Moreover, the complexity of near-field algorithms is higher than that of far-field algorithms, since the far-field methods only need to recover the AoA parameters while the near-field schemes have to detect the AoA and distance parameters simultaneously. To illustrate, the complexity difference comes from the first item in ([\[eq:comp\]](#eq:comp){reference-type="ref" reference="eq:comp"}), i.e., \(\mathcal{O}(\hat{L}N_aN_d P N_{\text{RF}}M)\) for near-field methods and \(\mathcal{O}(\hat{L}N_d P N_{\text{RF}}M)\) for far-field methods. As the far-field channel estimation solely cares about the AoA information, the number of sampled distances \(N_d\) in can be regarded as 1, giving rise to the complexity of \(\mathcal{O}(\hat{L}N_d P N_{\text{RF}}M)\). Despite the lower computational complexity, the far-field algorithms neglect the essential distance information, so they can hardly achieve accurate XL-MIMO channel estimation, which is demonstrated in the next section.
# Simulation Results {#sec:4}
In this section, we present simulation results to demonstrate the performance of the proposed algorithm. A wideband XL-MIMO system is considered, and some of the simulation configurations are shown in Table [2](#tab2){reference-type="ref" reference="tab2"}. According to these configurations, the number of BS's antennas is \(N = 256\) and the carrier wavelength is \(\lambda_c = \frac{c}{f_c} = 3\) mm. Therefore, the antenna spacing is \(d = \frac{\lambda_c}{2} = 1.5\) mm, and the BS array aperture is \(D = (N-1)d \approx N d = \frac{N\lambda_c}{2} = 0.384\) m. Then, it can be derived that the Fraunhofer distance is \(\frac{2D^2}{\lambda_c} = 98.3\) m. In addition, the distance parameters \(r_l\) in ([\[eq:Channel\]](#eq:Channel){reference-type="ref" reference="eq:Channel"}) are randomly generated from the uniform distribution \(\mathcal{U}(R_{\min}, R_{\max})\), where the smallest distance \(R_{\min}\) and the largest distance \(R_{\min}\) have different values for different simulations. The signal-to-noise radio is defined as \(\text{SNR} = \mathbb{E}\left( \| \mb{H} \|_F^2 / \| \bar{\mb{N}} \|_F^2\right)\). The compared benchmark channel estimation algorithms are as follows: 1) the LS algorithm; 2) the far-field narrowband channel estimation schemes, including the angle-domain OMP and SOMP algorithms; 3) the BSPD based far-field wideband channel estimation ; 4) the near-field narrowband channel estimation schemes, including the polar-domain OMP and SOMP algorithms . We use the performance of normalized mean square error (NMSE) to evaluate different algorithms, which is defined as \(\text{NMSE} = \mathbb{E} \left( \frac{\| \mb{H}-\hat{\mb{H}}\|_F^2 } {\| \mb{H} \|_F^2} \right)\). Besides, 300 Monte Carlo experiments are carried out to plot each figure. First of all, the NMSE performance against distance is shown in Fig. [\[fig:NMSE_R\]](#fig:NMSE_R){reference-type="ref" reference="fig:NMSE_R"} to explain the influence of near-field effect. In this figure, we assume the distance \(r_l\) is growing from \(5\) meters to 100 meters, where we set \(R_{\min} = R_{\max} = r_l\). The other parameter settings are as follows: 1) the \(\text{SNR}\) is fixed as \(5\) dB, 2) the bandwidth is \(B = 10\) GHz, 3) and the pilot overhead is \(P = 32\), with an observation dimension of \(P N_{\text{RF}} = 128\). One can observe from Fig. [\[fig:NMSE_R\]](#fig:NMSE_R){reference-type="ref" reference="fig:NMSE_R"} that the achieved NMSE of all far-field channel estimation methods (angle-domain OMP, SOMP, and BSPD) degrades with the decrease in distance. This is because these algorithms tailored for far-field communications neglect the impact of spherical wavefront. Besides, as the LS algorithm is able to work on all kinds of channels, its performance is robust to different distances. Similarly, all near-field channel estimation algorithms also have a stable NMSE in both far-field and near-field environments by exploiting the polar-domain sparsity. Despite this distance-robust advantage, the achieved NMSE of the LS, polar-domain OMP, and SOMP methods are not satisfactory, because they fail to utilize the underlying beam split structure in wideband systems. Fortunately, the proposed BPD algorithm well captures the polar-domain frequency-dependent sparse support structure, and thus it outperforms all benchmark algorithms. Then, Fig. [\[fig:NMSE_B\]](#fig:NMSE_B){reference-type="ref" reference="fig:NMSE_B"} illustrates the NMSE performance against system bandwidth, which is growing from \(100\) MHz to 10 GHz. The other parameters settings are as follows: \(\text{SNR} = 5\) dB, the pilot overhead is \(P = 32\), \(R_{\min} = 10\) meters, and \(R_{\max} = 30\) meters. As the common sparse support set characteristic is not valid in wideband systems, the NMSE performance of angle-domain and polar-domain SOMP algorithms becomes worse and worse with the increase of bandwidth. In addition, the LS, angle-domain OMP, and polar-domain OMP schemes make no assumption on the wideband channel structure, where the channel of each subcarrier is independently estimated. As a consequence, these algorithms achieve a stable but non-satisfactory channel estimation accuracy for different bandwidth conditions. Moreover, the BSPD method accurately describes the angle-domain beam split pattern but ignores the distance-domain beam split pattern, so it is able to slightly improve the estimation accuracy for different bandwidth cases. On the other hand, we can observe from Fig. [\[fig:NMSE_B\]](#fig:NMSE_B){reference-type="ref" reference="fig:NMSE_B"} that the proposed BPD algorithm realizes both stable and the most accurate estimation performance in all bandwidth cases. This is because the BPD scheme fully exploits the frequency-dependent sparse structure resulting from the near-field beam split effect. In Fig. [\[fig:NMSE_SNR\]](#fig:NMSE_SNR){reference-type="ref" reference="fig:NMSE_SNR"}, we evaluate the achieved NMSE performance in different SNR conditions. Here, the SNR grows from-5 dB to 15 dB. The other parameter settings are set as below: \(\text{B} = 10\) GHz, \(R_{\min} = 10\) meters, \(R_{\max} = 30\) meters, and \(P = 32\). The accuracy achieved by all considered algorithms improves with the increase of SNR. It is clear that the proposed BPD method significantly outperforms all compared benchmarks, especially in high SNR cases. For example, when the SNR is 7 dB, around 5 dB improvement of NMSE is realized compared to the BSPD method and the polar-domain OMP method. Eventually, in Fig. [\[fig:NMSE_P\]](#fig:NMSE_P){reference-type="ref" reference="fig:NMSE_P"}, the NMSE performance with respect to the pilot overhead \(P\) is plotted. As illustrated in Fig. [\[fig:NMSE_P\]](#fig:NMSE_P){reference-type="ref" reference="fig:NMSE_P"}, the pilot overhead \(P\) is increasing from 4 to 32, which corresponds to a compression ratio \(\frac{N}{PN_{\text{RF}}}\) declining from \(16\) to \(2\). Then, the other parameters are set as: \(\text{SNR} = 5\) dB, \(R_{\min} = 10\) meters, \(R_{\max} = 30\) meters, and \(B = 10\) GHz. It is clear that the proposed BPD based method outperforms other existing channel estimation algorithms for all considered pilot lengths. This fact implies that the proposed method can be used to reduce the pilot overhead. Specifically, take the NMSE of-9 dB as a baseline, around 36 pilot overhead is required for the BSPD to achieve this NMSE baseline. On the contrary, only 12 pilot length is enough for the proposed BPD method to reach this baseline. In this case, our scheme has the capability of reducing the pilot overhead by \(66\%\).
# Conclusions {#sec:5}
In this paper, we have investigated the channel estimation for wideband XL-MIMO communications in the presence of near-field beam split effect. Specifically, we first revealed the bilinear pattern of the near-field beam split effect, which indicated that the polar-domain sparse support set for each near-field channel path shows a linear structure over frequencies. Then, we have proposed a BPD based channel estimation algorithm to recover each near-field channel path, by using the bilinear pattern to accumulate the largest polar-domain power from the entire bandwidth. Simulation results demonstrated our scheme is capable of achieving high channel estimation accuracy in all far-field/near-field/narrowband/wideband conditions. For future works, the bilinear pattern discovered in this paper can potentially be extended to tackle the relevant channel estimation issues in various near-field wideband communication scenarios, such as reconfigurable intelligent surface (RIS) communications and cell-free massive MIMO communications .
[^1]: Simulation codes of this paper are available at http://oa.ee.tsinghua.edu.cn/dailinglong/publications/publications.html.
[^2]: Notice that the proposed scheme can achieve higher accuracy on the estimation of the location parameter \((\vartheta_l, r_l)\) at all frequencies than existing channel estimation schemes. Thus, our scheme can also be used to improve the time of arrival (ToA) estimation accuracy. | {'timestamp': '2022-12-19T02:10:03', 'yymm': '2212', 'arxiv_id': '2212.08401', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08401'} |
# Introduction {#sec:intro}
An efficient and accurate video anomaly detection algorithm can help maintain social security and stability. Therefore, video anomaly detection has high practical value and broad application prospects. With the development of weakly supervised learning, the weakly supervised video anomaly detection (WSVAD) method is an effective method for detecting anomalies, which uses weakly labeled training data containing both normal and anomalous videos to train the model. Recently, WSVAD has been formulated as a multiple instance learning (MIL) task. Sultani et al. constructed a large-scale anomaly dataset and proposed a deep MIL ranking based method for WSVAD. Wan et al. replaced the max anomaly score selection policy with a k-max value selection policy. Li et al. selected the sequence with the highest sum of anomaly scores instead of selecting the instance with the highest anomaly score. Gong et al. introduced temporal continuity of multiple neighboring instances at different time scales. However, most of the existing work have only used MIL-based classification loss. Although MIL-based classification loss ensures the inter-class separability of the learned features to some extent, it is not sufficient for accurate anomaly detection. Therefore, we propose a batch clustering based loss to further increase the discriminative power of the features, and our framework is shown in Figure [\[fig:model\]](#fig:model){reference-type="ref" reference="fig:model"}. The abnormal videos in a batch are clustered into two clusters, and then the loss encourages the network to maximize the distance between these two clusters. Meanwhile, for normal videos in a batch, the loss encourages the network to minimize the distance between these two clusters. Compared to clustering each video individually, batch clustering improves the robustness of the model and reduces the influence of the model affected by the noise. Considering the data in WSVAD task is highly unbalanced, it may negatively affect model training when only using a small portion of data. Previous mini-batches can provide valuable knowledge enabling the model to better understand the underlying distribution of the data. Therefore, we use a cross-batch learning strategy to provide guidance for the current batch clustering by introducing clustering results from previous batches. The introduced cross-batch learning strategy can make the clustering results more accurate, model the temporal-spatial distribution of the data better, and improve the adaptability of the model to unbalanced samples. In addition, the knowledge that the model can learn in WSVAD directly from supervised learning is limited, since only video-level labels are available. The clustering branch can obtain potential information reflecting the similarity of video segments. Therefore, we propose anomaly score generation based on batch clustering guidance to generate more discriminative anomaly scores and further improve the model performance. We first generate pseudo labels for video segments based on the batch clustering results, and then uses the pseudo labels to guide the backbone network to rectify the estimated anomaly scores for video segments. In summary, our main contributions are as follows: 1) We propose a loss based on batch clustering to complement and enhance the separability of the features guided by MIL-based classification loss. 2) We propose a cross-batch learning strategy to generate more accurate clustering results. 3) We propose to use the knowledge learned from batch clustering to guide the predition of more discriminative anomaly scores.
# Proposed Method {#method}
## Backbone network
Our approach employs a backbone netwwork based on graph convolutional neural network (GCN) to model video sequences. The backbone network consists of a feature extraction module and a graph convolution module. The Inflated 3D (I3D) pretrained on the Kinetics dataset is used as the feature extraction network to extract the appearance and motion information of the video segments. Before each video \(V_i\) is fed into the feature extraction module, the video is divided into non-overlapping segments containing 16 consecutive frames, and we denote the number of segments by \(T_i\). The graph convolution module consists three graph convolution layers, where the first two layers are followed by a ReLU activation function and a dropout layer, and the last layer is followed by a Sigmoid activation function. For each video \(V_i\), the input layer receives the temporal-spatial features extracted from the feature extraction module and the adjacency matrix of a global graph constructed based on feature similarity and temporal proximity of the video segments. The output layer produces the anomaly score vector of the video \(\tilde{S}_{i}=\left\{\tilde{s}_{i, j}\right\}_{j=1}^{T_{i}}\). The network is trained using video-level labels. \(y_{i} \in\{0,1\}\) is the video-level label of video \(V_i\), where \(y_i=0\) indicates that video \(V_i\) is a normal video and \(y_i=1\) indicates that \(V_i\) is an abnormal video.
## Batch clustering based on K-means
Although MIL-based classification loss ensures the inter-class separability of learned features to a certain extent, it cannot ensure a more discriminative power of features since there is no explicit supervision. Several studies on unsupervised anomaly detection have enlightened us, in which normal samples are compulsorily clustered in a compact space such that they can be kept away from the anomaly space. Therefore, we have reasonable grounds to believe that the normal activities should be compact in the feature space. Consequently, we use batch clustering to cluster normal video segments to enhance the intra-class compactness of normal features. A larger inter-class distance in abnormal video indicates that normal and abnormal are separated by a higher probability. Therefore, we perform batch clustering on abnormal video segments to enhance the inter-class dispersion of normal and abnormal features. Here, we propose batch clustering to provide supervision to enhance the discriminative power of features as shown in Figure [\[fig:model\]](#fig:model){reference-type="ref" reference="fig:model"}. For all normal videos in a batch, the feature representations of each video are clustered into two clusters. Since all segments in normal videos are normal, we try to close the distance of the two clusters to ensure the intra-class compactness of the normal features. All the abnormal videos in a batch are also clustered into two clusters. Since there are both abnormal and normal video segments in the abnormal videos, we try to push the centers of the two clusters away from each other to achieve the inter-class dispersion of normal and abnormal features. Specifically, for the abnormal or normal videos in a batch, we use the K-means algorithm to cluster the normalized intermediate feature representations, which is the output of the first layer of GCN. The loss based on batch clustering is shown below: \[L_{b c}=\left\{\begin{array}{c} \min (d, \mu), \text { if }\left\{V_{i}\right\}_{i=1}^{b} \text { are normal videos } \\ \frac{1}{d}, \text { if }\left\{V_{i}\right\}_{i=1}^{b} \text { are abnormal videos } \end{array}\right.\] where \(d=c_{1}-c_{2}\) is the distance between two cluster centers, and \(c_{1}\), \(c_{2}\) are the two cluster centers. \(\mu\) is an upper bound that helps the model to be robust to different videos, and \(b\) is the batch size. We train the model using the \(k\)-max loss function to expand the inter-class distance between abnormal and normal segments, denoted as follows: \[L_{k-\max }=-\frac{1}{k_{i}} \sum_{s_{i, j} \in p_{i}}\left[y_{i} \log \left(s_{i, j}\right)+(1-y_{i}) \log \left(1-s_{i, j}\right)\right]\] where \(p_i\) is the first \(k_i\) large elements in \(S_i\) of video \(V_i\), \(k_{i}=\left\lfloor\frac{T_{i}}{8}+1\right\rfloor\), \(y_i\) is the video-level label of video \(V_i\). So, the total loss function is expressed as: \[\label{loss2} L=L_{k-\max }+\lambda_{1} L_{b c}\] where \(\lambda_{1}\) is a trade-off hyperparameter to keep the balance between two losses.
## Cross-batch learning strategy {#batch}
The data in WSVAD task is highly imbalanced. However, only a mini-batch of samples can be accessed in each iteration. The performance of anomaly detection models can be improved when the batch size becomes larger on large-scale datasets. However, simply scaling up the batch is not an ideal solution because batch size is limited by GPU memory and computational cost. A simple solution to collect rich information is to introduce information from previous batches at each training iteration to enable the model to better understand the underlying distribution of the data. Thus, video segments from past batches can also serve as an important reference when performing \(K\)-means clustering on video segments of the current batch. Previously, we perform batch clustering based on \(K\)-means by directly selecting any two video segments from all video segments as the initial clustering centers, but the final clustering results of the \(K\)-means depend on the selection of the initial clustering centers heavily. In order to select the most appropriate initial clustering centers, we introduce the clustering results of previous batches to provide guidance for the current batch clustering, which can help model training. Specifically, at the \(t\)-th epoch, \(C^a\) and \(C^n\) are constructed in the iteration process to store the learned knowledge. That is, during the \(i\)-th iteration, we cluster all abnormal video segments in a batch, and add the two clustering centers into \(C^a\). We will obtain \(C^{a}=\left\{c_{i, 1}^{a}, c_{i, 2}^{a}\right\}_{i=1}^{m}\) at the end of the \(t\)-th epoch, where \(m\) is the number of iterations in an epoch. We also do the above operation for all normal video segments in a batch to get \(C^{n}=\left\{c_{i, 1}^{n}, c_{i, 2}^{n}\right\}_{i=1}^{m}\). At the \((t+1)\)-th epoch, we use the stored information to guide the batch clustering. We cluster the data in \(C^a\) to obtain the clustering centers \(c_1^a\), \(c_2^a\), which are used as the initial clustering centers for the batch clustering of all abnormal video segments in each batch of the current epoch. Meanwhile, the data in \(C^n\) are clustered to obtain the clustering centers \(c_1^n\), \(c_2^n\), which are used as the initial clustering centers for all normal video segments in each batch of the current epoch. Dynamically updating the data in \(C^a\) and \(C^n\) among different epochs and introducing them into the batch clustering as a priori information can improve the effect of batch clustering and further improve the performance of the anomaly detection task.
## Anomaly score generation based on batch clustering guidance
When the available labels are limited, the labeled samples often do not provide sufficient supervised information for the model, so the deep model is prone to overfitting. Since WSVAD only considers video-level labels, the knowledge that the model can learn will be limited. To address this problem, we use cluster labels obtained from batch clustering to guide backbone network to predict the anomaly scores of segments. The unsupervised information is effectively transferred to the weakly supervised learning process to improve the performance of the anomaly detection task. For the labeled normal videos, each segment of these videos can simply be annotated as normal because there are no abnormal events in it. However, in the case of abnormal videos, there are also some normal events, so we use batch clustering results to generate pseudo label for each segment of the abnormal video. All segments of abnormal videos are clustered into two clusters assuming that one cluster would contain normal segments, while the other would contain abnormal. Therefore, we need to analyze which of the two clusters contains mostly normal segments and which contains mostly abnormal segments so that we can assign the appropriate pseudo labels for video segments. We obtain the pseudo labels of the segments by the similarity score between the anomaly scores and the clustering labels. Specifically, we compute the cosine similarity score \(S_1\) between the anomaly score \(s_i\in[0,1]\) of video \(V_i\) predicted by the backbone network and the label \(y_{i}^{c} \in\{0,1\}\) generated by clustering, and the cosine similarity score \(S_2\) between \(s_i\) and the inverted clustering label \(\neg y_{i}^{c}\). Finally, the pseudo label \(y_{i, j}^{p}\) of the \(j\)-th segment of video \(V_i\) is given by the following equation: \[y_{i, j}^{p}=\left\{\begin{array}{cc} y_{i, j}^{c}, & \text { if } s_{1}>s_{2} \\ \neg y_{i, j}^{c}, & \text { otherwise } \end{array}\right.\] The pseudo labels generated by batch clustering for video segments and the anomaly scores generated by adaptive graph convolutional networks for video segments can complement each other in the anomaly detection task. In an anomalous video, if the pseudo label of a video segment is 1, the segment has a high probability of being abnormal. Therefore, we expand the anomaly score of the segments with a pseudo label 1 in the abnormal video. In particular, in an abnormal video, if the pseudo label of the video segment is 1, the abnormal score of the video segment becomes \(\min \left(\alpha \times s_{i, j}, 1\right)\); if the pseudo label of the video segment is 0, the abnormal score of the video segment remains unchanged. The anomaly score of each video segment is given by the following equation: \[s_{i, j}=\left\{\begin{array}{c} \min \left(\alpha \times \tilde{s}_{i, j}, 1\right), \text { if } y_{i}=1 \text { and } y_{i, j}^{p}=1 \\ \tilde{s}_{i, j}, \text { otherwise } \end{array}\right.\] where \(\alpha\) is the expansion factor.
# Experimental results
## Datasets and evaluation metrics
UCF-Crime is a large-scale dataset which spans over 128 hours of surveillance videos. It covers 13 realistic anomalies. The entire dataset contains 1,900 long untrimmed videos, of which 1610 videos with video-level label are used for training and the rest for testing. ShanghaiTech is a medium-scale campus surveillance dataset containing 437 videos. Following Zhong et al., we split the data into two subsets: the training set consisting of 175 normal videos and 63 anomalous videos, and the test set containing 155 normal videos and 44 anomalous videos. Following previous work, we use the area under curve (AUC) of the receiver operating characteristic curve (ROC) at the frame level as the criterion for the model, and a higher AUC value indicates a better detection of the model.
:::
## Experimental details
We extract the 2048-dimensional features from the "mix 5c" layer of I3D. Following previous work, we extract \(T\) segments from the video uniformly to represent the whole video. By default, we set \(T\) to 150 for UCF-Crime and 100 for ShanghaiTech. The fully connected layer in the model has 512 nodes, and the graph convolutional network layer has 128, 32 and 1 nodes, respectively. Our model is trained with a mini-batch size of 64 using the Adam optimizer. For hyper-parameters, \(\alpha\) is set to 1.3.
## Experimental results
We compare our method with the current methods on two datasets, and show results in Table [1](#table_1){reference-type="ref" reference="table_1"}. For the UCF-Crime dataset, comparison results show that our method outperforms all comparison methods. Using the same I3D-RGB features, our method outperforms the previous GCN-based method by 4.79% over Zhong et al. and 3.43% over Wu et al.. Our method also surpasses previous methods using clustering by 6.33% over SRF and 2.84% over CLAWS. It also achieves better performance compared to previous weakly supervised methods on the ShanghaiTech dataset. It is 20.01% higher than the GCN-based method, 12.29% and 6.78% higher than the clustering-based method and, respectively.
## Ablation experiments
We perform ablation study to investigate the contribution of each component of our proposed method. We start by evaluating only the backbone network and observe the performance gain while adding different modules. The result on UCF-Crime is shown in Table [2](#bottomup1){reference-type="ref" reference="bottomup1"}. The backbone network achieves 84.67% AUC while the addition of batch clustering based loss improves the performance to 85.21% and the addition of cross-batch learning strategy further improves the performance to 85.56% which validates their effectiveness. Addition of the anomaly score generation based on batch clustering guidance improves the performance to 85.87%, demonstrating that the pseudo labels generated by batch clustering can guide the backbone network. We select different network layer outputs for batch clustering to investigate their effects on model performance. We choose the outputs of FC layer (FC), the first GCN layer (GCN-1), and the second GCN layer (GCN-2), respectively. As shown in Table [3](#layer){reference-type="ref" reference="layer"}, using GCN-1 achieves the best result with 0.25% higher than that using FC, probably because GCN-1 exploits the temporal relationship between video segments. The performance using GCN-2 is 0.94% lower than that using GCN-1 because GCN-1 contains more information due to its higher dimensionality compared to GCN-2. We conduct experiments and show comparison results to evaluate the effects of different cross-batch learning strategies. The first method is the one mentioned in Section [2.3](#batch){reference-type="ref" reference="batch"}. The second method is to save the clustering centers obtained from the previous batch clustering as the initial clustering centers when clustering the current batch. The third method is to save the clustering centers obtained from all previous batch clustering and use the clustering centers obtained by clustering them again as the initial clustering centers for the current batch clustering. The fourth method is to save the clustering centers of all the batches from the previous epoch as clustering samples to participate in the clustering of each batch in the current epoch. The Table [4](#cross_batch){reference-type="ref" reference="cross_batch"} shows that all our proposed cross-batch learning strategies improve the model performance. The first method has the best performance, and the fourth method performing less well, with only a small improvement. We also conduct experiment to further investigate the expansion factor \(\alpha\), and show the change of performance with different expansion factors \(\alpha\) on UCF-Crime in Figure [\[fig:expand\]](#fig:expand){reference-type="ref" reference="fig:expand"}. We can observe that the AUC first increases and then decreases as the value of \(\alpha\) increases. The model achieves the best performance when \(\alpha\) is 1.3.
## Qualitative Result and Analysis
To further evaluate the performance of our method, we visualize the anomaly score curves, as shown in Figure [\[fig:res\]](#fig:res){reference-type="ref" reference="fig:res"}. The figure shows the ground truth and the predicted anomaly scores of three abnormal videos and one normal video from UCF-Crime. It is obviously that our model exactly localizes the anomalous events, showing the effectiveness and robustness of our model. Our method successfully predicts both short-term anomalous events and long-term anomalous events. In addition, our method can also detect multiple anomalous events in a video.
# Conclusion
In this work, we propose a novel WSVAD model based on cross-batch clustering guidance. The method enhances feature discrimination by binary batch clustering for normal and abnormal videos within a batch separately. In addition, a cross-batch learning strategy is incorporated to solve the data imbalance problem caused by the mini-batch training strategy, allowing the model to better capture the potential distribution of the data. Finally, the pseudo labels generated by batch clustering guide the backbone network to generate the anomaly scores, which further enhances the separability of normal and anomalous. The experimental results show that the proposed method achieves significant improvements on commonly-used WSVAD datasets. | {'timestamp': '2022-12-19T02:12:55', 'yymm': '2212', 'arxiv_id': '2212.08506', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08506'} |
null | null |
# Introduction
Many large eddy simulation (LES) studies of turbomachinery flows focus on the low-pressure turbine (LPT) due to its low Reynolds number regime of \(10^5\). To further reduce the simulation costs, the assumption of a statistically 2D flow at the midspan of a turbine blade is often made by applying periodic boundary conditions in the spanwise direction. However, a significant amount of the aerodynamic losses is generated in the secondary flow regions influenced by the interaction of end wall and blade boundary layers . Hence, the next logical step to evaluate the performance of the cascade is to conduct 3D simulations, including the endwall boundary layers, e.g. . The MTU T161, considered in this work, is representative of high lift low-pressure turbine airfoils used in modern jet engines . The blades with a chord length of \(C = 0.069935\,\mathrm{m}\) and an average aspect ratio of 2.65 are staggered at an angle of \(61.72^\circ\). The cascade is arranged with pitch to chord ratio of \(t/C = 0.956\). It features diverging end walls at an angle of \(12^\circ\), such that the flow cannot be studied using a simple spanwise periodic setup. Its geometry and boundary conditions are in the public domain and it has been the subject of both experimental and numerical investigations. The numerical investigations have focused on operating points with a Mach number of 0.6 and Reynolds numbers of 90,000 and 200,000 based on isentropic exit conditions. Müller-Schindewolffs et al. performed a direct numerical simulation of a section of the profile in which the effect of the diverging end walls was modelled using inviscid walls (termed quasi 3D, Q3D). Recently, results obtained with a second order FV code were presented, which included the end wall boundary layers but the analysis was focussed on the flow physics in the mid-section . Various computations of the full 3D configuration were conducted using high-order codes during the EU project TILDA . However, due to the specification of laminar end wall boundary layers and no freestream turbulence at the inflow, no satisfactory results could be obtained. With a full 3D LES including appropriate turbulent end wall boundary layers and freestream turbulence obtained with a high order Discontinuous Galerkin (DG) method, we aim to provide a high-quality reference dataset for this configuration.
# Numerical method
We use DLR's flow solver for turbomachinery applications, TRACE, to perform an implicit LES with a kinetic-energy-preserving DG scheme for the spatial discretisation of the implicitly filtered Navier-Stokes equations . The scheme is based on the collocated nodal Discontinuous Galerkin Spectral Element Method (DGSEM) on Legendre-Gauss-Lobatto nodes with a polynomial order of 5. The anti-aliasing is performed by the split-formulation of Kennedy and Gruber , cf. . Due to the non-uniqueness of the solution at the element interfaces, Roe's approximate Riemann solver is applied for the advective part and the viscous terms are discretised by the Bassi-Rebay 1 scheme . To advance in time, a third-order explicit Runge-Kutta scheme of has been used. Resolved turbulent scales are injected at the inflow boundary using a synthetic turbulence generation (STG) method based on randomized Fourier modes . The fluctuating velocity is introduced using a Riemann boundary condition and, in contrast to the inner faces, no numerical flux function is used at the inflow faces.
# Inflow boundary conditions
The operating conditions are specified by centerline inlet flow angle of \(\alpha_1 = 41^\circ\), total pressure of \(p_{t, 1} = 11636 \, \mathrm{Pa}\) and total temperature of \(T_{t, 1} = 303.25 \, \mathrm{K}\) as well as outlet Mach number \(Ma_{2,\mathrm{th}} = 0.6\) and Reynolds number \(Re_{2,\mathrm{th}} = 90,000\) based on isentropic relations computed with the centerline outflow pressure. Furthermore it is required to match the momentum thickness of the incoming end wall boundary layers as well as the decay of freestream turbulence. The boundary conditions are then set by a spanwise varying profiles of \(p_{t, 1}(z)\) and \(T_{t, 1}(z)\) with a constant \(\alpha_1\) as well as spanwise varying Reynolds stress tensor \(\overline{u_i' u_j'}(z)\) and turbulent length scale \(L_T(z)\). All these profiles will be obtained by scaling a boundary layer profile in wall units at \(Re_\theta = 670\) (<https://www.mech.kth.se/~pschlatt/DATA/>, ) and combining it with freestream turbulence values. We perform a precursor channel flow simulation to determine the distance from the inlet required for the boundary layer to develop and meet the specifications. Starting from the above centerline inflow conditions, we use the isentropic relations \[\frac{T_t(z)}{T_{1}} = \left( \frac{p_t(z)}{p_{1}} \right)^{\frac{\gamma-1}{\gamma}} = \left( 1 + \frac{\gamma-1}{2} Ma(z)^2\right) \label{eqn:isentropic}\] to obtain a spanwise variation. Here, we assume a constant static temperature \(T_{1}\) and pressure \(p_{1}\) throughout the boundary layer. These two quantities can be computed from ([\[eqn:isentropic\]](#eqn:isentropic){reference-type="ref" reference="eqn:isentropic"}) by introducing a centerline Mach number \(Ma_{1,\mathrm{center}}\), which essentially controls the value of the \(T_t\) and \(p_t\) at the end wall. It has to be chosen such that the total pressure is always greater than the static pressure resulting from the simulation. Hence, an iteration might be necessary. It has to be noted, though, that small adaptations of \(Ma_{1,\mathrm{center}}\) mainly influence the total pressure profile close to the wall and do not have a significant impact on the quantities of interest in the simulation. For this concrete case, we chose \(Ma_{1,\mathrm{center}} = 0.362\). The discrete DNS velocity profile \(u_{i,\mathrm{DNS}}^+\) and the corresponding distances to the wall \(z^+_{i,\mathrm{DNS}}\) are now transformed to obtain a Mach number profile \(Ma(z)\) simply using the definitions of \(u^+\) and \(z^+\) via \[\label{eqn:wallUnitScaling} Ma(z_i) = u^+_{i, \mathrm{DNS}} \cdot \frac{u_\tau}{a_1}, \quad z_i = z^+_{i,\mathrm{DNS}} \cdot \frac{\nu_1}{u_\tau} \quad \mathrm{with} \quad u_\tau = a_1 \cdot Ma_{1,\mathrm{center}} / u_{\infty, \mathrm{DNS}}^+\] and the DNS freestream velocity \(u^+_{\infty, \mathrm{DNS}}\). The speed of sound \(a_1\) and viscosity \(\nu_1\) can be computed from \(T_1\) and \(p_1\) using ideal gas and Sutherland laws, respectively. This allows to describe the spanwise variation of \(T_t\) and \(p_t\) using ([\[eqn:isentropic\]](#eqn:isentropic){reference-type="ref" reference="eqn:isentropic"}). The non-dimensional turbulent stress tensor is scaled analogously via \(\overline{u_j' u_k'}(z_i) = \overline{u_j' u_k'}^+|_{i,\mathrm{DNS}} \cdot u_\tau^2\). Finally, the integral turbulent length scale in the boundary layer is estimated from the turbulent kinetic energy and dissipation rate as \(L_T^+ = (k^+)^{3/2} / \epsilon^+\) and dimensionalised as the \(z\)-coordinate in ([\[eqn:wallUnitScaling\]](#eqn:wallUnitScaling){reference-type="ref" reference="eqn:wallUnitScaling"}). Optimally, the freestream Reynolds stresses and turbulent length scale should be chosen such that the measured turbulent decay is matched. However, in this case, this would result in a length scale in the order of cascade pitch. This could, in principle, be accommodated for by simulating more than one blade at the respective expense of computational effort. Instead, we chose to decrease the turbulent length scale and scale the non-zero components of the Reynolds stress tensor at the inflow of the domain, to reproduce the turbulence intensity in the blade leading edge plane according to the experiment. This leads to a stronger decay of turbulent structures and has to be considered when assessing the quality of the results. The boundary layer profile and freestream values of the turbulence quantities are combined where they intersect at the edge of the boundary layer. Because of the large freestream turbulent length scale compared to the smaller length scale in the boundary layer, we use \(L_T = \max (L_{T, \mathrm{BL}}, L_{T, \mathrm{freestream}})\) for distances to the wall greater than \(\delta_{99}\). Finally, the Reynolds stress tensor is rotated from the streamline-aligned to a Cartesian coordinate system. Fig. [\[fig:inflow_and_mesh\]](#fig:inflow_and_mesh){reference-type="ref" reference="fig:inflow_and_mesh"} (*left*) shows both the development of momentum thickness Reynolds number \(Re_\theta\) and freestream turbulence intensity \(Tu\) in comparison to fits to the measured data. As discussed, the turbulent decay is steeper in our setup but the turbulence intensity at the blade leading edge (dashed vertical line) is well captured.
# Validation and analysis of flow structures
The mesh consisting of 876,960 hexahedral elements (geometry polynomial order \(q=4\), solution polynomial order \(p=5\)) with 189.4M degrees of freedom (DoF) for the final computation was designed to meet widely accepted criteria for wall resolution required for LES. This is demonstrated in Fig. [\[fig:inflow_and_mesh\]](#fig:inflow_and_mesh){reference-type="ref" reference="fig:inflow_and_mesh"} (*right*) which shows the streamwise, wall normal and spanwise solution point distances to be below 30, 1 and 30, respectively. The resolution in the free stream was ensured by the ratio of solution point distance and estimated Kolmogorov scale below 6 along a mid passage streamline. To be able to assess mesh dependence, results from a preliminary study with a significantly coarser mesh in spanwise direction with 312,424 elements (\(q=4\), \(p=5\)) and 67.9M DoF are also included. After clearing the initial transient, the statistics on the fine mesh were sampled for 100 convective time units based on chord length and outflow velocity. Fig. [\[fig:midspan_results\]](#fig:midspan_results){reference-type="ref" reference="fig:midspan_results"} shows a comparison of the midspan blade surface pressure distribution and wake total pressure losses with the experiment. Both simulations agree very well within the 99% confidence intervals for these first order statistics. While LES and experiment agree very well on the pressure side, a difference in surface pressure can be found on the suction side between \(x/C_\mathrm{ax} = 0.1\) and 0.6. Similar offsets have been found in previous studies of this configuration, e.g. . The laminar separation bubble and subsequent turbulent reattachment indicated by the pressure plateau between 0.7 and 0.9 and recovery of the base pressure, on the other hand, is captured very well. The total pressure loss coefficient \(\zeta\) is computed from time-averaged primitive variables in the wake plane at \(x/C_\mathrm{ax} = 1.4\) and the upstream reference plane at \(x/C_\mathrm{ax} =-0.099\). The pitch coordinate \(y\) is given with respect to the point \(y_\mathrm{LE}\) on the blade at its minimum axial coordinate \(x/C_\mathrm{ax} = 0\). The simulation on the coarse mesh shows larger confidence intervals in the highly turbulent region of the wake due to its shorter averaging time of only 31 convective time units. Both simulations and the experimental data agree very well within the statistical confidence intervals. With the simulation setup validated against the experiment, we can now discuss the system of secondary flows developing at the intersection of blade and end wall. It is shown in Fig. [\[fig:secondary_flows\]](#fig:secondary_flows){reference-type="ref" reference="fig:secondary_flows"} (*left* and *middle*) in terms of time-averaged surface streaklines visualised with line integral convolution and vortices visualised using an isosurface of \(Q = 10^7 \mathrm{s}^{-2}\) coloured with streamwise vorticity \(\omega_\mathrm{sw}\) to indicate the sense of rotation. Limiting surface lines have been added by hand to ease the interpretation. The horseshoe vortex develops due to a roll-up of the incoming boundary layer. While its suction side leg (blue) dissipates well before the suction peak, its pressure side leg follows the passage cross flow towards the suction side of the next blade lifting off the end wall and becoming the passage vortex (PV). Behind the blade, a second large structure can be identified as the trailing shed vortex (TSV), which rotates in the opposite direction. Both vortices persist well downstream of the blade and become visible as the two regions of strong loss in Fig. [\[fig:secondary_flows\]](#fig:secondary_flows){reference-type="ref" reference="fig:secondary_flows"} (*right*). The loss distribution in the wake plane also agrees very well with the measured data. Another small but very distinct vortex developing along the pressure side of the blade is the leading edge corner vortex. The pressure side features a short separation bubble close to the leading edge. Due to the diverging end walls, the backflow within the bubble is driven towards the endwalls where it rolls up, lifts off and mixes with the newly developing boundary layer in the passage between the passage vortex and the pressure side of the blade.
# Conclusions
We have presented a new well-resolved dataset of the MTU T161 at \(Ma_{2,\mathrm{th}} = 0.6\), \(Re_{2,\mathrm{th}} = 90,000\) and \(\alpha_1 = 41^\circ\) with focus on appropriate reproduction of inflow turbulent boundary layers and freestream turbulence. Average blade loading and losses distribution due to secondary flow features agree well with the experiment underlining the validity of the presented approach. After the brief discussion of the secondary flow structures in this paper, more in depth analysis of the unsteady data set is ongoing. Future analyses will introduce a numerical experiment featuring unsteady wakes at engine-relevant Strouhal numbers. | {'timestamp': '2022-12-19T02:10:51', 'yymm': '2212', 'arxiv_id': '2212.08432', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08432'} |
# Introduction
In this paper, motivated by considerations from Littlewood--Offord theory, we study the intersection of two classical combinatorial problems in the hypercube, namely that of finding large antichains and that of finding large distance-\(r\) codes. A family of sets \(A\subset 2^{[n]}\) is an *antichain* if \(x\not\subset y\) for any distinct \(x,y\in A\). For example, the layer \[\binom{[n]}{k}=\{x\subset [n]:|x|=k\}\] is an antichain for all \(0 \le k \le n\), and it is a classical result of Sperner that every antichain in the hypercube \(2^{[n]}\) has size at most \(\binom{n}{\lfloor n/2\rfloor}\). There are a huge number of strengthenings and variants of Sperner's theorem; we refer the reader to for more background. A family of vectors \(B \subset \{0,1\}^n\) is called a *distance-\(r\) code* if the Hamming distance between any pair of vectors in \(B\) is at least \(r\); identifying \(\{0,1\}^n\) and \(2^{[n]}\) in the natural way, we call a family \(A\subset 2^{[n]}\) a *distance-\(r\) code* if the symmetric difference \(x\,\triangle\, y\) of any two distinct \(x,y\in A\) always has size at least \(r\). One of the central problems of coding theory is to find large distance-\(r\) codes with various desirable properties, and the existence of such codes has many applications in both pure and applied problems. We refer the reader to for more on coding theory, and mention only the basic fact (as evidenced by BCH codes) that the largest possible distance-\(r\) codes in \(2^{[n]}\) have size \(\Theta(2^n n^{-\lfloor(r-1)/2\rfloor})\). Here, we aim to answer the following natural question: how large can the cardinality of an *antichain code* in \(2^{[n]}\) be? After some thought, one finds that it is difficult to do much better than taking the intersection of a large code and a large antichain; our main result shows that such constructions are indeed optimal.
This result is best-possible up to multiplicative constants. Indeed, as mentioned above, we know that there exists a distance-\((2r+1)\) code \(A\subset 2^{[n]}\) with \(|A|=\Theta(2^nn^{-r})\). By simple averaging, it is easy to show that there exists some \(x\subset[n]\) for which \(A \,\triangle\, x=\{a\,\triangle\, x:a\in A\}\), which is also a distance-\((2r+1)\) code, intersects the middle layer \(\binom{[n]}{\lfloor n/2\rfloor}\) in at least \(\Theta(n^{-r} \binom{n}{\lfloor n/2\rfloor})=\Theta(2^n n^{-r-1/2})\) sets; then \(( A\,\triangle\, x)\cap \binom{[n]}{\lfloor n/2\rfloor}\) gives us an antichain code of the desired size. As mentioned earlier, the primary motivation for Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} comes from the Littlewood--Offord theory of anticoncentration. In particular, Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} may be viewed as a purely combinatorial abstraction of an important result of Halász that is widely used in the study of random matrices; see and the many references therein. To explain this connection, we need to fill in some background, a task to which we now turn. Recall that the Littlewood--Offord problem asks the following: given a vector \(\mathbf{a}=(a_1,\ldots,a_n)\) of non-zero real numbers, estimate \[\rho(\mathbf{a})=\max_{\alpha\in \mathbb{R}} \Prob[\epsilon_1 a_1+\cdots +\epsilon_n a_n = \alpha],\] where the \(\epsilon_i\)'s are independent Bernoulli random variables with \(\Prob[\epsilon_i=0]=\Prob[\epsilon_i=1]\). In their study of random polynomials, Littlewood and Offord showed that \(\rho(\mathbf{a})= O(n^{-1/2} \log n)\) for any such \(\mathbf{a}\), and soon after, Erdős used Sperner's theorem to give a simple combinatorial proof of the sharp estimate \(\rho(\mathbf{a}) \le 2^{-n}\binom{n}{n/2} = O(n^{-1/2})\). There has since been considerable interest in establishing better bounds on \(\rho(\mathbf{a})\) under stronger assumptions on the arithmetic structure of \(\mathbf{a}\). For example, Erdős and Moser proved that \(\rho(\mathbf{a})=O(n^{-3/2}\log n)\) whenever all of the entries of \(\mathbf{a}\) are distinct. Sárközy and Szemerédi improved this to \(\rho(\mathbf{a})=O(n^{-3/2})\), which is asymptotically best possible, and subsequently, Stanley used algebraic arguments to establish sharp bounds for this problem. Of particular interest to us is a far-reaching generalisation of the Sárközy--Szemerédi theorem due to Halász , one formulation of which is as follows.
Note in particular that the \(r=1\) case is equivalent to saying \(a_i\ne a_j\) for any \(i\ne j\), so the result in this case reduces to the Sárközy--Szemerédi theorem on the Erdős--Moser problem. Halász's theorem has since become a widely used tool in the study of random matrices and random polynomials. All known proofs of Halász's theorem use Fourier analysis, and are very much arithmetic in nature. While searching for an analogue of Halász's theorem for some other anticoncentration problems (in discussions with Berkowitz), it became clear that it would be of some help to find a purely combinatorial proof of this result, in the spirit of Erdős' classical arguments. Arguably, the primary motivation for Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} is that Halász's theorem is an easy corollary.
# Proof of the main result
We start with some brief comments on notation. We adopt the convention that lower case letters \(a,b,c,x,y,z\) represent subsets of \([n]\), and capital letters \(A,S\) represent families of sets, i.e., subsets of \(2^{[n]}\). If \(S\subset 2^{[n]}\) and \(r\ge 1\) is an integer, we write \(\partial^r S\) for the \(r\)-fold shadow of \(S\), i.e., the collection of sets which can be obtained by deleting \(r\) elements of \([n]\) from some set in \(S\). Similarly, we write \(\partial^{-r} S\) for the collection of sets which can be obtained by adding \(r\) elements to some set in \(S\). For singletons we abuse notation by writing \(\partial^{r} x\) for \(\partial^{r} \{x\}\) and similarly \(\partial^{-r}x\) for \(\partial^{-r}\{x\}\). Finally, we write \(x\,\triangle\, y\) for the symmetric difference of \(x\) and \(y\). Throughout, we omit floors and ceilings when they are not crucial. Before we state and prove the main lemma that drives the proof of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}, we recall one proof of Sperner's theorem that serves as our inspiration. The local-LYM inequality (see ) asserts that for any \(S \subset \binom{[n]}{k}\), we have \[|\partial S| \binom{n}{k-1}^{-1} \ge |S| \binom{n}{k}^{-1}.\] It is not hard to show using local-LYM that any antichain \(A \subset 2^{[n]}\) may be 'shifted', by means of taking shadows, into the middle layer without decreasing the size of the resulting family, whence we conclude that \(|A| \le \binom{n}{\lfloor n/2\rfloor}\). A natural approach to proving Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}, say for antichain codes of distance 3 to be concrete, is to proceed along similar lines as above, except using the distance condition instead of the local-LYM inequality to generate more 'local expansion'. Concretely, given \(A \subset 2^{[n]}\) that is both an antichain and a distance-3 code, it is easy to see for all \(k\) that \(|\partial A_k| \ge k |A_k|\), where \(A_k = A \cap \binom{[n]}{k}\); in particular, for \(k \approx n/2\), this tells us that \(|\partial A_k| \gtrsim n|A_k|/2\), which is a significant improvement over the rather modest bound \(|\partial A_k| \gtrsim |A_k|\) promised by local-LYM. It is then natural to attempt to transform a given antichain code \(A\) of distance 3, by means of taking shadows, into a family that lives in the middle layer that is about \(n\) times larger, which would then show that \(|A| \lesssim \binom{n}{\lfloor n/2\rfloor}/n = O(2^n n^{-3/2})\), as desired. To implement such an idea, we need to deal with how the shadows of the different \(A_k\)'s overlap as we repeatedly take shadows to move \(A\) into the middle layer. For example, an estimate of the following form would be ideal: for any \(S \subset \binom{[n]}{k}\) disjoint from \(A_k\) (i.e., thinking of \(S\) as the shadow of all the \(A_\ell\)'s with \(\ell > k\) in layer \(k\)), we have \(|\partial (S \cup A_k)| \ge |S| + k|A_k|/100\). Unfortunately, this is too much to hope for: if there are no conditions on the arbitrary set \(S\), then it can be arranged so that \(\partial S\) contains the entirety of \(\partial A_k\). Nevertheless, the following lemma shows that something like this ideal estimate does in fact hold when one studies the expansion of antichain codes over (slightly) longer ranges.
We use these two claims to show that \(\partial^{2r}S\) has large distance-\(2r\) 'edge expansion' in \(\binom{[n]}{k-2r}\).
We finish by enumerating in two ways the set \(Q\) of pairs \((y,z)\in\binom{[n]}{k-2r}\times\binom{[n]}{k-3r}\) where \(z\in\partial^{3r}S\), \(z\in\partial^{r} y\), and \(y\not\in\partial^{2r}S\). On the one hand, any pair \((x,y)\) from the previous claim corresponds to such a pair \((y,z)\in Q\) by taking \(z=x\cap y\), and this correspondence is at most \(\binom{k-2r}{r}\)-to-one. By the previous claim, \[|Q|\ge\frac{n^{3r}|A|}{2\cdot 16^{r}r^{3r}\cdot \binom{k-2r}{r}}\ge\binom{n}{r}^{-1}\frac{n^{3r}|A|}{2\cdot 16^{r}r^{3r}}.\] On the other hand, the number of ways to pick \(z\) is \(|\partial^{3r}S|\), the number of ways to pick \(y\in\partial^{-r}z\) is \(\binom{n-k+3r}{r}\), and of these pairs at least \(\binom{k-2r}{r}|\partial^{2r}S|\) satisfy \(y\in\partial^{2r}S\). Thus, \[\begin{aligned}
|Q|&\le \binom{n-k+3r}{r}|\partial^{3r}S|-\binom{k-2r}{r}|\partial^{2r}S|<\binom{n/2}{r}(|\partial^{3r}S|-|\partial^{2r}S|)\\&\le\binom{n/2}{r}(|\partial^{3r}S\cup\partial^{2r}A|-|S|)\le \binom{n/2}{r}\frac{n^{r}|A|}{4 (2r)^{3r}}. \end{aligned}\] To get a contradiction, it suffices to show \[\binom{n/2}{r}\binom{n}{r}< 2\cdot 2^{-r}n^{2r},\] which follows from \(\binom{n/2}{r}\le (n/2)^{r}\) and \(\binom{n}{r}\le n^{r}\). ◻ | {'timestamp': '2022-12-19T02:10:11', 'yymm': '2212', 'arxiv_id': '2212.08406', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08406'} |
# EVALUATION {#sec:exp}
In this section, we will show the details of the used hardware, performance metric and the evaluation of the proposed system in real environments.
## Hardware
Figure [\[fig:hardware\]](#fig:hardware){reference-type="ref" reference="fig:hardware"} shows the hardware used in the evaluation. We use one Google Nest WiFi router and three Google Nest WiFi points as the WiFi access points. The mobile device we use is Google Pixel 6. Other devices supporting IEEE 802.11mc can also be used for the same purpose. The full list of devices supporting IEEE 802.11mc can be found in. WiFi access points are deployed on one floor of a house. There is no specific configuration for Google Nest WiFi routers and Nest WiFi points other than enabling WiFi RTT functions using the official Google WiFi App. For Google Pixel 6, the sampling rate of the IMU is set as 100 Hz, and the sampling rate of WiFi RTT is empirically set as 5 Hz. A higher sampling rate could be set in the mobile device, but that will lead to packet loss because of the burden of processing the ranging requests in the router end[^1].
## Performance Metric
Our proposed system will focus on 2D localisation in the indoor environment. In this section, we use the following metrics to evaluate the proposed indoor localisation system.
- **Trajectories**: Trajectories using the proposed method, IMU based PDR and that using traditional SLAM as well as ground truth will be shown using figures, which will provide an intuitive image of the good performance of the proposed method.
- **Root mean square of errors of the trajectory**: \(RMSE = \sqrt{ \Sigma^{i_e}_{i=i_s}{ e(i)^2 }/{(i_e-i_s)} }\), where \(i_s\) and \(i_e\) are the start point and end point of the whole trajectory, and \(e(i)\) the error at the time point \(i\). \(e(i) = ||l(i)-g(t)||_2\), where \(l(i)\) and \(g(t)\) are estimated locations and ground truth in 2D, respectively.
- **Cumulative distribution function (CDF) of localisation errors**: The CDF of localisation errors demonstrates the frequency that a localisation error is less than or equal to a corresponding error.
- **Localisation error of the end point**: As the proposed localisation method is based on the trajectory, a good quality localisation system should not have a huge cumulative error in the end.
## Experiment in Real Environment
To evaluate the proposed system, the user walks on one floor in a house within 10 m \(\times\) 5 m area. Figure [\[fig:traj_gt\]](#fig:traj_gt){reference-type="ref" reference="fig:traj_gt"} shows the ground truth of the trajectory. The situation is similar to walking in a house, where there is furniture, so that people can only walk in a certain area. The ground truth is obtained by manually adding loop closures to key points, such as narrow corridors where people have to walk by. In this experiment, the total trajectory path is 308.8 m. Figure [\[fig:traj_imu\]](#fig:traj_imu){reference-type="ref" reference="fig:traj_imu"} shows the trajectory only using IMU. Compared with Figure [\[fig:traj_gt\]](#fig:traj_gt){reference-type="ref" reference="fig:traj_gt"}, there are two observations: (1) the shape of the trajectory is similar to the ground truth; (2) the trajectory has accumulative bias compared with the ground truth. This confirms that it is feasible to use an IMU based PDR method for the indoor localisation system, and a secondary mechanism, such as the loop closure method, is required to calibrate the raw trajectory. After considering WiFi RTT based loop closures, false positives are introduced, i.e., due to the noise and multipath effect, distant points could be clustered falsely for further loop closure. When using the traditional pose graph SLAM, as shown in Figure [\[fig:traj_noslam\]](#fig:traj_noslam){reference-type="ref" reference="fig:traj_noslam"}, it cannot provide accurate localisation performance. Different from traditional SLAM, the robust pose graph SLAM, as discussed, uses various weights for loop closures during the optimisation process, so the false positives of loop closures can be given less weight. Figure [\[fig:traj_slam\]](#fig:traj_slam){reference-type="ref" reference="fig:traj_slam"} shows the trajectory using the proposed method, which clearly shows its good performance and outperforms the IMU based PDR method. The trajectory in Figure [\[fig:traj_slam\]](#fig:traj_slam){reference-type="ref" reference="fig:traj_slam"} also confirms that the robust SLAM has the capability to handle the false positives of loop closures and obvious enhancement compared with the trajectory using traditional SLAM shown in Figure [\[fig:traj_noslam\]](#fig:traj_noslam){reference-type="ref" reference="fig:traj_noslam"}. When looking at RMSE of the trajectory, the use of IMU PDR can achieve 2.67 m. When using traditional SLAM, the RMSE of the trajectory is 5.95 m. When using the proposed method, the RMSE of the trajectory can reach 0.13 m, which has significantly been increased compared with the use of IMU and traditional SLAM. The good performance of the proposed method can be further confirmed by the localisation error of the end point. When using IMU based PDR, the localisation error of the end point is 9.84 m. The use of the traditional SLAM and our proposed method can achieve nearly 0 m. The good performance in this metric for the proposed method and traditional SLAM is due to loop closures near the end point. Although the use of traditional SLAM can achieve good performance in this metric, it can still be clearly seen in Figure [\[fig:traj_noslam\]](#fig:traj_noslam){reference-type="ref" reference="fig:traj_noslam"} that, for the most part of its estimate trajectory, the use of the traditional SLAM cannot achieve satisfactory localisation performance. Figure [\[fig:cdf\]](#fig:cdf){reference-type="ref" reference="fig:cdf"} shows the CDF of errors using different methods. It demonstrates that the 90th percentile tracking errors for the IMU based PDR, traditional SLAM and the proposed method are 9.12 m and 4.46 m, and 0.21 m, respectively. This means, when using the CDF of localisation errors as a metric, the proposed method can have 97.6% and 95.2% enhancement compared with the PDR and traditional SLAM, respectively.
# INTRODUCTION {#sec:intro}
Localisation is a fundamental research topic with a wide range of applications, such as mobile robotics, self-driving cars, and smart buildings. Localisation also plays an important role in mobile systems, such as mobile phones, to support many services. Global Positioning System (GPS) is a common localisation solution, where devices with GPS receive signals from satellites to localise themselves. However, due to the weak GPS signal perceived in the indoor environment, the GPS localisation accuracy significantly drops, which cannot provide sufficiently good performance. To solve these challenges, many indoor localisation solutions have been designed for mobile devices. Inertial Measurement Units (IMU) are common sensors integrated into mobile devices as motion sensors and leveraged for Pedestrian Dead Reckoning (PDR). Although it can provide promising localisation performance, the IMU sensor bias generates unavoidable accumulated trajectory drift. In order to correct the drift in IMU-based PDR solutions, supplementary sensors, such as cameras, mmWave radar, etc., are used. Since cameras and mmWave radar can capture high-quality and high-resolution features of the surrounding environment, they can achieve good and stable localisation performance. However, cameras and mmWave radar cost a high volume of energy, which is not suitable for long-term usage on mobile devices. Other radio signal based solutions have been designed, such as Ultra-wide Band (UWB) based and magneto-inductive device based, but they are not commonly equipped within mobile devices. WiFi, as a common communication technique, has been widely integrated into mobile systems. Received Signal Strength Indicator (RSSI) from WiFi has been used with IMU sensors for indoor localisation, such as. The localisation performances of these systems are promising, but they need heavy signal processing methods to mitigate the WiFi signal variance caused by multi-path effect. Although RSSI has a certain relationship with the distance between a device and a WiFi access point, the RSSI variance in a certain location is too huge for accurate localisation. This is because RSSI was not designed for localisation purposes, and multi-path effect is very common in the indoor environment. Recently, a new protocol IEEE 802.11mc with WiFi Round Trip Time (RTT) feature has been designed for ranging between an RTT capable WiFi access point and a mobile device. With the deployment of multiple APs with the WiFI RTT (at least 3 access points for 2D localisation and 4 access points for 3D localisation), a mobile system can locate itself in the area of interest. However, since trilateration is usually used on the ranging results between a mobile device and APs, the exact locations of APs are required. This condition requires frequent reconfiguration and the information exchanged in a new environment or a dynamic environment, such as when WiFi routers are moved. The configuration, i.e., surveying and updating the map of WiFi access points, is labour expensive, especially in a large and/or dynamic environment, not to mention localisation in an uncooperative environment. Additionally, due to multiple path effects in an indoor environment, current localisation solutions only relying on WiFI RTT ranging results may decrease the indoor localisation performance. In this paper, we propose a SLAM system based on data fusion on WiFi RTT and inertial measurements, which achieves good accuracy in indoor localisation. Different from existing research, we do not need the prior knowledge of the locations of WiFi access points and the dependence on accurate ranging results. Instead, the ranging measurements are used as **observations** directly for clustering. As discussed, the IMU only based PDR can provide an accurate trajectory in a relatively short term, but with unavoidable long term drift. In our paper, we propose to use clusters from WiFi ranging measurements for loop closure and the PDR drift correction. One novelty of the system is the removal of the configuration requirement of the deployment of WiFi access points, which enables the proposed system suitable for a dynamic environment, such as home environment, where APs are not fixed and/or easily moved, or a new environment, where mobile devices do not contain the location information of WiFi access points. Furthermore, we consider the false positives of loop closures and use a robust pose graph SLAM framework to provide high quality indoor localisation performance. To summarise, the contributions of this paper are as follows:
- We design an indoor localisation system based on the observations of WiFi ranging measurements within a pose graph SLAM framework. Our designed system removes the prior knowledge of the locations of WiFi access points and the dependence on accurate ranging results.
- We design an effective clustering method for loop closure along with a robust SLAM framework to tolerate false positives of loop closing constraints and provide accurate localisation performance.
- We have conducted extensive experiments in a real environment, which shows our proposed method has good localisation performance.
The rest part of this paper is organised as follows. Section [4](#sec:related){reference-type="ref" reference="sec:related"} discusses the related works, and we show the overview of the designed system in Section [5](#sec:system){reference-type="ref" reference="sec:system"}. Section [3](#sec:method){reference-type="ref" reference="sec:method"} details the proposed accurate indoor localisation system. Next, in Section [1](#sec:exp){reference-type="ref" reference="sec:exp"}, extensive experiments are performed in a real environment, which shows a good and robust localisation performance of the proposed system. Finally, we conclude the paper in Section [6](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}.
# METHOD {#sec:method}
In this section, we will show the details of the proposed indoor localisation method, which includes Pedestrian Dead Reckoning, an efficient loop closure method based on WiFi RTT measurements and a robust SLAM.
## Pedestrian Dead Reckoning
The first step of the system is to provide an initial trajectory based on inertial navigation. A robust Pedestrian Dead Reckoning (PDR) method is firstly applied on the collected acceleration, angular velocity and magnetic field measurements from the accelerometer, gyroscope and magnetometer. Theoretically, an IMU in a mobile device, containing an accelerometer, a gyroscope and a magnetometer, can estimate the carrier's trajectory based on acceleration and gyroscope. However, the low specification IMU in mobile devices cannot provide sufficiently accurate measurements because of the noise in the accelerometer and long term drift in the gyroscope. To solve this problem, the whole trajectory estimate is split into multiple step motion estimates. This is based on the fact that the velocity is 0 when both feet are on the ground during walking and the long term bias can be mitigated at a certain level by dividing the long term estimate into smaller chunks. Therefore, the PDR can be conducted by the following two steps, (1) performing step detection and (2) estimating the length \(L\) and attitude \(\alpha\) of each step. In our system, acceleration magnitudes are used for step detection. To be specific, the norm of \(i\)th 3D acceleration measurement \((a_{xi}, a_{yi}, a_{zi})\) is calculated as \(a_i = (a_{xi}^2+a_{yi}^2+a_{zi}^2)^{\frac{1}{2}}\). A threshold of standard deviations within a window size \(w\) is set to detect the step (relatively steady and the corresponding standard deviation is below the threshold) and moving (the corresponding standard deviation is above the threshold). Once the steps are detected, the step length is estimated using Weinberg algorithm by using vertical movement of the hip. It is empirically shown in that the \(j\)th step length can be calculated in Equation [\[equ:step\]](#equ:step){reference-type="ref" reference="equ:step"}. \[\label{equ:step} L_j = h \times ( \text{max}(\widetilde{a}(j))-\text{min}(\widetilde{a}(j)) )^{1/4}\] ,where \(a(j)\) represents the vectors of the magnitude of acceleration measurements (i.e., the norm of 3D acceleration measurements) between \(j\)th step and (\(j\)-1)th step after a low pass filter to remove the measurement noise. The maximum and minimal values of this vector are obtained by using max() and min(), respectively. \(h\) is a constant. In the meanwhile, the attitude \(\alpha\) of each step is estimated by the angular velocity in the global reference. Please note that the raw data from IMUs are based on the IMU reference system. Therefore, the measurements of angular velocity from the IMU reference system need to be transformed to the global reference with the assistance of the magnetometer [^2].
## Efficient WiFi RTT Measurement Clustering for Loop Closure
Along with trajectory from the PDR method, WiFi RTT measurements are used as observations directly for the loop closure. WiFi RTT was initially designed for indoor localisation based on its ranging capability between a mobile device and a WiFi access point. The common solution using WiFi RTT measurements is trilateration, i.e., the location estimation based on the locations of WiFi access points and the corresponding WiFi RTT ranging measurements. However, due to multipath effect, the ranging from WiFi RTT may not be sufficiently accurate. Different from existing solutions, we do not rely on the knowledge of locations of WiFi access points or accurate ranging results. Instead, we use the WiFi RTT measurements as observations to directly identify loop closures, based on the fact that the mobile device can perceive similar WiFi RTT measurements at the same place, as shown in Figure [\[fig:cluster\]](#fig:cluster){reference-type="ref" reference="fig:cluster"}. These similar observations will be used to indicate loop closures, i.e., the user revisits the same place. Since the locations of WiFi access points are not known in our setting, it is impossible to calculate the exact locations based on WiFi RTT ranging measurements. Therefore, we cannot estimate the relevant locations between two observations except if they are in the vicinity. When a mobile phone revisits the same place, the observations from all WiFi RTT measurements are similar. Our loop closure methods are based on the similarity of the observations containing all the measurements from WiFi access points. To enable loop closure, we design an efficient clustering method for WiFi RTT observations. The \(i\)th WiFi RTT observation is denoted as \(w_i = (w_{1i},w_{2i}...,w_{ki}...,w_{ni})\), where \(n\) is the number of WiFi access points and \(w_{ki}\) is the WiFi RTT measurement between the mobile device and the \(k\)th WiFi access point in this \(i\)th observation. Once the \(i\)th WiFi RTT observation \(w_i\) is collected, it is used for clustering with previous WiFi RTT observations, i.e., \(\{w_1, w_2,..., w_{i-1}\}\), by using a Euclidean distance \(d_{i,k} = ||w_i-w_k ||_2\). When \(d_{i,k}\) is below a threshold, the corresponding points when \(i\)th and \(k\)th WiFi RTT measurements in the raw estimate trajectory are labelled as a constraint for loop closure, which means they are candidates of places in the vicinity. However, due to the multipath effect, WiFi RTT measurements are not necessarily accurate, and many false positives of loop closures exist during the clustering method. To solve this, we use a robust SLAM framework to achieve global optimisation.
## Robust SLAM Framework
In this section, we demonstrate the used pose graph SLAM framework in detail. The traditional pose graph SLAM conducts optimisation based on all the constraints equally by solving the following optimisation problem in Equation [\[equ:slam\]](#equ:slam){reference-type="ref" reference="equ:slam"}. \[\label{equ:slam} \operatorname*{argmin}_A \sum_{i \in L} r_i^T M_i r_i\], where \(A\) is the estimated trajectory for the localisation purpose, and \(L\) is the constraint set, including IMU raw trajectory and loop closures. \(M\) is the information matrix of all constraints, and \(r\) represents the corresponding error terms. When applying the traditional SLAM framework to our application, the false positives of constraints from WiFi RTT observations are equally treated, which will lead to unnecessary loop closures and decrease indoor localisation performance significantly. In the example in Figure [\[equ:slam\]](#equ:slam){reference-type="ref" reference="equ:slam"}(d), the estimated trajectory is collapsed due to the false positive of loop closures. The robust pose graph SLAM optimiser solves this problem by introducing scaling factors for loop closures. Equation [\[equ_slam_robust\]](#equ_slam_robust){reference-type="ref" reference="equ_slam_robust"} shows the optimiser of the used robust pose graph SLAM. \[\label{equ_slam_robust} \operatorname*{argmin}_A \underbrace{ \sum_{i \in L_{traj}} r_i^T M_i r_i }_{\text{(Trajectory Constraints)}} + \underbrace{ \sum_{i \in L_{loop}}s_i^2 r_i^T M_i r_i }_{\text{(Loop Closures)}}\] When using the robust pose graph SLAM optimiser, constraints from raw IMU based trajectories and loop closures. Constraints from raw IMU trajectories are used as that in the traditional graph SLAM, while the consideration of loop closures involves a scaling factor \(s\) defined in [\[equ:scalor\]](#equ:scalor){reference-type="ref" reference="equ:scalor"}. \[\label{equ:scalor} s_i = \operatorname*{min}(1, \frac{2C}{C + \Theta_i^2} )\], where \(C\) is a free parameter, and \(\Theta\) is also the error term of loop constraints. The error terms for loop closures depend on the used threshold for the loop closure detection. The use of the scaling factor \(s\) can eliminate the impact of false positives of loop closure by providing less weight to them and relying more on the constraints from IMU based trajectories when conducting the optimisation. Please refer to for more technical details regarding the traditional pose graph SLAM and robust pose graph SLAM.
# RELATED WORKS {#sec:related}
In this section, we will discuss the related works regarding indoor localisation systems. As a ubiquitous signal, WiFi has been studied for indoor localisation. used WiFi received signal strength along with IMUs for indoor localisation. However, due to the multipath effect, the use of received signal strength cannot maintain stable quality for indoor localisation. The wireless communication research community has also been aware of the importance of WiFi for indoor localisation, standard IEEE 802.11mc has been designed for ranging and WiFi routers and mobile devices have started to integrate the new standard. The common methods using WiFi RTT measurements are based on multilateration, i.e., localisation based on estimate distances between the mobile device and WiFi access points. The data fusion of WiFi RTT and IMU measurements has also been studied. Different from it, our proposed method does not rely on accurate ranging estimate from WiFi RTT measurements. Because the multiple path effect still exists, albeit having better performance compared with the use of received strength, it is not easy to estimate the ranging quality of WiFi RTT. Instead, we focus on the localisation of the mobile device rather than other factors, such as the location estimate of WiFi access points. We use WiFi RTT measurements directly as observations for loop closure and apply the pose GraphSLAM to obtain an accurate trajectory. Other wireless signals have also been used for localisation. Bluetooth, as another common wireless communication technique, was also employed to indicate the locations. However, due to the short communication range of Bluetooth, a large volume of Bluetooth anchor points are needed for localisation purposes. UWB is another wireless communication technique. As UWB is designed for indoor localisation, it can achieve centimetre-level accuracy. Along with IMU measurements, it can deliver a high-precision context awareness system, including localisation. The basic principle of UWB for localisation is also multilateration based on the ranging between mobile devices and WiFi access points. One fallback for UWB based applications is that UWB is not as common as WiFi devices, which needs specialised devices and extra installation. Additional to wireless technologies, other sensors with high resolutions are used for localisation, such as laser based, mmWave radar based, camera based. Although they can provide very high accuracy localisation performance, their high energy consumption is one constraint for the large-scale application in mobile devices. Researchers have also investigated other signals for localisation purposes, such as magnetic field, photovoltage, etc. There also exist radio based device free localisation methods, i.e., users do not need to carry any devices and the system localises the user in the area of interest based on the radio interference by human beings. Localisation is also the core technology for self-driving cars, which uses multiple sensors, such as GPS, IMU, Lidar, and cameras. With the assistance of a powerful computer, self-driving cars can achieve good localisation performance. Different from localisation techniques used in self-driving cars, our paper focuses on the use of readily available sensors on mobile phones, i.e., WiFi and IMU, for indoor localisation.
# SYSTEM OVERVIEW {#sec:system}
In this section, we overview the whole structure of the system rWiFiSLAM, and the flowchart is shown in Figure [\[fig:overview\]](#fig:overview){reference-type="ref" reference="fig:overview"}. The user, carrying a mobile phone, walks in an area of interest with the deployment of WiFi access points. The mobile phone, equipped with our designed system, can localise itself. To enable the indoor localisation system, the mobile system continuously collects acceleration, angular velocity measurements and magnetic field measurements from its own IMU, which provides an initial trajectory estimate. WiFi RTT measurements are collected simultaneously, and a clustering method will be applied on WiFi RTT measurements for the loop closure detection and adjusting the biased initial trajectory. A robust pose graph SLAM will then be applied on both initial trajectory and WiFi RTT constraints, which provides an accurate indoor localisation performance.
# CONCLUSIONS {#sec:conclusion}
In this paper, we use WiFi RTT measurements as observations for loop closure. A robust pose graph SLAM system is applied on both types of constraints, i.e., raw trajectories from IMU measurements and loop closures from WiFi RTT measurements. An efficient clustering method for WiFi RTT measurements has been proposed for loop closure, which, different from traditional methods, does not rely on the knowledge of the locations of WiFi access points and accurate ranging results. The robust pose graph SLAM can also help handle false positives of loop closures, which provides a good indoor localisation performance. To evaluate our proposed method, we have conducted experiments in a real indoor environment, which can achieve 0.13 m accuracy and clearly shows the proposed method outperforms the IMU based PDR and the localisation using traditional SLAM by more than 90%.
[^1]: One WiFi ranging measurement is averaged over a sequence of the ranging requests in the lower level, but the API to access each measurement is not available in the current Android Version Android 12 (version 8.4).
[^2]: More details of the reference transformation can be found in. | {'timestamp': '2022-12-19T02:10:34', 'yymm': '2212', 'arxiv_id': '2212.08418', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08418'} |
# The problem and the first solution {#sec:problem}
In 1977, P. Yang asked the following (see ):
Completeness is a very natural condition to impose on a Riemannian manifold when one is interested in its global properties. Recall that an immersed submanifold \(\varphi\colon M\to\mathbb{C}^n\) is said to be complete if the Riemannian metric \(\varphi^*d\sigma^2\) induced on \(M\) by pulling back the Euclidean metric \(d\sigma^2\) on \(\mathbb{C}^n\) by the immersion \(\varphi\) is a complete metric on \(M\): geodesics go on indefinitely. By the Hopf-Rinow theorem, \(\varphi\) is complete if and only if \(\varphi\circ\gamma\colon[0,1)\to\mathbb{C}^n\) has infinite Euclidean length for every divergent path \(\gamma\colon[0,1)\to M\). Every compact Riemannian manifold is complete, but compact complex manifolds cannot be found in \(\mathbb{C}^n\) by the maximum principle for holomorphic functions, so the question in Problem [\[qu:Yang\]](#qu:Yang){reference-type="ref" reference="qu:Yang"} is in order. The main original motivation for P. Yang to pose the aforementioned question is that a positive answer would prove the existence of complete immersed complex submanifolds \(M^k\to\mathbb{C}^{2n}\) with strongly negative holomorphic sectional curvature. On the other hand, since complex submanifolds of \(\mathbb{C}^n\) are minimal (i.e., critical points for the volume functional; see e.g., among many others, for an introduction to the subject), the Yang problem is also related to the so-called Calabi-Yau problem, which dates back to E. Calabi's conjectures from 1965 and asked whether there are complete bounded minimal hypersurfaces in \(\mathbb{R}^n\) \((n\ge 3)\); we refer to for a recent survey on this fascinating topic. Nevertheless, the Yang problem for complex submanifolds has become an active focus of interest in its own right, having received many important contributions in the last decade; see for a brief introduction to the topic. We shall denote by \(\mathbb{D}=\{\zeta\in\mathbb{C}\colon |\zeta|<1\}\) the open unit disc in \(\mathbb{C}\) and by \(\mathbb{B}_n=\{z=(z_1,\ldots,z_n)\in\mathbb{C}^n\colon |z|<1\}\) the open unit ball in \(\mathbb{C}^n\) for \(n\ge 2\). The first affirmative answer to the question in Problem [\[qu:Yang\]](#qu:Yang){reference-type="ref" reference="qu:Yang"} was given only two years later, in 1979, by P.W. Jones. Recall that a map \(f\colon X\to Y\) between topological spaces is said to be proper if \(f^{-1}(K)\subset X\) is compact for every compact set \(K\subset Y\).
P.W. Jones' construction method is strongly complex analytic. It relies on using the BMO duality theorem in order to find a pair of bounded holomorphic functions \(f_1\) and \(f_2\) on \(\mathbb{D}\) satisfying the property that \[\int_\gamma\big( |f_1'(\zeta)|+|f_2'(\zeta)|\big)\, d\sigma(\zeta)=+\infty\] for all paths \(\gamma\subset\mathbb{D}\) terminating on \(\mathbb{S}^1=b\mathbb{D}\), where \(\sigma\) denotes Euclidean arc length. It follows that \(\mathbb{D}\ni\zeta\mapsto(\zeta,f_1(\zeta),f_2(\zeta))\) is a complete bounded embedded complex disc in \(\mathbb{C}^3\); the other two assertions in the theorem are obtained by slight modifications of this procedure.
# Curves {#sec:curves}
Despite having to wait more than three decades for it, Theorem [\[th:Jones\]](#th:Jones){reference-type="ref" reference="th:Jones"} has been generalized in several directions. The first extension of P.W. Jones' existence result was given by A. Alarcón and F.J. López in 2013 and concerns the topology of the examples.
The embeddedness condition for \(n\ge 3\) in this statement was not explicitly stated in; nevertheless, it trivially follows from a standard transversality argument, as was later pointed out by A. Alarcón and F. Forstnerič in. We emphasize that the examples in Theorem [\[th:AL2013\]](#th:AL2013){reference-type="ref" reference="th:AL2013"} may have any topological type, even infinite. The proof in is completely different from that in. In particular, it is much more geometric, and is reminiscent of the method developed by N. Nadirashvili in his seminal paper for constructing a complete bounded minimal disc in \(\mathbb{R}^3\); see also. The construction goes by induction, the rough idea being the following. In the step \(j\in\mathbb{N}\) we begin with a smoothly bounded compact complex curve, say \(X_{j-1}\), whose boundary \(bX_{j-1}\) lies inside the ball \(R_{j-1}\mathbb{B}_n\) of some radius \(R_{j-1}>0\) but close to its boundary sphere. Then, we apply to \(X_{j-1}\) a deformation that is arbitrarily small outside a neighborhood of \(bX_{j-1}\) and pushes each boundary point \(x\in bX_{j-1}\) a distance approximately \(1/j\) in a direction perpendicular to the position vector of \(x\) in \(\mathbb{C}^n\). In this way we increase the boundary distance from a fixed interior point of the curve an amount of approximately \(1/j\), while the extrinsic diameter is increased, by Pythagoras' theorem, an amount of the order of \(1/j^2\). Moreover, we ensure that the boundary of the new complex curve, \(X_j\), lies inside the ball of radius \[R_j=\sqrt{R_{j-1}^2+1/j^2}>R_{j-1}\] but close to its boundary sphere. See Figure [\[fig:Pi\]](#fig:Pi){reference-type="ref" reference="fig:Pi"}. Since \[\sum_{j\ge 1}1/j=\infty\quad \text{and}\quad \sum_{j\ge 1}1/j^2<\infty,\] if we arrange this process in the right way then we obtain in the limit a complete complex curve that is contained in the ball of radius \(R=\lim_{j\to\infty}R_j<\infty\) and is proper in it. In order to prescribe the topology of the curve, we begin with a disc and at each step of the inductive construction apply a surgery which enables us to add either a handle or a boundary component to a given compact bordered complex curve in \(\mathbb{C}^n\). Finally, if \(n\ge 3\), then a general position argument allows to guarantee that all complex curves in the sequence are embedded, and hence the limit one can be granted to be embedded as well. The main tool in order to make the described deformations is the theory of uniform approximation for holomorphic functions on open Riemann surfaces; in particular, the Runge-Mergelyan theorem (see E. Bishop or e.g. Theorem 5 in the survey on holomorphic approximation by J.E. Fornæss, F. Forstnerič, and E.F. Wold ). A different construction of complete bounded complex curves in \(\mathbb{C}^2\) can be found in, where F. Martín, M. Umehara, and K. Yamada gave examples with arbitrary finite genus and finitely many ends. Their technique relies on the existence of a simply connected complete bounded holomorphic null curve in \(\mathbb{C}^3\) (such a curve was first constructed in, an alternative construction was given later by L. Ferrer, F. Martín, M. Umehara, and K. Yamada in; see the mentioned sources or e.g. for definitions) and modifies a method developed by F.J. López in for constructing complete minimal surfaces in \(\mathbb{R}^3\) of hyperbolic conformal type. Despite the flexibility of the methods in, none of them allows to control the complex structure on the curve, except of course in the simply connected case: every bounded immersed complex curve in \(\mathbb{C}^n\) must have hyperbolic conformal type, and hence if it is simply connected it is biholomorphic to the disc \(\mathbb{D}\). Indeed, since the construction in relies on Runge's theorem, at a certain stage one does not have enough information on the placement in \(\mathbb{C}^n\) of some parts of the curve, and hence one is forced to cut away some small pieces of the curve to ensure its boundedness. This makes one to lose the control of the complex structure of the curve. This difficulty was overcome by A. Alarcón and F. Forstnerič in, also published in 2013, where two additional complex analytic tools were introduced into the game, namely, the method of F. Forstnerič and E.F. Wold for exposing boundary points of a complex curve in \(\mathbb{C}^n\) and the use of approximate solutions to Riemann-Hilbert boundary value problems. (The former is a modern technique that has led to important progress in the classical Forster-Bell-Narasimhan Conjecture asking whether every open Riemann surface admits a proper holomorphic embedding into \(\mathbb{C}^2\); we refer to F. Forstnerič for a survey on this long-standing open problem. On the other hand, the use of the Riemann-Hilbert problem for constructing proper holomorphic maps has a long history; we refer to F. Forstnerič and J. Globevnik, B. Drinovec Drnovšek and F. Forstnerič, and the references therein.) The implementation of these new tools enabled to substantially simplify the construction in and, moreover, to control the complex structure on the curve. Recall that a bordered Riemann surface is an open connected Riemann surface \(M\) that is the interior, \(M=\overline M\setminus b\overline M\), of a compact one dimensional complex manifold \(\overline M\) with smooth boundary \(b\overline M\) consisting of finitely many closed Jordan curves; such an \(\overline M\) is called a compact bordered Riemann surface.
More generally, it is shown in that if \(X\) is a Stein manifold of dimension \(n\ge 2\) endowed with a hermitian metric, then every bordered Riemann surface \(M\) admits a complete proper holomorphic immersion \(M\to X\) that can be chosen an embedding if \(n\ge 3\). Recall that a Stein manifold is the same thing as a closed complex submanifold of a complex Euclidean space; we refer to for a comprehensive monograph on the theory of Stein manifolds. A domain \(D\) in \(\mathbb{C}^n\) \((n\ge 2)\) is a Stein manifold if and only if it is pseudoconvex, meaning that there is a strictly plurisubharmonic exhaustion function \(D\to\mathbb{R}\). This happens if and only if \(D\) is a domain of holomorphy and if and only if \(D\) is holomorphically convex. For instance, every convex domain in \(\mathbb{C}^n\) is pseudoconvex. We refer to the monographs by R.M. Range (see also the introductory note ) and L. Hörmander for background on the subject. More recently, the construction technique in has been refined to produce complete bounded complex curves with control on the complex structure and with some further control on the asymptotic behavior. In particular, there are such curves bounded by Jordan curves. The following is a compilation of results by A. Alarcón, I. Castro Infantes, B. Drinovec Drnovšek, F. Forstnerič, and F.J. López.
Summarizing, by the year 2015 there were available in the literature several constructions of complete bounded complex curves immersed in \(\mathbb{C}^2\) and embedded in \(\mathbb{C}^3\), allowing a high control on the asymptotic behavior (proper in the ball or in a given pseudoconvex domain, bounded by Jordan curves, etc), on the topology, and on the complex structure of the examples. However, the construction of complete bounded embedded complex curves in \(\mathbb{C}^2\) turns out to be a much more challenging undertaking, and the question whether such curves exist remained open (see ). Recall that complex curves are generically embedded in \(\mathbb{C}^n\) for \(n\ge 3\), meaning that self-intersections can be removed by applying small deformations, while self-intersections of complex curves in \(\mathbb{C}^2\), which are generically double points, are stable under such deformations. That is the main reason why the task is a more difficult one. The Yang problem for embedded complex curves in the affine plane \(\mathbb{C}^2\) was finally settled by A. Alarcón and F.J. López in a paper published in 2016.
The proof goes by induction and involves an approximation process by embedded complex curves in \(\mathbb{C}^2\). The main step in the construction is to prove that every compact embedded complex curve \(X\) in \(\mathbb{C}^2\) with the boundary \(bX\) lying in the boundary \(bD\) of a regular strictly convex domain \(D\subset\mathbb{C}^2\) may be approximated by another compact embedded complex curve \(\tilde X\) with \(b\tilde X\subset b\tilde D\), for any given convex domain \(\tilde D\subset\mathbb{C}^2\) with \(\overline D\subset\tilde D\). The new curve \(\tilde X\) is ensured to contain a biholomorphic copy of \(X\), which we denote by \(X\) as well, and the main point is to guarantee that \(\tilde X\setminus X\subset\tilde D\setminus D\) and the intrinsic Euclidean distance in \(\tilde X\) from \(X\) to \(b\tilde X\) is suitably larger (in a Pythagorical way similar to that explained in Figure [\[fig:Pi\]](#fig:Pi){reference-type="ref" reference="fig:Pi"}) than the distance from \(D\) to \(b\tilde D\). These conditions are the key for obtaining embeddedness, completeness, and properness of the limit complex curve in the limit convex domain. In order to guarantee the embeddedness of \(\tilde X\) a standard self-intersection removal method consisting of replacing every normal crossing in an immersed complex curve in \(\mathbb{C}^2\) by an embedded annulus is applied. This surgery may generate shortcuts in the arising desingularized curve \(\tilde X\), thereby giving rise to divergent paths of shorter length. This is an important difficulty for ensuring completeness; for instance, if one applies this surgery at each step in the inductive construction in the proof of Theorem [\[th:AL2013\]](#th:AL2013){reference-type="ref" reference="th:AL2013"} or Theorem [\[th:AF2013\]](#th:AF2013){reference-type="ref" reference="th:AF2013"}, then one still obtains in the limit a properly embedded complex curve in \(\mathbb{B}_2\) (or, more generally, in any given pseudoconvex domain of \(\mathbb{C}^2\)), but it need not be complete. A main novelty in the construction in is provided a good enough (say, in a Pythagorical sense) estimate of the growth of the intrinsic Euclidean diameter of the desingularized, embedded complex curve \(\tilde X\); this is achieved by keeping a stronger control on the placement of \(\tilde X\setminus X\) in \(\tilde D\setminus D\). The main tool in this construction continues to be the classical Runge-Mergelyan approximation theorem for holomorphic functions on open Riemann surfaces. The aforementioned surgery may increase the topological genus of the curve, and so there is no control on the topology of the examples in Theorem [\[th:AL2016\]](#th:AL2016){reference-type="ref" reference="th:AL2016"}. They actually seem to have infinite genus, and can be ensured to have infinite topology. It therefore remained an open question whether there are complete bounded embedded complex curves in \(\mathbb{C}^2\) of finite topology (see ).
# Submanifolds of arbitrary dimension {#sec:higher}
All examples of complete bounded complex submanifolds we have discussed so far are of complex dimension one; i.e., complex curves in \(\mathbb{C}^n\). The first known such submanifolds of higher dimension were also given in.
The corollary is obtained by the following simple trick which was pointed out by J.E. Fornæss. In, B. Drinovec Drnovšek and F. Forstnerič proved that every domain \(D\) as in the statement admits a proper holomorphic immersion \(g\colon D\to \mathbb{D}^{2k}\) into the polydisc \(\mathbb{D}^{2k}=\mathbb{D}\times \stackrel{2k}{\ldots}\times\mathbb{D}\subset\mathbb{C}^{2k}\). Choose a complete proper holomorphic immersion \(\varphi\colon \mathbb{D}\to\mathbb{B}_2\) provided by Theorem [\[th:AF2013\]](#th:AF2013){reference-type="ref" reference="th:AF2013"}, and consider the proper holomorphic map \(\varphi^{2k}\colon \mathbb{D}^{2k}\to (\mathbb{B}_2)^{2k}\) given by \[\varphi^{2k}(\zeta_1,\ldots,\zeta_{2k})=(\varphi(\zeta_1),\ldots,\varphi(\zeta_{2k})),\quad (\zeta_1,\ldots,\zeta_{2k})\in \mathbb{D}^{2k}.\] It is then easily checked that \(\varphi^{2k}\circ g\colon D\to (\mathbb{B}_2)^{2k}\) is a complete proper holomorphic immersion. A slight modification of this argument using a proper holomorphic embedding \(D\hookrightarrow \mathbb{D}^{2k+1}\) (existence of such is also proved in ) provides a proper holomorphic embedding \(D\hookrightarrow (\mathbb{B}_3)^{2k+1}\). This same trick together with Theorem [\[th:AL2016\]](#th:AL2016){reference-type="ref" reference="th:AL2016"} allows to prove the following.
The ad hoc construction of the high dimensional examples in Corollaries [\[co:AF2013\]](#co:AF2013){reference-type="ref" reference="co:AF2013"} and [\[co:AL2016\]](#co:AL2016){reference-type="ref" reference="co:AL2016"} seemed to give very particular solutions to the Yang problem for complex submanifolds of dimension \(\ge 2\). This led to some new questions, as whether the dimension \(2k\) in Corollary [\[co:AL2016\]](#co:AL2016){reference-type="ref" reference="co:AL2016"} is optimal, and exposed the need of looking for new construction methods other than those based on the existence of complete bounded complex curves. For instance, the natural question whether the ball \(\mathbb{B}_k\) \((k\ge 2)\) admits a complete proper holomorphic embedding into the ball \(\mathbb{B}_n\) for some \(n>k\), and, in particular, for \(n=k+1\) appeared; see and. An affirmative answer to this question in the case of sufficiently high codimension was given in 2015 by B. Drinovec Drnovšek.
The proof in continues to use the geometric idea of deforming a compact submanifold near the boundary in orthogonal directions to the position vector (see Figure [\[fig:Pi\]](#fig:Pi){reference-type="ref" reference="fig:Pi"}), but it exploits different tools. The main new ingredients are holomorphic peak functions, going back to ideas of M. Hakim and N. Sibony and E. Løw, and the construction of inner functions on the ball, as well as J.E. Fornæss embedding theorem. The construction relies on suitably modifying earlier methods by F. Forstnerič and E. Løw for constructing proper holomorphic maps from a strictly convex domain with \(\mathcal{C}^2\) boundary in \(\mathbb{C}^k\) into a unit ball of some Euclidean space of higher dimension, in order to make them complete. The construction method in requires a sufficiently high codimension, and hence the following question remains open; see and.
# Hypersurfaces {#sec:hypersurfaces}
Recall that complex submanifolds of dimension \(k\in\mathbb{N}\) in \(\mathbb{C}^n\) are generically embedded for \(n\ge 2k+1\) (meaning that one can get rid of their self-intersections by applying small deformations on compact pieces), while for \(n\le 2k\) self-intersections of \(k\)-dimensional complex submanifolds in \(\mathbb{C}^n\) are stable under small deformations. This is the main reason why the Yang problem for embedded submanifolds is much more difficult in low codimension; in particular, for hypersurfaces. In the lowest dimensional case of \(n=2\) (i.e., for complex curves in \(\mathbb{C}^2\)), this question was first solved by A. Alarcón and F.J. López in (see Theorem [\[th:AL2016\]](#th:AL2016){reference-type="ref" reference="th:AL2016"} above), but the question whether there are complete bounded embedded complex hypersurfaces in \(\mathbb{C}^n\), or even whether there are such complex submanifolds of dimension \(k\) with \(2k\ge n\), remained open for every \(n\ge 3\). Indeed, note that the codimension in all examples of complete bounded embedded complex submanifold of dimension \(\ge 2\) which have been mentioned so far is high. It was J. Globevnik who, in a pair of landmark papers in 2015-2016, positively settled Yang's question in Problem [\[qu:Yang\]](#qu:Yang){reference-type="ref" reference="qu:Yang"} for embeddings in arbitrary dimension and codimension; in particular, for hypersurfaces.
Globevnik's approach is completely different from any previous method used in the study of the Yang problem, and it was a major breakthrough in this topic. In particular, his construction of a complete closed complex hypersurface in a given pseudoconvex domain \(D\subset\mathbb{C}^n\) is implicit: the examples are obtained as level sets of highly oscillating holomorphic functions on \(D\). (The existence of complete closed complex submanifolds of any higher codimension is then an obvious consequence.) To be more precise, Globevnik proved the following.
If a function \(f\colon D\to\mathbb{C}\) on a domain \(D\subset\mathbb{C}^n\) is holomorphic and nonconstant, then all its nonempty level sets \(f^{-1}(c)=\{z\in D\colon f(z)=c\}\) \((c\in\mathbb{C})\) are closed complex hypersurfaces of \(D\), possibly with singularities. Nevertheless, by Sard's theorem most of them are smooth, and hence properly embedded. So, if \(c\in\mathbb{C}\) is such that \(f^{-1}(c)\neq\varnothing\), then every divergent path \(\gamma\colon [0,1)\to f^{-1}(c)\) diverges on \(D\) as well. If \(D\) is pseudoconvex and the function \(f\) is as in Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} then, since \(\Re(f)\) is constant (and hence bounded) on \(\gamma\), we have that \(\gamma\) has infinite length, and \(f^{-1}(c)\) is thus complete. Therefore, the level sets \(f^{-1}(c)\) \((c\in\mathbb{C})\) of \(f\) form a (possibly singular) holomorphic foliation of \(D\) by complete closed complex hypersurfaces. Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} thus implies the following corollary which, in turn, implies Theorem [\[th:PikoAM\]](#th:PikoAM){reference-type="ref" reference="th:PikoAM"}.
Let us outline the proof of Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} in the case of \(D=\mathbb{B}_n\) given in. Recall that a convex polytope \(P\) in \(\mathbb{R}^d\), \(d\ge 2\), is a compact convex set which is the intersection of finitely many closed half-spaces. A face of \(P\) is a closed convex subset \(F\subset P\) such that every closed segment in \(P\) whose relative interior intersects \(F\) is contained in \(F\). The boundary \(bP\) of \(P\) is the union of its faces of dimension \(d-1\), while the skeleton \({\rm skel}(P)\) of \(P\) is the union of all \((d-2)\)-dimensional faces of \(P\). Most of the work in the proof consists of constructing a sequence of convex polytopes \(P_j\) in \(\mathbb{C}^n=\mathbb{R}^{2n}\) and positive numbers \(\theta_j\) \((j\in\mathbb{N})\) satisfying the following conditions:
(i) \(P_1\subset {\rm Int}(P_2)\subset P_2\subset {\rm Int}(P_3)\subset\cdots\subset\bigcup_{j\in\mathbb{N}}P_j=\mathbb{B}_n\).
(ii) Denote by \(U_j\) the \(\theta_j\)-neighborhood of \({\rm skel}(P_j)\) in \(bP_j\), and set \(V_j=(bP_j)\setminus U_j\), \(j\in\mathbb{N}\). If \(\gamma\colon[0,1)\to \mathbb{B}_n\) is a divergent path such that \(\gamma([0,1))\cap V_j=\varnothing\) for all \(j\ge j_0\) for some \(j_0\in\mathbb{N}\), then \(\gamma\) has infinite length.
Each set \(V_j\) \((j\in\mathbb{N})\) is compact and its connected components are closed convex sets in hyperplanes of \(\mathbb{R}^{2n}\). The union \[\label{eq:Lab} L=\bigcup_{j\in\mathbb{N}} V_j\] of all of them is a sort of labyrinth of compact connected \((2n-1)\)-dimensional convex sets in \(\mathbb{B}_n\subset\mathbb{R}^{2n}\) with the property that every divergent path \(\gamma\colon[0,1)\to\mathbb{B}_n\) meeting at most finitely many components of \(L\) has infinite length. See Figure [\[fig:L-G\]](#fig:L-G){reference-type="ref" reference="fig:L-G"}. The construction of the labyrinth in is very involved and belongs to convex geometry. With the labyrinth in hand, to complete the proof of Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} an idea of J. Globevnik and E.L. Stout from is used in order to construct, via Runge's theorem, a sequence of holomorphic polynomials \(f_j\colon \mathbb{C}^n\to\mathbb{C}\), \(j\in\mathbb{N}\), such that the following conditions hold for each \(j\in\mathbb{N}\):
(i) \(\Re(f_j(z))\ge j+1\) for all \(z\in V_j\).
(ii) \(|f_{j+1}(z)-f_j(z)|\le 1/2^{j+1}\) for all \(z\in P_j\).
Conditions (i) and (iv) ensure that there is a limit holomorphic function \[f=\lim_{j\to\infty}f_j\colon\mathbb{B}_n\to\mathbb{C}.\] By conditions (iii) and (iv), we have \(|f(z)-f_j(z)|<1\) and hence \(\Re(f(z))\ge j\) for all \(z\in V_j\), \(j\in\mathbb{N}\). Since every divergent path \(\gamma\colon[0,1)\to\mathbb{B}_n\) of finite length meets \(V_j\) for infinitely many \(j\)'s by condition (ii), we obtain that \(\Re(f)\) is unbounded above on every such \(\gamma\), proving Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} in the case of \(D=\mathbb{B}_n\). The proof of the theorem for an arbitrary pseudoconvex domain in \(\mathbb{C}^n\) \((n\ge 2)\) given in relies on a modification of this technique, exploiting the well known fact that every such domain is biholomorphic to a properly embedded complex submanifold of a complex Euclidean space; see. (In fact, if \(n\ge 2\) then every \(n\)-dimensional Stein manifold admits a proper holomorphic embedding into \(\mathbb{C}^N\) for \(N=[\frac{3n}{2}]+1\); see Y. Eliashberg and M. Gromov for even \(n\) and J. Schürmann for odd \(n\ge 3\). The dimension \([\frac{3n}{2}]+1\) in this assertion is the lowest prossible for all \(n\ge 2\) due to purely topological reasons, as O. Forster pointed out by examples in.) Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} also holds for \(n=1\) the proof being much easier. In fact, it is not hard to show, as an application of the classical Runge theorem in one variable, that every domain \(\Omega\subset\mathbb{C}\) admits a holomorphic function whose real part is unbounded both above and below on every divergent path \([0, 1) \to\Omega\) of finite length (recall that every domain in \(\mathbb{C}\) is pseudoconvex). In this direction, P.W. Jones proved back in 1980 that the unit disc \(\mathbb{D}\) admits a bounded holomorphic function \(h\) such that all level sets of \(|h|\colon \mathbb{D}\to\mathbb[0,\infty)\) have infinite length. Obviously, the functions in Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} are not bounded, hence the following remains an open (but likely difficult) question.
## Controlling the topology of the examples {#sec:topology}
Neither the construction in of complete properly embedded complex curves in the ball \(\mathbb{B}_2\) nor that in of complete properly embedded complex hypersurfaces in \(\mathbb{B}_n\) for arbitrary \(n\ge 2\) provide any control on the topology of the examples, which, in principle, could be very complicated. In particular, it remained an open question whether there are complete properly embedded complex curves of finite topology in the ball of \(\mathbb{C}^2\), and, more precisely, whether there are simply connected ones; see and. The first construction of complete properly embedded complex hypersurfaces in the ball with control of the topology was carried out shortly after by A. Alarcón, J. Globevnik, and F.J. López, who proved the following.
In the case of \(n=2\), Theorem [\[th:AGL\]](#th:AGL){reference-type="ref" reference="th:AGL"} trivially implies the following more precise result. Recall that all open orientable surfaces can be realized as Runge domains of properly embedded complex curves in \(\mathbb{C}^2\); see M. Černe and F. Forstnerič for the finite topological case and A. Alarcón and F.J. López for the arbitrary one.
This settled in the positive the aforementioned questions and. The construction method in is different from the ones in. In particular, it introduces a new tool into the game: the use of holomorphic automorphisms of \(\mathbb{C}^n\). In order to prove Theorem [\[th:AGL\]](#th:AGL){reference-type="ref" reference="th:AGL"} one begins with the given properly embedded hypersurface \(Z\) in \(\mathbb{C}^n\) and the natural embedding \(Z\hookrightarrow \mathbb{C}^n\) given by the inclusion map. Then, in a recursive way, one composes this initial embedding with a sequence of holomorphic automorphisms \(\Phi_j\) of \(\mathbb{C}^n\), \(j\in\mathbb{N}\), which converges uniformly on compact subsets of \(\mathbb{B}_n\). One obtains in this way a sequence of proper holomorphic embeddings \(\Psi_j\colon Z\hookrightarrow\mathbb{C}^n\), where \(\Psi_j=\Phi_j\circ\Phi_{j-1}\circ\cdots\circ\Phi_1\), \(j\in\mathbb{N}\), whose images converge uniformly on compact subsets of \(\mathbb{B}_n\) to a properly embedded complex hypersurface of \(\mathbb{B}_n\). If this process is carried out in the right way, then one can guarantee that a connected component \(V\) of the limit hypersurface is biholomorphic to a pseudoconvex Runge domain \(X\) in \(Z\) which is closed in \(\mathbb{B}_n\) and contains \(K\). The main point is that the sequence of holomorphic automorphisms \(\Phi_j\) of \(\mathbb{C}^n\) can be chosen to make sure that \(V\) is disjoint from a labyrinth \(L\) of compact sets in \(\mathbb{B}_n\) as that constructed by J. Globevnik in (see [\[eq:Lab\]](#eq:Lab){reference-type="eqref" reference="eq:Lab"} and Figure [\[fig:L-G\]](#fig:L-G){reference-type="ref" reference="fig:L-G"}), and hence \(V\) is in addition complete. Indeed, denote by \(\{L_j\}_{j\in\mathbb{N}}\) the family of connected components of \(L\) and note that, up to reordering, the following properties are satisfied for all \(j\in\mathbb{N}\):
(a) \(L_j\) is a compact convex set contained in an affine real hyperplane \(H_j\) of \(\mathbb{C}^n\).
(b) There exists a compact convex set \(K_j\subset\mathbb{C}^n\) such that \(\bigcup_{i=1}^{j-1} L_i\subset K_j\) and \(K_j\cap H_l=\varnothing\) for all \(l\ge j\).
Moreover, up to removing finitely many pieces of the labyrinth if necessary, we may also assume that:
(a) There is a compact convex set \(K_0\subset\mathbb{C}^n\) such that \(K\subset K_0\subset \mathbb{C}^n\setminus\bigcup_{l\in\mathbb{N}} H_l\).
By Kallin's lemma (see E. Kallin or E.L. Stout ) and the Oka-Weil theorem (see ), if \(C\) and \(T\) are disjoint compact convex sets in \(\mathbb{C}^n\) and \(A\subset C\) is a compact polynomially convex set, then the union \(A\cup T\) is polynomially convex as well; we refer to for a monograph on polynomial convexity. This easily implies in view of conditions (a), (b), and (c) that:
(a) The set \(K_j\cup L_j\) is polynomially convex for all \(j\in\mathbb{N}\cup\{0\}\). More generally, the compact set \(K_j\cup \bigcup_{i=j}^k L_k\) is polynomially convex for all \(k\ge j\), \(j\in\mathbb{N}\cup\{0\}\).
In this situation, the Andersén-Lempert theory (see E. Andersén and L. Lempert and F. Forstnerič and J.-P. Rosay ) enables one to construct a sequence \(\{\Phi_j\}_{j\in\mathbb{N}}\) of holomorphic automorphisms of \(\mathbb{C}^n\) such that if \[\Psi_j=\Phi_j\circ\cdots\circ\Phi_1,\quad j\in\mathbb{N},\] then \(|\Phi_j(z)-z|<\epsilon_j\) for all \(z\in K_j\) and \[\Psi_j(Z)\cap\bigcup_{i=1}^j L_j=\varnothing\quad \text{for all }j\in\mathbb{N}.\] Here, each \(\epsilon_j\) is a positive number which may be chosen arbitrarily small at each step in the inductive construction. In fact, setting \(\Psi_0=\Phi_0={\rm Id}_{\mathbb{C}^n}\) and assuming that we have such automorphisms \(\Phi_1,\ldots,\Phi_{j-1}\) for some \(j\in\mathbb{N}\), up to a suitable affine complex change of coordinates we can even choose \(\Phi_j\) to be a shear map of the form \[\Phi_j(\zeta,w)=(\zeta,e^{\phi_j(\zeta)}w),\quad \zeta\in\mathbb{C},\; w\in\mathbb{C}^{n-1},\] with a holomorphic function \(\phi_j\colon\mathbb{C}\to\mathbb{C}\). Choosing the number \(\epsilon_j>0\) sufficiently small at each step, we have that \(\lim_{j\to\infty}\Psi_j=\Psi\) exists uniformly on compact subsets in \(\Omega:=\bigcup_{j\in\mathbb{N}}\Psi_j^{-1}(K_j)\), and \(\Psi\colon\Omega\to\mathbb{B}_n\) is a biholomorphic map such that \[\Psi(Z\cap\Omega)\cap L=\varnothing.\] This implies that \(X=Z\cap\Omega\) is a pseudoconvex Runge domain in \(Z\), \(K\subset X\), and the closed complex hypersurface \(V=\Psi(X)\) of \(\mathbb{B}_n\) is complete (since it is disjoint from the labyrinth \(L\) in [\[eq:Lab\]](#eq:Lab){reference-type="eqref" reference="eq:Lab"}; see condition (ii) in the description of \(L\)), thereby proving Theorem [\[th:AGL\]](#th:AGL){reference-type="ref" reference="th:AGL"}. Despite the fact that the method developed in works fine if one uses the labyrinths constructed in, the authors introduced in new labyrinths of compact sets in the ball with properties analogous to those in (i)-(iv) and (a)-(d) above. The main reason for building these new labyrinths is that they are constructed in a much simpler way; only basic trigonometry is required.
These new labyrinths are much more symmetric than Globevnik's ones in, and its pieces are all round balls. See Figure [\[fig:tidy\]](#fig:tidy){reference-type="ref" reference="fig:tidy"}. The main two properties of tangent labyrinths are the following:
(I) If \(\mathcal{L}\) is a tangent labyrinth in the spherical shell \(R\mathbb{B}_n\setminus r\overline{\mathbb{B}}_n\) \((0<r<R,\; n\ge 2)\), then the compact set \(r\overline{\mathbb{B}}_n\cup\mathcal{L}\subset\mathbb{C}^n\) is polynomially convex.
(II) For any given numbers \(0<r<R\), \(\delta>0\) (big), and \(\tau>0\) (small) there exists a tangent labyrinth \(\mathcal{L}\) in \(R\mathbb{B}_n\setminus r\overline{\mathbb{B}}_n\) \((n\ge 2)\) enjoying the following conditions:
- If \(\gamma\colon[0,1]\to\mathbb{B}_n\) is a path crossing the shell (i.e., \(|\gamma(0)|\le r\) and \(|\gamma(1)|\ge R\)) and \(\gamma\) is disjoint from the labyrinth (i.e., \(\gamma([0,1])\cap\mathcal{L}=\varnothing\)), then the Euclidean length of \(\gamma\) is greater than \(\delta\).
- Each component \(T\) of \(\mathcal{L}\) has Euclidean diameter smaller than \(\tau\); i.e. the radius of the Euclidean ball \(T\) is smaller than \(\tau/2\).
Condition (I) can be seen as a consequence of the Kallin lemma and the Oka-Weil theorem, as above; this was observed by A. Alarcón and F. Forstnerič in, see also. On the other hand, condition (II) was granted in; see also. Note that the tangent labyrinths \(\mathcal{L}\) in Definition [\[def:L\]](#def:L){reference-type="ref" reference="def:L"} are compact and have finitely many components, while those \(L\) in [\[eq:Lab\]](#eq:Lab){reference-type="eqref" reference="eq:Lab"} are closed in the ball and have infinitely many components. It is however obvious in view of condition (II) that for any increasing sequence \(0<r_1<R_1<r_2<R_2<\cdots\) with \(\lim_{j\to\infty}r_j=1\) there is a sequence of tangent labyrinths \(\mathcal{L}_j\) in \(R_j\mathbb{B}_n\setminus r_j\overline{\mathbb{B}}_n\), \(j\in\mathbb{N}\), with the property that every divergent path \(\gamma\colon[0,1)\to\mathbb{B}_n\) meeting at most finitely many components of the infinite labyrinth \[\label{eq:infiniteL} \mathcal{L}=\bigcup_{j\in\mathbb{N}}\mathcal{L}_j\subset\mathbb{B}_n\] has infinite length; cf. property (ii) of the Globevnik labyrinth \(L\) in [\[eq:Lab\]](#eq:Lab){reference-type="eqref" reference="eq:Lab"}. The new construction method developed in, based on the use of tangent labyrinths and holomorphic automorphisms of \(\mathbb{C}^n\), has given rise to some further developments. For instance, A. Alarcón and J. Globevnik extended Corollary [\[co:AGL\]](#co:AGL){reference-type="ref" reference="co:AGL"} to complex curves of arbitrary topology, possibly infinite, including in addition a hitting condition.
Choosing the set \(\Lambda\) such that the limit set \(\overline{\Lambda}\setminus\Lambda\) equals the boundary sphere \(b\mathbb{B}_n=\{z\in\mathbb{C}^n\colon |z|=1\}\), one obtains the following immediate corollary.
In the case when \(M\) is simply-connected, and hence the Riemann surface \((M,J)\) is biholomorphic to the disc \(\mathbb{D}\), Theorem [\[th:AG-C2\]](#th:AG-C2){reference-type="ref" reference="th:AG-C2"} was previously obtained by J. Globevnik in. The proof of Theorem [\[th:AG-C2\]](#th:AG-C2){reference-type="ref" reference="th:AG-C2"} combines ideas from and, the latter relying on the method of exposing boundary points of bordered Riemann surfaces in \(\mathbb{C}^2\) as well as on the use of Fatou-Bieberbach domains of \(\mathbb{C}^2\). Theorem [\[th:AG-C2\]](#th:AG-C2){reference-type="ref" reference="th:AG-C2"} is somehow analogous to the main result in where properly embedded complex curves in \(\mathbb{C}^2\) with arbitrary topology are constructed. Also using tangent labyrinths and holomorphic automorphisms of \(\mathbb{C}^n\) as in, A. Alarcón and F. Forstnerič obtained complete complex submanifolds in \(\mathbb{B}_n\) with controlled topology, having no self-intersections, and hitting any given countable subset (not necessarily closed or discrete), thereby complementing Theorem [\[th:AGL\]](#th:AGL){reference-type="ref" reference="th:AGL"}.
In the case of complex curves, this has the following immediate corollary; compare with Theorem [\[th:AG-C2\]](#th:AG-C2){reference-type="ref" reference="th:AG-C2"} and statement 3 in Theorem [\[th:Jordan\]](#th:Jordan){reference-type="ref" reference="th:Jordan"}.
The first result which provided control of the topology of all leaves in a holomorphic foliation of the ball by complete closed complex submanifolds was subsequently given by A. Alarcón and F. Forstnerič.
The topology of the leaves in the foliations given in Theorem [\[th:AF2020\]](#th:AF2020){reference-type="ref" reference="th:AF2020"} is the simplest possible one: discs. It is not known whether a comparable result holds for leaves with prescribed but more complicated topology. The proof follows a similar approach as that in, but with a more careful use of the Andersén-Lempert theory. The idea is, using holomorphic automorphisms, to successively twist a holomorphic foliation of \(\mathbb{C}^n\) by complex lines in order to make bigger and bigger parts of the foliation avoid more and more pieces of an infinite labyrinth in \(\mathbb{B}_n\), as those in [\[eq:Lab\]](#eq:Lab){reference-type="eqref" reference="eq:Lab"} or [\[eq:infiniteL\]](#eq:infiniteL){reference-type="eqref" reference="eq:infiniteL"}. Each component of the intersection of the ball with a properly embedded complex line is Runge in the complex line, and hence simply connected. Since it is in addition bounded, it is a properly embedded complex disc in the ball. More precisely, the construction provides a sequence of foliations \(\{\mathcal{F}_j\}_{j\in\mathbb{N}}\) of \(\mathbb{B}_n\) by properly embedded complex discs, converging to a limit foliation \(\mathcal{F}\) whose all leaves are discs and each one meets at most finitely many components of the labyrinth, so it is complete. None of the methods discussed in this section enables one to control the complex structure of the examples, except of course if the curve is simply connected. In particular, the following remains an open question; see.
F. Forstnerič and E.F. Wold proved in that every bordered Riemann surface \(M\) as in Problem [\[pr:compact\]](#pr:compact){reference-type="ref" reference="pr:compact"} admits a proper holomorphic embedding \(M\hookrightarrow\mathbb{C}^2\). Obviously, such an embedding is complete but unbounded. Here is a less ambitious but still appealing question in this direction.
## Complete complex hypersurfaces in the ball come in foliations
In principle, it was not clear whether the construction of Globevnik in gives, say, a rare type or a very special type of solution to the embedded Yang problem for hypersurfaces in the ball. Note that his examples of complete closed complex hypersurfaces in \(\mathbb{B}_n\) are obtained as leaves of a holomorphic foliation of \(\mathbb{B}_n\) by complex hypersurfaces of the same sort. Thus, the question whether every such hypersurface can be embedded into a foliation of the ball by hypersurfaces with the same properties naturally appeared. Recall that, at that time, there were already available in the literature several constructions of complete properly embedded complex submanifolds in balls, besides the implicit one, so one could be inclined to think that the answer to this question is negative. However, the answer is positive, as it follows from the following more general result by A. Alarcón dealing with submanifolds of arbitrary codimension.
The family of components of the fibres \(f^{-1}(c)\) \((c\in\mathbb{C}^q)\) of the holomorphic submersion \(f\) given by the theorem form a nonsingular holomorphic submersion foliation of \(\mathbb{B}_n\) by smooth connected closed complex submanifolds of codimension \(q\) all which, except perhaps those contained in the initial submanifold \(V\subset f^{-1}(0)\), are complete. If the given \(V\) is complete, then all the leaves in the foliation are complete. This provided the first known example of a foliation of the ball \(\mathbb{B}_n\) by complete connected closed complex submanifolds of codimension \(q>1\), as well as the first example of a nonsingular holomorphic foliation of \(\mathbb{B}_n\) by such submanifolds for any dimension \(n\ge 2\) and codimension \(1\le q<n\). Somewhat surprisingly, the hypersurface \(V\) in Theorem [\[th:JDG\]](#th:JDG){reference-type="ref" reference="th:JDG"} need not be complete, so the following holds.
The following result concerning the existence of foliations of the ball having both many complete leaves and many incomplete leaves was subsequently obtained by a generalization of the construction method in .
The assumption in Theorem [\[th:JDG\]](#th:JDG){reference-type="ref" reference="th:JDG"} that the given hypersurface \(V\) is contained in a fibre of a holomorphic submersion is necessary and cannot be relaxed. In the case of \(q=1\), this condition is always satisfied (see F. Forstnerič ), so the theorem applies to every smooth closed complex hypersurface in \(\mathbb{B}_n\). More precisely, it is known that every closed complex hypersurface \(V\) (possibly with singularities) in \(\mathbb{B}_n\) is defined by a holomorphic function on \(\mathbb{B}_n\) which is noncritical out of the singular set of \(V\). (This actually holds true with the ball replaced by any Stein manifold \(X\) with the vanishing second cohomology group \(H^2(X;Z)=0\) ; which extends the classical result by J.-P. Serre from 1953 that every divisor in such a Stein manifold \(X\) is a principal divisor.) The following more precise version of Theorem [\[th:JDG\]](#th:JDG){reference-type="ref" reference="th:JDG"} in the case of hypersurfaces, possibly with singularities, was also obtained in.
In this case, the family of components of the level sets \(f^{-1}(c)\) \((c\in\mathbb{C})\) of the holomorphic function \(f\) furnished by the theorem form a holomorphic foliation of \(\mathbb{B}_n\) by connected closed complex hypersurfaces all which, except perhaps those contained in the given hypersurface \(V= f^{-1}(0)\), are smooth and complete. If the given \(V\) is smooth, then \(f\) is noncritical and hence the foliation nonsingular; while if \(V\) is complete, then all the leaves in the foliation are complete. This established a sort of converse to Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"}, namely, it proves that, in fact, every complete closed complex hypersurface in \(\mathbb{B}_n\) is defined by a holomorphic function on \(\mathbb{B}_n\) whose level sets are all complete (i.e., as those constructed by Globevnik in ), and hence it can be embedded as a leaf in a holomorphic foliation of the ball by complete closed complex hypersurfaces. Moreover, the defining function can be chosen noncritical, and hence the foliation nonsingular, whenever the given hypersurface is smooth. The method used in the proof of Theorem [\[th:JDG\]](#th:JDG){reference-type="ref" reference="th:JDG"} broadly follows the one developed by Globevnik for proving Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} in, but it presents some major differences and novelties. The holomorphic submersion \(f\colon\mathbb{B}_n\to\mathbb{C}^q\) satisfying the conclusion of the theorem is obtained as the limit of a sequence of holomorphic submersions \(f_j\colon\mathbb{B}_n\to\mathbb{C}^q\) \((j\in\mathbb{N})\), and it is granted that each fibre of \(f\), except perhaps the one over \(0\in\mathbb{C}^q\), meets at most finitely many components of a suitable infinite labyrinth of compact sets in \(\mathbb{B}_n\), thereby making sure that every such fibre is complete. A main novelty is that the proof requires to construct the labyrinth and the limit holomorphic submersion \(f\) at the same time in an inductive procedure; that is, the labyrinth cannot be fixed beforehand, as in the proofs in and in the aforementioned subsequent papers. Moreover, the labyrinth depends on the given submanifold \(V\), and the components of the labyrinth that meet \(V\) are treated in a completely different way than those that do not; in particular, the former components must be chosen with small diameter. Another novelty of the proof is that it relies on a rather sophisticated Oka-Weil-Cartan type approximation with interpolation theorem for holomorphic submersions due to F. Forstnerič (see or ), instead of the more basic Runge-Weierstrass theorem. We next spell out some of the details. Let us briefly outline the proof of Theorem [\[th:JDG\]](#th:JDG){reference-type="ref" reference="th:JDG"} in the slightly simpler case when there is a holomorphic submersion \(f_0\colon\mathbb{B}_n\to\mathbb{C}^q\) such that \(V=f_0^{-1}(0)\); that is always the case if \(q=1\). Choose any increasing sequence of positive numbers \(0<r_1<R_1<r_2<R_2<\cdots\to 1\), fix \(\epsilon_0>0\), choose \(0<R_0<r_1\), and call \(L_0=\varnothing\). The proof consists of inductively constructing a sequence \(S_j=\{f_j,\epsilon_j,\mathcal{L}_j\}_{j\in\mathbb{N}}\), where
- \(f_j\colon\mathbb{B}_n\to\mathbb{C}^q\) is a holomorphic submersion,
- \(\epsilon_j>0\) is a positive number, and
- \(\mathcal{L}_j\) is a tangent labyrinth in the spherical shell \(R_j\mathbb{B}_n\setminus r_j\overline{\mathbb{B}}_n\) (see Definition [\[def:L\]](#def:L){reference-type="ref" reference="def:L"}),
such that the following conditions are satisfied for each \(j\in\mathbb{N}\):
1. \(|f_j(z)-f_{j-1}(z)|<\epsilon_j\) for all \(z\in r_j\overline{\mathbb{B}}_n\).
2. \(f_j^{-1}(0)=V\).
3. \(0<2\epsilon_j<\epsilon_{j-1}\) and if \(f\colon\mathbb{B}_n\to\mathbb{C}^q\) is a holomorphic map such that \(|f(z)-f_{j-1}(z)|<2\epsilon_j\) for all \(z\in r_j\overline{\mathbb{B}}_n\), then \(f\) is submersive everywhere on \(R_{j-1}\overline{\mathbb{B}}_n\).
4. If \(\gamma\colon[0,1]\to\mathbb{B}_n\) is a path satisfying that \(|\gamma(0)|\le r_j\), \(|\gamma(1)|\ge R_j\), and \(\gamma([0,1])\cap\mathcal{L}_j=\varnothing\), then \({\rm length}(\gamma)>1\).
5. If \(z\in\mathcal{L}_j\), then either \(|f_j(z)|>j\) or \(|f_j(z)|<1/j\).
Carrying out this process in the right way, one obtains a holomorphic submersion \[f=\lim_{j\to\infty}f_j\colon\mathbb{B}_n\to\mathbb{C}^q,\quad j\in\mathbb{N},\] as well as an infinite labyrinth of compact sets \[\mathcal{L}=\bigcup_{j\in\mathbb{N}}\mathcal{L}_j\subset\mathbb{B}_n,\] satisfying the following conditions:
(A) \(f^{-1}(0)=V\).
(B) Every divergent path \([0,1)\to\mathbb{B}_n\) hitting at most finitely many components of \(\mathcal{L}\) has infinite length.
(C) \(f^{-1}(c)\) meets at most finitely many components of \(\mathcal{L}\) for each \(c\in\mathbb{C}^q\setminus\{0\}\).
Note that conditions (B) and (C) are granted by properties (4\(_j\)) and (5\(_j\)), respectively. It is clear that (B) and (C) ensure that \(f^{-1}(c)\) is a smooth complete closed complex submanifold of pure codimension \(q\) in \(\mathbb{B}_n\) for all \(c\in\mathbb{C}^q\setminus\{0\}\), and hence, in view of condition (A), \(f\) satisfies the conclusion of the theorem. For the inductive step, fix \(j\in\mathbb{N}\) and assume that we already have a suitable triple \(S_{j-1}=\{f_{j-1},\epsilon_{j-1},\mathcal{L}_{j-1}\}\). First of all, choose a number \(\epsilon_j>0\) so small that (3\(_j\)) holds; that is always possible by the Cauchy estimates. Also choose, as we may by (2\(_{j-1}\)), a number \(\tau>0\) so small that \[\label{eq:tau} |f_{j-1}(z)|<1/j\quad \text{for all \(z\in R_j\overline{\mathbb{B}}_n\) with \({\rm dist}(z,V)<\tau\)}.\] Next, take a tangent labyrinth \(\mathcal{L}_j\) in \(R_j\mathbb{B}_n\setminus r_j\overline{\mathbb{B}}_n\) satisfying (4\(_j\)) and such that \[\label{eq:tau-1} {\rm diam}(T)<\tau\quad \text{for each component \(T\) of \(\mathcal{L}_j\)}.\] (See property (II) below Definition [\[def:L\]](#def:L){reference-type="ref" reference="def:L"}.) We then split the labyrinth \(\mathcal{L}_j\) into two disjoint parts, \(\mathcal{L}_j=\Lambda_V\cup\Lambda_0\), where \(\Lambda_V\) is the union of those components of \(\mathcal{L}_j\) which intersect \(V\) and \(\Lambda_0\) is the union of those which are disjoint from \(V\). Both \(\Lambda_V\) and \(\Lambda_0\) consist of finitely many components and \(\Lambda_V\cap\Lambda_0=\varnothing\). So, we have in view of [\[eq:tau\]](#eq:tau){reference-type="eqref" reference="eq:tau"} and [\[eq:tau-1\]](#eq:tau-1){reference-type="eqref" reference="eq:tau-1"} that \[\label{eq:LambdaV} |f_{j-1}(z)|<1/j\quad \text{for all \(z\in \Lambda_V\)}.\] In this situation, a standard application of the aforementioned Oka-Weil-Cartan-Forstnerič theorem for holomorphic submersions (see for the precise statement that we need here), taking into account that \(r_j\overline{\mathbb{B}}_n\cup \mathcal{L}_j\) is a polynomially convex compact set (see property (I) below Definition [\[def:L\]](#def:L){reference-type="ref" reference="def:L"}), furnishes us with a holomorphic submersion \(f_j\colon \mathbb{B}_n\to\mathbb{C}^q\) satisfying the following conditions:
- \(|f_j(z)-f_{j-1}(z)|<\delta\) for all \(z\in r_j\overline{\mathbb{B}}_n\cup\Lambda_V\) for any given \(\delta>0\).
- \(f_j^{-1}(0)=V\).
- \(|f_j(z)|>j\) for all \(z\in \Lambda_0\).
It is clear that these properties, together with [\[eq:LambdaV\]](#eq:LambdaV){reference-type="eqref" reference="eq:LambdaV"} and the fact that \(\mathcal{L}_j=\Lambda_V\cup\Lambda_0\), guarantee that \(f_j\) satisfies conditions (1\(_j\)), (2\(_j\)), and (5\(_j\)) provided that the number \(\delta>0\) is chosen sufficiently small. This closes the induction and completes the sketch of the proof of Theorem [\[th:JDG\]](#th:JDG){reference-type="ref" reference="th:JDG"}. Complete complex hypersurfaces in a Stein manifold equipped with a Riemannian metric come in foliations as well, as the following result points out.
Completeness in this statement refers to the metric induced in the hypersurface by that in \(X\). Note that the assumption on the normal bundle \(N_{V/X}\) is necessary by the F. Docquier and H. Grauert tubular neighborhood theorem. The connected components of the level sets \(f^{-1}(c)\subset X\) \((c\in\mathbb{C})\) of the holomorphic function \(f\) provided by Theorem [\[th:JDG-Stein\]](#th:JDG-Stein){reference-type="ref" reference="th:JDG-Stein"} form a (possibly singular) holomorphic foliation of the Stein manifold \(X\) by connected closed complex hypersurfaces all which, except perhaps those contained in \(V\), are complete. If \(V\) is complete, then all the leaves are complete. Moreover, most level sets of \(f\) are smooth by Sard's theorem, and hence we obtain the following extension of Corollary [\[co:PikoMA\]](#co:PikoMA){reference-type="ref" reference="co:PikoMA"}. (Recall that pseudoconvex domains are Stein manifolds, but not every Stein manifold is biholomorphic to a pseudoconvex domain in a complex Euclidean space.)
The proof of Theorem [\[th:JDG-Stein\]](#th:JDG-Stein){reference-type="ref" reference="th:JDG-Stein"} combines the construction method in with a slight refinement of the trick used by Globevnik in to show that every pseudoconvex domain (other than the ball) in \(\mathbb{C}^n\) admits a holomorphic foliation by complete closed complex hypersurfaces (see Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"}). Although every Stein manifold admits a holomorphic function without critical points, as was shown by F. Forstnerič in, the method of proof in does not allow to guarantee that the function \(f\) furnished by Theorem [\[th:JDG-Stein\]](#th:JDG-Stein){reference-type="ref" reference="th:JDG-Stein"} be noncritical. Therefore, the following remains an open problem (see ).
## Labyrinths in pseudoconvex domains
Recently, S. Charpentier and Ł. Kosiński constructed labyrinths of compact sets in any given pseudoconvex Runge domain \(D\subset\mathbb{C}^n\) with properties analogous to those of the tangent labyrinths in balls (see Definition [\[def:L\]](#def:L){reference-type="ref" reference="def:L"} and properties (I) and (II) below it).
Since every pseudoconvex domain \(D\subset\mathbb{C}^n\) admits a normal exhaustion \(\Omega_1\Subset \Omega_2\Subset\cdots\subset\bigcup_{j\in\mathbb{N}}\Omega_j=D\) by smooth \(\mathcal{O}(D)\)-convex strongly pseudoconvex domains, which can be taken to be Runge if \(D\) is Runge (see ), an inductive application of Theorem [\[th:CK\]](#th:CK){reference-type="ref" reference="th:CK"} provides holomorphically contractible compact sets \(\Gamma_j\subset \Omega_{j+1}\setminus \overline{\Omega}_j\), \(j\in\mathbb{N}\), satisfying the following conditions (see ):
- \(\Gamma_j\) has finitely many components and \(\overline{\Omega}_j\cup \Gamma_j\) is \(\mathcal{O}(D)\)-convex for all \(j\in\mathbb{B}_n\).
- If \(\gamma\colon[0,1)\to D\) is a divergent path meeting at most finitely many components of \(\Gamma=\bigcup_{j\in\mathbb{N}}\Gamma_j\), then \(\gamma\) has infinite length.
- If \(\Omega\) is a Runge domain, then the components of \(\Gamma\) can be chosen to be images of Euclidean balls of real dimension \(2n-1\) under \(\mathbb{R}\)-affine isomorphisms and automorphisms of \(\mathbb{C}^n\).
This can be used to extend from the ball \(\mathbb{B}_n\) to any pseudoconvex Runge domain \(D\) in \(\mathbb{C}^n\) many of the results mentioned in this survey, just by using the labyrinths in Theorem [\[th:CK\]](#th:CK){reference-type="ref" reference="th:CK"} instead of the tangent labyrinths in Definition [\[def:L\]](#def:L){reference-type="ref" reference="def:L"}. In particular, all results in Section [4.1](#sec:topology){reference-type="ref" reference="sec:topology"}, as well as Theorem [\[th:JDG\]](#th:JDG){reference-type="ref" reference="th:JDG"}, Theorem [\[th:JDG-H\]](#th:JDG-H){reference-type="ref" reference="th:JDG-H"}, and the first assertion in Corollary [\[co:JDG\]](#co:JDG){reference-type="ref" reference="co:JDG"}, hold true with the ball replaced by any pseudoconvex Runge domain (for the extension of Theorem [\[th:AF2020\]](#th:AF2020){reference-type="ref" reference="th:AF2020"} it is also required that the domain \(D\) be hyperbolic, meaning that it does not contain any biholomorphic copy of \(\mathbb{C}\)). Note, in particular, that this provides an affirmative answer to Problem [\[pr:Stein\]](#pr:Stein){reference-type="ref" reference="pr:Stein"} in the special case when the Stein manifold \(X\) is a pseudoconvex Runge domain in \(\mathbb{C}^n\). Likewise, Corollary [\[co:Indiana\]](#co:Indiana){reference-type="ref" reference="co:Indiana"} and assertions 2 and 3 in Corollary [\[co:JDG\]](#co:JDG){reference-type="ref" reference="co:JDG"} hold true with the ball replaced by any such domain \(D\) if we assume in addition that \(D\) contains no smooth closed complex submanifold of \(\mathbb{C}^n\); this happens for instance if \(D\subset\mathbb{C}^n\) is relatively compact. In the subsequent paper, S. Charpentier and Ł. Kosiński extended Globevnik's existence result for highly oscillating holomorphic functions in Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"}, to the existence of functions with an even wilder asymptotic behavior.
Note that the functions \(f\) provided by Theorem [\[th:CK-wild\]](#th:CK-wild){reference-type="ref" reference="th:CK-wild"} also satisfy the conclusion of Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"}. In particular, the components of the level sets \(f^{-1}(c)\) \((c\in\mathbb{C})\) form a (possibly singular) holomorphic foliation of \(D\) by complete connected closed complex hypersurfaces (possibly with singularities), as those in Corollary [\[co:PikoMA\]](#co:PikoMA){reference-type="ref" reference="co:PikoMA"}. We emphasize that the functions \(f\) in Theorem [\[th:CK-wild\]](#th:CK-wild){reference-type="ref" reference="th:CK-wild"} are abundant; in particular, they form a dense subset of the space of all holomorphic functions on \(D\). Therefore, the functions in Theorem [\[th:PikoMA\]](#th:PikoMA){reference-type="ref" reference="th:PikoMA"} are abundant as well. | {'timestamp': '2022-12-19T02:13:28', 'yymm': '2212', 'arxiv_id': '2212.08521', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08521'} |
# Introduction
Astronomical spectral clustering has attracted increasing attention in astronomy. It helps us to analyse the birth, formation and evolution of the cosmic and astronomical objects. Nowadays, rapid developing sky survey projects have obtained a large amount of astronomical data, which poses great challenges for effective astronomical clustering. In view of the above challenges, a wide range of clustering methods (partition-based clustering methods, density-based clustering methods and so on) have been applied to diverse astronomical tasks. However, clustering methods perform differently on various data sources, making it difficult to evaluate the performance and determine their applicable scenarios. To explore the advantages of each type of clustering method, we investigate a large number of clustering methods applied to astronomical spectra data and analyse their applications, core ideas, merits and caveats. We then compare the performance of classical clustering methods on unified datasets and give objective appraisals. The datasets are constructed considering four aspects, as follows. (i) Different data types of some spectral feature extraction methods. Four types of data including 1D spectra, PCA (Principal Component Analysis, a widely used dimensionality reduction method) features of spectra, line indices and stellar parameters are used to test different methods.\
(ii) The quality of spectra. Three signal-to-noise ratios (S/Ns) are used to analyse the robustness of clustering algorithms.\
(iii) Data volume. Four sizes of spectral datasets are constructed including 8000, 20000, 40000 and 80000, however some algorithms cannot run on the size of 80000.\
(iv) Outliers detection. We know that some clustering methods can be used to detect outliers, so datasets containing normal spectra and rare objects are constructed to test the ability of some algorithms to detect outliers. After each experiment, an objective analysis is given. Moreover, experimental source codes in this paper and a brief manual about usage of the source codes are available to readers. This work is organized as follows: In Section 2, we briefly describe the clustering methods from applications and theory analysis. In Section 3, experiments on four astronomical tasks-A/F/G/K stars classification, star/galaxy/quasar classification, subclasses of A-type star classification and outliers detection-are carried out. Section 4 puts forward source codes of the aforementioned experiments and a manual about the usage of the source codes. Finally, a discussion and our future work are presented in Section 5.
# Investigation of Clustering Methods on Astronomical Spectra Data
As a prevalent task, clustering has been applied in many fields, such as imaging processing , social networks, finance security, biological fields , and others . It attempts to divide data into different groups according to certain criteria. Data in the same group appears to be more similar than data in different groups. The classical clustering process can be divided into the following steps.
(1) Data preprocessing: extract and choose the most representative features;
(2) Clustering: design suitable clustering algorithms and train models with unlabelled data to cluster the real data;
(3) Result analysis: explain the clustering result and evaluate the clustering method.
Clustering methods play a vital role in astronomical spectral data analysis . Traditional clustering algorithms can be divided into partition-based, density-based, hierarchical, grid-based, and model-based. In recent years, lots of new clustering algorithms have appeared, like fuzzy theory-based clustering algorithm, kernel-based clustering algorithm, etc. The next subsections will introduce the above typical clustering algorithms, considering applicable scenarios, core ideas, merits, and caveats.
## Partition-Based Clustering Algorithm
Partition-based clustering methods classify the data into different clusters by finding optimal cluster centers and K-means is a widely used partition-based algorithm because of its simplicity and efficiency. Astronomical investigations based on partition-based clustering algorithm are shown in Table [\[table_partitioned_related\]](#table_partitioned_related){reference-type="ref" reference="table_partitioned_related"}. used K-means to classify galaxies from SDSS DR7. However, some spectra appeared to be between classes, so they proposed a K-means-based method to identify marginal galaxies. analysed the application of K-means to massive Apache Point Observatory Galactic Evolution Experiment (APOGEE) high resolution spectra. The results showed that K-means was able to separate the bulge and halo populations and distinguish dwarfs, subgiants, red clump (RC) stars, and red giant branch (RGB) stars. Clustering of star, galaxy, quasar and subclasses have been carried out extensively to analyse physical properties, chemical abundances, boundary spectra, etc ( @2018RAA....18...73C [@7984705; @Hogg_2016; @Kheirdastan_2016; @2011ApJ...743...77M], @2013ApJ...763...50S). Furthermore, clustering results contain valuable information, such as structure of Galaxy and evolutionary stages of the Universe, which provides convenience for astronomers . Initialization of clusters centers and the number of clusters are two factors that affect K-means results greatly, and many improved methods have been proposed. In order to optimize clustering,, and used new approaches to initialize cluster centers and results suggested that clustering with better initial centers are of high quality. optimized K-means by influence space and applied ISBFK-means on astronomical spectral data of LAMOST. Results presented a good performance on stability and speed. Another common application is special objects identification (e.g. rare objects, light curves, star clusters) and retrieval tasks , and the detected special objects are a significant supplement to the astronomy category. used K-means clustering to successfully divide the confirmed exoplanets from NASA dataset into different clusters. In astronomical applications, effective feature selection not only affects clustering accuracy and time, but is also closely related to celestial properties. and adopt line indices as spectral features to improve data quality and results indicated that they were consistent with internal properties of stars. Optimal feature extraction (color space, chemical pattern, etc.) can provide more information for clustering without additional prior knowledge . Among partition-based clustering algorithms, K-means and K-mediods are two typical methods; variants such as K-means++, intelligent K-means, genetic K-means, K-modes, and kernel K-means also perform well. The goal of partition-based clustering algorithms is that data in the same group appears more similar than those in different groups. The pseudo-code about partition-based clustering methods is given in Algorithm [\[Algorithm_1\]](#Algorithm_1){reference-type="ref" reference="Algorithm_1"} and Fig. [\[fig:kmeans_structure\]](#fig:kmeans_structure){reference-type="ref" reference="fig:kmeans_structure"} shows the structure of K-means.
At present, density-based clustering algorithms are promising for exploring the universe. Table [\[table_DBSCAN_related\]](#table_DBSCAN_related){reference-type="ref" reference="table_DBSCAN_related"} depicts applications of density-based methods. Density-based clustering methods on spectra data are used to classify celestial objects and they are more frequently used in chemical abundance, photometric data, etc. classified star/galaxy/quasar from photometric data with hierarchical DBSCAN and the accuracy was high. used DBSCAN and DENCLUE to classify galaxies in the Galaxy Zoo dataset. combined DBSCAN and nearest neighbor algorithm to analyse quasars; the results showed that gravitational interaction may cause more low redshifts quasars clustering. Another common application of density-based methods is the detection and analysis of spatial structure of objects, like galaxy clusters, molecular clouds and others. used OPTICS to identify the structure of galaxy/halo and the low-density data points can be identified effectively. HDBSCAN was applied to systematically find members which are up to 50 pc from the cluster centers, and GMM was used to fit the two dimensional distribution of cluster members in this work; these two methods helped to find 389 open clusters. determined membership of open cluster NGC 188 and found that DBSCAN could efficiently select possible member stars in 3D kinematic space without making any assumptions about the distribution of cluster or field stars. used DBSCAN to identify the statistical overdensities of stellar objects, which help to find more open clusters. The structure, kinematics, and ages of the young stellar populations in the Orion region can also be analysed by DBSCAN. Besides this, many researchers committed to multi-space, which is a new perspective of the universe, and software engineering also provides many useful tools for clustering analysis. Some density-based clustering algorithms can search for outliers and cluster centers, such as Clustering by fast search and find of density peaks (CFSFDP). used CFSFDP to detect special samples in the massive low S/N spectra. In astronomical tasks, 'data level' and 'algorithm level' are two major approaches to achieve desirable clustering performance. and found optimal parameters and distance measurement methods to obtain satisfactory results from 'algorithm level'. Astronomical spectra can be represented in many forms, such as chemical abundance, line indices, color space and other features. Efficient feature extraction is a particularly interesting way to improve performance from 'data level'. Many new density-based clustering algorithms have also been proposed to improve the efficiency and robustness of clustering. Density-based clustering algorithms group data according to region density. High density areas are defined as classes sharing similar information. Always, they require two input parameters: the density radius (Eps) and the minimum number of data (MinPts) to build a cluster. DBSCAN, DENCLUE, OPTICS, CFSPDP are classic density-based methods. In addition, some optimization approaches have been explored to improve clustering performance, such as manhattan distance density algorithm (MD-DBSCAN) , normalised euclidean distance (NED-DBSCAN) and so on. The procedure of density-based clustering algorithm is presented in Algorithm [\[Algorithm_2\]](#Algorithm_2){reference-type="ref" reference="Algorithm_2"}.
Investigations of model-based clustering algorithm on astronomical spectra are shown in Table [\[table_model_related\]](#table_model_related){reference-type="ref" reference="table_model_related"}. Model-based methods mainly contain neural network clustering algorithms and probability-based model clustering algorithms. Both of them have already been developed for star, galaxy and peculiar objects clustering. Self-organizing map (SOM) is a special type of neural network which maps the data into a two-dimensional grid. It is always used to visualize the distribution of data, and is similar to the dimensionality reduction visualization methods like t-SNE and UMAP. used SOM to map 158 spectra into a 13 \(\times\) 13 grid and made a classification. The two-dimensional map contains aggregated information of the data and the gaps in the map can be regarded as the separation of different clusters. used SOM to distinguish different objects and divided the grid with Fuzzy C-Means (FCM) which is a fuzzy clustering algorithms based on K-means. trained SOM using a set of galactic templates which covering the wavelength range from far ultraviolet to near-infrared. Compared with other methods, the spectra of galaxies grouped together by the SOM are more similar. And the order of sample categories on the map corresponded to the order of physical properties. That is, in addition to representing the type, the order of samples on the map can also represent many other information, like physical parameters distribution. For clustering, researchers divided the 2D map into many clusters or mapped the data into 1 \(\times\) n where n is the number of categories. For easier use of SOM, some researchers built interactive programs to analyse objects efficiently. Many other applications like outlier analysis, exploration of the spectroscopic diversity of Type Ia supernovae can also be handled or assisted by SOM. Gaussian mixture model (GMM) is one of the most commonly used probability models. It uses multiple multidimensional Gaussian distributions to fit data, and each distribution represents a cluster. In astronomy, GMM can be used in two ways, one is to cluster data based on features, such as 1D spectra, chemical element abundance, etc. The other is to use Gaussian distributions to fit targets directly, such as star clusters and molecular clouds. reduced X-ray spectra which are obtained from Chandra X-ray Observatory from Tycho's supernova remnant using a variational autoencoder (VAE) and clustered the low-dimension features with GMM to analyse spatial structures in Tycho's supernova remnant. used an infinite GMM to detect variability in large amounts of astronomical time series data. GMM was also used to classify different kinds of gamma-ray bursts. used PCA+GMM to identify tidal tail around the old open cluster NGC 2506 based on Gaia-DR2. GMM can also detect and analyze objects which shaped like Gaussian distribution. used a mixture model to probe gas motions in the intra-cluster medium and showed that the mixture parameters can be accurately constrained by Astro-H spectra. proposed a probabilistic HzQ selection method which used GMM to obtain likelihoods and used a Bayesian framework for poster probabilities of HzQ and contaminating sources. It is useful to find more complete HzQ samples. SOM and GMM are the most widely used model-based algorithms. Compared with K-means, they can obtain more valuable information. The idea of SOM is to establish a mapping between high and low dimensions. An illustration of SOM structure is given in Fig. [\[fig:som_structure\]](#fig:som_structure){reference-type="ref" reference="fig:som_structure"}. Algorithm [\[Algorithm_3\]](#Algorithm_3){reference-type="ref" reference="Algorithm_3"} is the main idea of SOM.
We evaluate their performances by ROC curve (receiver operating characteristic curve, or sensitivity curve), which is a widely used index to measure the ability of outliers detection. The horizontal axis of the curve is false positive rate, it is the ratio of normal samples in detected targets to all normal samples. Vertical axis is true positive rate which is the ratio of true outliers in detected targets to all true outliers. The closer the curve is to the upper left corner of the graph, indicates the method has a stronger ability to detect outliers. From the ROC curves of results (Fig. [\[fig:ROC_star\]](#fig:ROC_star){reference-type="ref" reference="fig:ROC_star"} and Fig. [\[fig:ROC_allgalaxy\]](#fig:ROC_allgalaxy){reference-type="ref" reference="fig:ROC_allgalaxy"}), we can know that compared with the classic outlier algorithm LOF, clustering algorithms also perform well in outlier detection tasks. It is easier to detect outliers in stellar spectra than in all spectra (including stars, galaxies, and quasars). The reason is that spectra of quasars are also unusual and very few in number, so they are often detected as outliers too. The performance of algorithms on detecting spectra with bad-merge in red and blue segment is better than carbon star and double star only except GMM. Some badly merged spectra have a wavelength range without flux values, and the Euclidean distance between them and normal spectra is far, so they can be detected easily. GMM uses several Gaussian distributions to fit data and does not measure the similarity between spectra, so its ability to detect badly merged spectra is poor. Fig. [\[fig:ROC_diff_feature\]](#fig:ROC_diff_feature){reference-type="ref" reference="fig:ROC_diff_feature"} reveals desirable results of five algorithms on finding outliers on different spectra features. Algorithms perform worse on 1D spectra than on PCA features, because normal 1D spectra sometimes contain few incomplete spectra which are usually regarded as outliers. In Fig. [\[fig:ROC_diff_feature\]](#fig:ROC_diff_feature){reference-type="ref" reference="fig:ROC_diff_feature"}, clustering algorithms can get more satisfactory consequences than LOF algorithm. It indicates that clustering methods also have good ability to outliers detection. Another interesting thing is that though DBSCAN has difficulty on clustering spectra, it is able to detect outliers. K-means and GMM select spectra far away from cluster centers as outliers and they both have good performances. In our experiments, we find that it is not a good idea to regard the samples in clusters with small number in the clustering results as outliers. By clustering methods, clusters or cluster centers are found, and samples farthest from the cluster centers are more likely to be outliers. Fig. [\[fig:ROC_diff_snr\]](#fig:ROC_diff_snr){reference-type="ref" reference="fig:ROC_diff_snr"} reveals the effect of S/N on outlier detection, and the quality of spectra has a great impact on the results. In the high S/N spectra, each algorithm can find outliers well. But in the low S/N spectra, algorithms basically have no ability to find outliers. K-means and DBSCAN are inferior to others on S/N \> 30. While K-means is superior among all algorithms on S/N \< 10 and S/N:10-30 that shows a little robustness. In the experiment of data volumes (Fig. [\[fig:ROC_diff_volume\]](#fig:ROC_diff_volume){reference-type="ref" reference="fig:ROC_diff_volume"}), the proportion of normal spectra and outliers is 1000:5. When the size of dataset is large, clustering algorithms are better than LOF, since there are more abnormal spectra in large dataset, and LOF will treat them as outliers. Rare objects make significant sense to research in astronomy. To search some kinds of rare objects, we can use supervised classification methods. However, it is difficult to obtain a robust model due to the small number of rare samples. In this situation, clustering methods are good choices to deal with this task. And SOM is also a commonly used method, which can map similar spectra to the same or close locations on two-dimensional competing layer with topology structure. Candidates of rare objects can be adjacent spectra of known rare spectra in the map. In Fig. [\[fig:SOM_spectra\]](#fig:SOM_spectra){reference-type="ref" reference="fig:SOM_spectra"}, 400 spectra including A, F, G and K stars are mapped on a complete layer of 20 \(\times\) 20. Deeper background means more spectra are mapped on this position. Spectra with similar shapes are closer in the map and spectra separated by blank are more different than adjacent spectra. In this graph, A stars, F stars, G stars and K stars are located in different positions respectively. And there are gaps between anomalous spectra and normal spectra, so we also can use SOM to detect outliers. Fig. [\[fig:SOM_rare\]](#fig:SOM_rare){reference-type="ref" reference="fig:SOM_rare"} shows the idea of searching carbon stars by SOM. In the figure, grey points represent normal objects while red points are known carbon stars and purple points are unknown carbon stars. We can recognize adjacent data of red points as rare object candidates and then do further validation manually.
# Source Code and Manual
Source codes of clustering algorithms used in this paper is provided (<https://github.com/shichenhui/SpectraClustering>) and Table [\[table:code\]](#table:code){reference-type="ref" reference="table:code"} presents main codes list. The specific steps and precautions of code usage are also given in the above link. Clustering results of some algorithms will be influenced by the parameters so we adjust parameters and find optimal parameters. The code is written in Python which is widely used for machine learning and data analysis. The dependent packages of the codes are: numpy, sklearn, matplotlib, pandas, scipy. Each algorithm is organized in the following steps: 1) load training datasets; 2) plot the t-SNE distribution of the data; 3) cluster data; 4) plot t-SNE distribution of results; 5) evaluate clustering results. These codes load data from \*.csv files which store tabular data in text and a row of data is a spectrum. Users need to convert their spectra data to such a format or modify the data loading mode. Some of algorithms are implemented directly through the sklearn package. It is efficient to adopt best algorithms for astronomical spectral analysis. The number of clusters K is the main parameter of partition-based methods and it has a great influence on the clustering results. Sometimes we can set the number of clusters K to be more than the actual number, because some rare objects or outliers will be classified as clusters. In our experiment, one or two more than actual number of clusters can be good results. If we have no idea of the number of clusters, t-SNE or UMAP method could be used to have a rough idea of how the data is aggregated. Euclidean distance is used in K-means to measure the similarity between samples and it can be replaced by other similarity measures, like Manhattan distance, distance in manifold space and others. This should be selected according to the characteristics of the data. K-mediod chooses a sample as cluster center which has the smallest sum of distances from all other samples in the cluster instead of an average point. This can reduce the impact of a small number of outliers on the cluster center. But the time complexity is higher than K-means. As for the selection of initial cluster centers, CFSFDP method is a good choice. Neighborhood radius and the number of min-samples in a cluster are two main parameters in DBSCAN. Generally, it requires multiple tests to obtain optimal values. To obtain good clustering results, DBSCAN has two requirements for data, one is the degree of sample aggregation of different clusters cannot be very different, otherwise suitable parameter neighborhood radius will not be found. The other is that there needs to be obvious separation between clusters, so that the clusters are not grouped together. DBSCAN performs bad on spectra because these two conditions are not satisfied. However, in the tasks of analysing spatial structure, DBSCAN can get good results. SOM needs to set the number of iterations, usually hundreds of times to achieve convergence results. It can map data into 1D or 2D space. 1D space is convenient for analyzing the physical properties and the input data must be a small number of relevant features. But for clustering, the results in 1D space are less stable. 2D space can effectively get the data distribution for clustering or outliers detection. The shape of grids mapped with SOM could be rectangular or hexagonal, and hexagonal grids can represent more detailed distribution information. An important parameter in GMM is covariance-type which describes the type of covariance, four types could be chosen, \"spherical\", \"tied\", \"diag\" and \"full\". \"full\" is the most flexible type meaning that each cluster can have its own arbitrary shape. But it may not give the best results in real applications. The choice of covariance-type requires experience or testing to obtain optimal results. Our experimental results found that the \"tied\" and \"diag\" were better than \"spherical\" and \"full\". Agglomerative and divisive are two models of hierarchical clustering. In our experiment, hierarchical clustering could not work well but it can be combined with other algorithms. CFSFDP calculates two values (\(\delta\) and \(\rho\)) for each sample to select cluster centers and outliers. And we can use them to cluster or detect outliers with decision graph method.
# Discussion {#sect:discussion}
Clustering methods used in recent astronomical spectroscopic studies are investigated in this paper and they are briefly introduced. Considering the different data used in respective research, it is difficult to find the strength of each method. So we construct some unified spectral datasets to analyse their advantages under different conditions. Some clustering methods can also be used to find outliers, so datasets also include normal spectra and rare spectra to test their performance on searching for outliers. Through the experiments, we found that GMM performs better than others on 1D spectra and PCA features. On stellar spectra line indices, GMM performs as well as partition-based methods. Spectra line indices can extract stellar spectra features effectively and the clustering results of many methods on line indices are better than 1D spectra. Density-based algorithms and hierarchical clustering perform poorly on spectra related datasets, although they have many advantages on benchmark datasets. The reason is that, in spectra dataset, there is no clear separation between different types of spectra and the density distribution of clusters may be different. So it is impossible to find appropriate parameters for efficient clustering. Although for supervised classification algorithms, overfitting can be reduced when the dataset is large. But for clustering, the amount of data has little influence on the clustering results. And some algorithms can not run on the large datasets. GMM works well on spectra, but its running time on large datasets is much higher than other methods. K-means is still a good choice if researchers want to make a fast clustering of the data. The experiments also show that clustering methods are very effective to find abnormal spectra. Multiple cluster centers need to be found first, then the samples far from the cluster centers can be regarded as outliers and these methods are very robust. When researchers want to observe the distribution of spectra data, dimensionality reduction and visualization methods are very intuitive, such as t-SNE, UMAP and SOM. And SOM is also widely used in astronomy to find special spectra. The purpose of this paper is to provide an analysis of the advantages and disadvantages of different algorithmic ideas. The methods we use are all basic algorithms and the data preprocessing is also general steps. Improved algorithms and better preprocessing will be more effective to improve clustering performance, such as extinction processing, feature extraction, etc. | {'timestamp': '2022-12-19T02:10:38', 'yymm': '2212', 'arxiv_id': '2212.08419', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08419'} |
null | null |
null | null |
# Introduction
Polar aprotic liquids are widely used as solvents on both a laboratory and industrial scale, with important applications across a wide range of chemistry, biochemistry and nanoscience. In this context, their relevant physicochemical properties include high dipole moments, high relative permittivities, and high boiling points, along with broad electrochemical stability windows when compared to their protic analogues. This combination of attributes makes these liquids highly effective for solvation of a wide spectrum of ions, small molecules, polymers, and nanostructures. For example, in electrochemistry their inertness and ability to solvate both metal ions and polymeric co-electrolytes under highly reducing conditions is critical for battery function and stability. In addition, polar aprotic liquids provide a unique arena in which to study and tune the fundamental nature of weak inter-molecular interactions, including C--H \(\cdots\) O and C--H \(\cdots\,\pi\) hydrogen bonds and both cyclic and acyclic \(\pi\) -- \(\pi\) effects. Dimethylformamide (DMF, \(\mathrm{Me_{2}NC(=O)H}\)) and dimethylacetamide (DMAc, \(\mathrm{Me_{2}NC(=O)Me}\)) are the simplest aprotic amides, in which the proton donor (protic) N--H groups present in formamide (FA, \(\mathrm{H_{2}NC(=O)H}\)), N-methylformamide (NMF, \(\mathrm{MeHNC(=O)H}\)), and N-methylacetamide (NMAc, \(\mathrm{MeHNC(=O)Me}\)) are replaced by N--Me (Figure [\[fgr:dipoles\]](#fgr:dipoles){reference-type="ref" reference="fgr:dipoles"}). The aprotic nature and the high dipolar character of these amides make them the ideal candidate for studying weak competitive interactions in the liquid state. Both DMF and DMAc are planar acyclic amides, where partial double bond character in the N--C=O framework arises from \(\pi\) electron delocalisation that enforces the planarity of the molecule. DMF and DMAc have similar dipole moments (\(\mu\) = 3.86 D and 3.72 D respectively) and relative permittivities (\(\epsilon_{r}\) = 36.8 and 37.8 at 20 C respectively, Table [1](#tb:TDparameters){reference-type="ref" reference="tb:TDparameters"}) which lead to strong dipole--dipole interactions and relative orientational effects in the liquid structure. Both molecules are regarded as weak Lewis bases, with donor numbers (DN) of 26.6 \(\mathrm{kcal\, mol^{-1}}\) and 27.8 \(\mathrm{kcal\, mol^{-1}}\) and acceptor numbers (AN) of 16.0 \(\mathrm{kcal\, mol^{-1}}\) and 13.6 \(\mathrm{kcal\, mol^{-1}}\) respectively for DMF and DMAc, Table [1](#tb:TDparameters){reference-type="ref" reference="tb:TDparameters"}. Furthermore, the presence of a C(=O)--H group in DMF raises the possibility of hydrogen bonding by a weakly donating formic H atom. On a practical level, this functionality also means that while DMF is one of the most heavily used solvents for chemical synthesis, it can react under highly basic conditions and with strong reducing and chlorinating agents. DMAc is usually more inert, and so has complementary applications for example in the production of pharmaceuticals and polymers. Neutron diffraction studies of liquid DMF and DMAc have shown well-defined local structures. For both cases, the coordination number of molecules in the first solvation shell is found to be around 13, with a clear second shell also present. In DMF, weak C(=O)--H \(\cdots\) O hydrogen bonds are observed that are thought to be electrostatic in nature. In DMF the first solvation shell shows the expected preference for anti-parallel dipole orientation between molecules, while in DMAc parallel dipoles maximise dispersion forces between the \(\pi\)-delocalised O=C--N backbones and methyl groups. These results are consistent with Raman, IR spectroscopy and molecular dynamics studies of liquid DMF, all of which reported a highly structured first solvation shell with weak hydrogen bonding from carbonyl and methyl H atoms, in line with the 'relatively short' C--H \(\cdots\) O=C contacts seen earlier by gas-phase electron diffraction. Dimethylsulfoxide (DMSO, \(\mathrm{Me_{2}S=O}\)) is a pyramidal molecule with high dipole moment (\(\mu\) = 3.96 D), weak Lewis base character (DN = 29.8 \(\mathrm{kcal\, mol^{-1}}\) and AN = 19.3 \(\mathrm{kcal\, mol^{-1}}\)), and a remarkably high permittivity for an aprotic solvent (\(\epsilon_{r}\) = 47.2 at 20 C). Its unique properties are due to the combination of a soft lone pair on the sulfur atom and the strong polarisation of the S=O bond. DMSO, like DMF and DMAc, is miscible with water and many organic solvents, and has an unique ability to solvate a wide range of chemical species from apolar hydrocarbons to entirely dissociated salts. DMSO is therefore extremely important in processing and technology. Moreover, DMSO is able to penetrate human skin with a non-destructive effect on tissues and is a keystone protectant in cryobiology. Additional theoretical interest stems from the long-standing question of how to best represent the sulfoxide bond: \(\mathrm{S^{+}}\)-- \(\mathrm{O^{-}}\) rather than a formal S=O double bond? Structural studies of liquid DMSO by X--ray and neutron diffraction have reported nearest neighbour coordination numbers in the range 11.5 -- 13.8 and have provided evidence for short-range antiparallel alignment of dipoles, with head-to-tail ordering at longer distances. In addition, weak methyl hydrogen to oxygen inter-molecular contacts were observed at distances of approximately 3.4 Å, but as such these probably do not constitute hydrogen bonds. While the bulk physicochemical properties of DMF, DMAc and DMSO are therefore similar (Table [1](#tb:TDparameters){reference-type="ref" reference="tb:TDparameters"}), the pure liquids exhibit contrasting local structures. This latter point immediately raises the question as to which interactions will dominate in mixtures of these molecules, and in particular how, and to what extent, DMSO is accommodated within the local solvation environments of DMF and DMAc (and *vice versa*). Previous studies of mixtures the polar protic solvent NMF in DMSO by neutron diffraction point to the formation of a strong N--H \(\cdots\) O=S hydrogen bond between the protic amine group of NMF and the oxygen of DMSO at 1.6 Å. The latter distance is considerably shorter than the typical strong hydrogen bonding in the liquid state: taking the interaction between water molecules as an example, the first O--H \(\cdots\) O contact is found at 1.85 Åat ambient conditions. NMF-DMSO hydrogen bonding is more similar in length to liquid HF, one of the shortest hydrogen bonds reported in a liquid. As a consequence, NMF and DMSO molecules form very stable dimers in the mixtures. This in turn results in well-organized first and second solvation shells with a preference for heteromolecular NMF--DMSO hydrogen bonds, rather than homomolecular NMF--NMF. NMF, though, is a protic, highly polar solvent (\(\mu\) = 3.86 D), with extremely high relative permittivity (\(\epsilon_r\) = 181 at 25 C) and the ability to act as both proton donor and acceptor via its N--H and C=O groups. In clear contrast to NMF, DMF and DMAc are aprotic solvents that only form weak hydrogen bonds via the C--H and methyl groups. As such, they will pose a very different conundrum for DMSO as a co-solvent when they are compared with NMF. In this study we have used neutron diffraction in conjunction with isotopic substitution of hydrogen (H) by deuterium (D) to study both the pure liquid amides DMF and DMAc and equimolar 50 : 50 mixtures of these with DMSO to understand the role of weak intermolecular interactions, such as weak hydrogen bonding and dispersion forces, on a molecular level and to reveal the role of DMSO as co-solvent in a aprotic environment. The use of the Empirical Potential Structure Refinement (EPSR) computations has allowed us to uncover the 3--dimensional site--site correlations in these systems.
# Theoretical Basis
The function of interest which can be extracted from a neutron diffraction measurement is know as the total structure factor, \(F(Q)\), which can be written as: \[F(Q)=\sum_{\alpha,\beta \geqslant \alpha }{(2-\delta_{\alpha \beta})\,b_{\alpha} b_{\beta} c_{\alpha} c_{\beta}\, (S_{\alpha\beta}(Q)-1 )} \label{eqn:tsf}\] where \(c_{\alpha}\) and \(c_{\beta}\), \(b_{\alpha}\) and \(b_{\beta}\) are respectively the fractional concentrations of the atomic species \(\alpha\) and \(\beta\) and the (isotope dependent) coherent neutron scattering lengths, \(Q=4\pi\frac{\sin\theta}{\lambda}\) is the magnitude of the neutron scattering vector, and \(S_{\alpha\beta}(Q)\) the Faber-Ziman partial structure factor for any two types of atoms. There is, therefore, a unique *F(Q)* for each isotopic composition (isotopologue) of a sample. In particular, we can exploit the difference in sign and magnitude between the coherent neutron scattering lengths of hydrogen (\(b_H =\)-- 3.74 fm) and deuterium (\(b_{D} =\) 6.72 fm) to distinguish between specific sites in a molecule and thereby measure multiple distinct \(F(Q)\)s. This approach constrains the overall structure refinement so that we can interrogate the individual site-site correlations needed to describe the structure of a complex liquid. The partial structure factors are related to the partial radial distribution functions (RDFs), \(g_{\alpha\beta}(r)\), via Fourier transformation: \[S_{\alpha\beta}(Q)= 1 + 4\pi \rho \int^{\infty}_{0} dr \; r^{2} \,[g_{\alpha \beta}(r)-1 ] \,\frac{sin(Qr)}{Qr}. \label{eqn:pdf}\] where \(\rho\) is the atomic number density and *r* is the distance between two species \(\alpha\) and \(\beta\). The \(g_{\alpha\beta}(r)\)s represent the probability density of finding, by spherical averaging, an atom of species \(\beta\) at distance *r* from an atom of species \(\alpha\) chosen as origin of the reference system. These functions therefore contain important site-specific structural information of the sample. In a liquid system, \(g_{\alpha\beta}(r)\) tends to 1 at large values of *r*. In order to quantify the average coordination number, \(N_{\alpha\beta}(r_{0})\), of sites of type \(\beta\) in proximity to a site of type \(\alpha\) up to a maximum distance \(r_{0}\), one can integrate the partial radial distribution function \(g_{\alpha \beta}(r)\) over separation distance, *r*: \[N_{\alpha\beta}(r_{0}) = 4\pi \int_{0}^{r_{0}} g_{\alpha\beta}(r)\, \rho_\beta \, r^{2} \, dr \label{eqn:coord}\] where \(\rho_{\beta}\) is the number density of species \(\beta\) and \(r_{0}\) is the maximum distance of integration. By definition, the first coordination number gives the average number of sites of species \(\beta\) present in a sphere of radius \(r_{0}\) centred on a site of species \(\alpha\). Traditionally, the upper limit of the integral in Equation [\[eqn:coord\]](#eqn:coord){reference-type="ref" reference="eqn:coord"} is the position of the first minimum of the partial \(g_{\alpha \beta}(r)\). Alternatively, the cumulative coordination numbers can be plotted as a function of the distance \(r\) from the central species. Beyond a one-dimensional analysis, the Spatial Density Functions (SDFs) are a three-dimensional map of the density of neighbouring molecules around a central molecule as a function of angular distance, *r*, and angular position \(\theta\). The SDFs therefore represent regions of space around a central molecule that are most likely to be occupied by a molecule of the same or another species at a given distance.
# Experimental Details
Experimental data have been acquired at the Near and InterMediate Range Order Diffractometer (NIMROD) at the ISIS Neutron and Muon Source (Didcot, UK) across a Q range of 0.05 Å\(^{-1}-\) 50 Å\(^{-1}\). DMF, DMAc and DMSO and their isotopes were purchased from Sigma Aldrich with purities \(\geq 99.5\%\) and handled under inert atmosphere. For pure DMF and DMAc, fully hydrogenated, fully deuterated and a 50:50 mixture of hydrogenated and deuterated liquids were loaded into \(\mathrm{Ti_{0.68}\,Zr_{0.32}}\) null scattering cells to give a total of 3 isotopically distinct samples for each liquid amide. To produce 50 : 50 amide/DMSO mixtures, the anhydrous liquids were mixed to obtain 7 isotopically distinct samples for DMF/DMSO and 5 for DMAc/DMSO, as summarised in Table [\[samples\]](#samples){reference-type="ref" reference="samples"}. The samples were inserted into flat-plate null coherent scattering titanium/zirconium cells, with 1 mm sample and wall thicknesses. Each composition was run for a minimum of 2 hours at 298 K. Data for the pure liquid DMF and DMAc have been reanalysed using the same methods and protocols as the mixture to allow for a more rigorous comparison. To allow data correction and calibration, scattering data were also collected from the empty instrument, empty sample cells, and an incoherent scattering vanadium-niobium reference slab of thickness 3 mm. Reduction of the experimental data, including absolute normalisation, background subtraction, and multiple and inelastic scattering corrections, has been conducted using standard procedures as implemented within the Gudrun package.
# Computational Details
The EPSR method consists of a classical Monte Carlo molecular simulation which takes initial seed potentials for modelling pairwise interactions, and subsequently refines these through the incorporation of an empirical potential. This empirical potential is calculated with reference to any mismatch between the experimental and simulated data, until a satisfactory agreement between the calculated structure factors and the measured neutron scattering data is reached. In this manner, a three-dimensional structural model of the system can be obtained which is consistent with the experimental data. The inter-molecular potential between two atomic sites \(\alpha\) and \(\beta\) is modelled in EPSR via a Lennard--Jones 12--6 function plus a Coulombic term: \[U_{\alpha \beta}(r_{ij}) = 4 \epsilon_{\alpha \beta} \Bigl[\Bigl(\frac{\sigma_{\alpha \beta}}{r_{ij}}\Bigr)^{12}-\Bigl(\frac{\sigma_{\alpha \beta}}{r_{ij}}\Bigr)^{6}\Bigr]+ \frac{q_{\alpha} q_{\beta}}{4 \pi \epsilon_{0} r_{ij}} \label{eqn:LJ}\] where \(q_{\alpha,\beta}\) are the atomic partial charges, \(\epsilon_{\alpha \beta}\) and \(\sigma_{\alpha \beta}\) are the well depth parameter and the range parameter respectively and are given by the Lorentz--Berthelot mixing rules in terms of their values of the individual atoms. All the molecules were generated in \(Avogadro\) and their geometry is optimised for 500 steps using the MMFF94 force field. (see Table S1 for intramolecular parameters) The intermolecular seed potentials are taken from the OPLS series of force-fields. The cubic EPSR boxes of side-length 44.80 Åand 47.58 Åcontain 700 molecules for pure liquid DMF and DMAc while cubic boxes of side 49.77 Åand 51.63 Åcontaining 1000 molecules of which 500 are DMF/DMAc and 500 are DMSO. The atomic number densities for the four systems are 0.0934 atoms/Å\(^{3}\) and 0.0975 atoms/Å\(^{3}\) for the DMF and DMAc pure liquid and 0.08925 atoms/Å\(^{3}\) and 0.09080 atoms/Å\(^{3}\) for DMF/DMSO and DMAc/DMSO mixtures. These values are obtained from a weighted average of the relevant bulk densities of the pure liquids and are verified by reference to the overall scattering levels of the experimental data. The labels assigned to atomic sites of the DMF, DMAc and DMSO molecules are shown for clarity in Figure [\[fgr:model3\]](#fgr:model3){reference-type="ref" reference="fgr:model3"}.
# Results and Discussion
The isotopically distinct experimental neutron diffraction structure factors, \(F(Q)\), are plotted for the pure liquid amides and their DMSO mixtures against the EPSR model in Figure [\[fgr:fit\]](#fgr:fit){reference-type="ref" reference="fgr:fit"}. Excellent agreement between the experimental data and model has been achieved for each data-set; the small discrepancies at low--\(Q\) in fully hydrogenated samples such as h-DMF and h-DMAc are attributed to a residual presence of inelastic and multiple scattering events. The rise in low--\(Q\) scattering for samples such as hd-DMF and hd-DMAc can then be attributed to the fact that these liquids are comprised of a mixture of fully hydrogenated and fully deuterated molecules, Table [\[samples\]](#samples){reference-type="ref" reference="samples"}. This isotopic partitioning on individual molecules leads to a genuine rise in elastic scattering that is well captured by the EPSR model. These neutron diffraction data (Figure [\[fgr:fit\]](#fgr:fit){reference-type="ref" reference="fgr:fit"} solid-line) show clearly that the liquid amides are fully miscible with DMSO at this concentration, as there is an absence of any residual low--\(Q\) signal that would indicate the presence of homomolecular clustering. This observation can be confirmed by examining the molecular Centre-of-Mass (CoM) radial distribution functions, \(g_{CoM-CoM}(r)\) obtained from the EPSR model.
Selected amide--DMSO \(g_{\alpha \beta}(r)\) partial RDFs are shown in Figure [\[fgr:g(r)amide-dmso\]](#fgr:g(r)amide-dmso){reference-type="ref" reference="fgr:g(r)amide-dmso"} for DMF/DMSO (left) and DMAc/DMSO (right) respectively, along with the corresponding peak positions and coordination numbers in Table [\[tb:coord-dmso\]](#tb:coord-dmso){reference-type="ref" reference="tb:coord-dmso"}. Further pairs are given in supporting information Figure S5. Regarding the heteromolecular hydrogen bonding, we note that for DMF, the formic proton \(\mathrm{H^{F}}\) to DMSO oxygen distribution shows a peak at \(\sim\) 2.5 Å. This is an indication of a weak, electrostatic hydrogen bond, with C(=O)--H \(\cdots\) O=S distance almost identical to that observed for C(=O)--H \(\cdots\) O=C between two DMF molecules (Table [\[tb:coord-amide\]](#tb:coord-amide){reference-type="ref" reference="tb:coord-amide"}). Weak hydrogen bonds are also observed between amide methyl protons (H\(\mathrm{^{A}}\), H\(\mathrm{^{Z}}\) and H\(\mathrm{^{E}}\)) and DMSO oxygen, and DMSO methyl protons and amide oxygen, at distances between \(\sim\) 2.6 and 2.8 Å. This methyl C--H \(\cdots\) O interaction is extremely weak for \(\mathrm{sp^{3}}\) carbon and in this case is facilitated by the high molecular dipoles of DMF, DMAc and DMSO. The partial RDFs relative to the amide O=C--N backbone and the O, S, and C sites of DMSO are presented in Figure [\[fgr:g(r)amide-dmso\]](#fgr:g(r)amide-dmso){reference-type="ref" reference="fgr:g(r)amide-dmso"}, with corresponding peak positions and coordination numbers in Table [\[tb:coord-dmso\]](#tb:coord-dmso){reference-type="ref" reference="tb:coord-dmso"}. In the case of DMF--DMSO, the N--O, O--O, C--O and O--C partial RDFs show only subtle differences when compared with their amide--amide counterparts. This is consistent with our observation that DMF--DMF backbone RDFs are very similar in the bulk and mixed liquids. We attribute this relative insensitivity to the presence of a formic proton, H\(\mathrm{^{F}}\), and consequent C(=O)--H \(\cdots\) O hydrogen bond as a dominant structural motif in DMF--DMSO mixtures. This C(=O)--H \(\cdots\) O interaction is absent in DMAc, where we observe only very weak Me--H \(\cdots\) O hydrogen bonds and dispersion interactions. In this case, DMAc--DMAc backbone interactions were displaced to longer distances in the mixture. We see faint indications in the DMAc--DMSO RDFs that DMSO may compensate for this effect. Specifically, we point to the lower separation shoulders in \(g_{N-O}(r)\) and \(g_{C-O}(r)\) occuring at around 3.4 Å. To obtain more detailed insight into the solvation shells, we need to look beyond the radially averaged representations of \(g_{\alpha\beta}(r)\) and \(N_{\alpha\beta}(r)\) and we therefore turn again to the 3--dimensional spatial density functions (SDFs). Site--specific spatial density functions are displayed for pure amides and amide--DMSO mixtures in Figure [\[fgr:amide-sdf\]](#fgr:amide-sdf){reference-type="ref" reference="fgr:amide-sdf"}. In each case, the reference molecule is shown at the CoM origin, and we focus on the 3--dimensional distribution of neighbouring C=O and S=O groups. In these SDFs, carbonyl carbon density is shown in dark grey, oxygen in pink, and sulfur in orange. The DMF--DMF functions (Figure [\[fgr:amide-sdf\]](#fgr:amide-sdf){reference-type="ref" reference="fgr:amide-sdf"}, panels a and b) show a band of oxygen density, with more distant carbonyl carbon, centred around the formic proton. This feature extends at longer distances towards the methyl Me\(\mathrm{^{Z}}\) groups, and, through a narrow strip, towards Me\(\mathrm{^{E}}\) and Me\(\mathrm{^{Z}}\). These features confirm the presence of C(=O)--H \(\cdots\) O and more distant Me--H \(\cdots\) O hydrogen bonds. Overall, the pure and mixed DMF systems have broadly similar amide--amide spatial density, as predicted from our analysis of the RDFs. Relatively delicate changes to the DMF SDFs occur on mixing, and are centred on the lobes corresponding to weaker interactions to the methyl groups, away from the formic proton. The DMF--DMSO SDFs (Figure [\[fgr:amide-sdf\]](#fgr:amide-sdf){reference-type="ref" reference="fgr:amide-sdf"}, panels e and f) support our asseveration that local solvation environment is occupied by DMF and DMSO in equal manner in the mixture. We note that the sulfonyl oxygen density tracks the main features of the amide oxygen, and, likewise, the sulfonyl sulfur and carbonyl carbon. The most likely location for DMSO methyl carbons is a band centred around the amide oxygen, indicative again of weak, directionally rather unconstrained, Me--H \(\cdots\) O hydrogen bonds. The DMAc--DMAc functions (Figure [\[fgr:amide-sdf\]](#fgr:amide-sdf){reference-type="ref" reference="fgr:amide-sdf"}, panels c and d) show clearly contrasting behaviour to those of DMF. In the pure liquid, the most likely DMAc oxygen and carbonyl carbon density is symmetrically above and below the plane of the O=C--N backbone, with a connecting strip directed towards the methyl protons H\(\mathrm{^{A}}\) and H\(\mathrm{^{E}}\). This picture is fundamentally different to that for DMF shown in Figure [\[fgr:amide-sdf\]](#fgr:amide-sdf){reference-type="ref" reference="fgr:amide-sdf"} panels a and b. Also in contrast to DMF, for DMAc there are clear differences in the amide--amide SDFs for the pure and mixed systems. Specifically, Figure [\[fgr:amide-sdf\]](#fgr:amide-sdf){reference-type="ref" reference="fgr:amide-sdf"} panel d shows an additional lobe directed towards the methyl protons H\(\mathrm{^{Z}}\) and H\(\mathrm{^{E}}\), and concomitant loss of density in the features seen in the pure system, panel c. The reason for this behaviour becomes clear when we examine the DMAc--DMSO SDFs. We see that in the mixture, DMSO sulfonyl oxygen and sulfur mimic the closest approach of DMAc carbonyl oxygen and carbon in the pure liquid. We conclude that DMSO prefers to approach DMAc from above and below the plane of the O=C--N backbone, or axially around Me\(\mathrm{^{A}}\) and Me\(\mathrm{^{E}}\). In doing this, DMSO molecules displaces some of the DMAc--DMAc interactions observed in the pure liquid, thereby displacing the DMAc--DMAc density in the mixture towards Me\(\mathrm{^{Z}}\) and Me\(\mathrm{^{E}}\). This is entirely consistent with the shifts observed in the DMAc--DMAc backbone RDFs (Figure 6). As with DMF, the most likely location for DMSO methyl carbons is a band centred around the amide oxygen, but broadened and cleft due to the presence of Me\(\mathrm{^{A}}\).
# Conclusions
Neutron diffraction augmented by isotopic substitution of hydrogen (H) for deuterium (D) has been used to study pure DMF and DMAc, and equimolar mixtures of DMF in DMSO and DMAc in DMSO in the liquid state. The atomistic resolution provided by neutron diffraction is critical for understanding weak inter--molecular interactions on a molecular level, as these bonding motifs are elusive and often invisible to many experimental techniques. Empirical Potential Structure Refinement (EPSR) has been used to generate 3--dimensional atomistic models that are consistent with the experimental data. This approach has enabled us to uncover individual site--site interactions, and to conduct detailed comparison between the pure and mixed systems. Our scattering data show that the amides and DMSO are mixed on the nanoscale. Analysis of the EPSR Centre--of--Mass (CoM) correlations shows that in all of our systems the coordination shell contains \(\sim\) 13 -- 14 molecules, and that in the mixed systems there is, on average, near ideal solvation shell sharing between amide and DMSO. Examination of site--specific correlations reveals a rich structural landscape, in which replacement of the formic proton H\(\mathrm{^{F}}\) in DMF by\(\) the methyl group Me\(\mathrm{^{A}}\) in DMAc leads to fundamentally different solvation environments for these amides in both the pure liquids and mixtures. Weak C--H \(\cdots\) O hydrogen bonds are formed by formic (DMF) and methyl (DMF, DMAc) protons, with distances around 2.5 Åand 3.0 Å. In addition, we observe dispersion interactions above and below the plane of the O=C--N amide \(\pi\)--backbones, particularly in the DMAc systems in which dispersive forces are expected to be predominant. By comparing pure amides with the liquid mixtures, we show that DMSO has a noteworthy ability to share in hydrogen bonding to both formic and methyl groups, matching the bond distances observed in the pure amides leading to similar solvation motifs and perfect mixing. As a result, DMSO is able to penetrate the amide--amide solvation shells while causing only subtle disruption to the amide--amide interactions. The new knowledge provided by our study is particularly important to tailor electrolytes in confined geometries such as battery electrodes and supercapacitors, since molecular mixing and local interactions are likely to impact on performance. | {'timestamp': '2022-12-19T02:13:39', 'yymm': '2212', 'arxiv_id': '2212.08530', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08530'} |
# Introduction
This paper aims at further investigating the interplay between reduction numbers, eventually taken relative to a given module, and the Castelnuovo-Mumford regularity of Rees and associated graded algebras and modules. The basic tool is the so-called Ratliff-Rush closure, a well-known operation which goes back to the classical work, and also for the relative case. These topics have been studied for decades and the literature about them is extensive; we mention, e.g.,, ,, ,, ,, ,, ,, ,, ,, , ,. Let us briefly describe our two main motivations. First, an interesting question raised by Rossi, Trung, and Trung concerning the Castelnuovo-Mumford regularity of the Rees algebra of a zero-dimensional ideal in a 2-dimensional Buchsbaum local ring, and second, a result due to Mafi on the Castelnuovo-Mumford regularity of the associated graded ring of a zero-dimensional ideal in a 2-dimensional Cohen-Macaulay local ring, both connected to the reduction number of the ideal. For the former (see Section [3](#R-T-T){reference-type="ref" reference="R-T-T"}) we provide answers in a couple of special cases, and we even ask whether one of the situations always occurs. For the latter (see Section [4](#gen-of-Mafi){reference-type="ref" reference="gen-of-Mafi"}), we are able to establish a generalization which extends dimension \(2\) to arbitrary dimension and in addition takes into account the more general context of blowup modules (relative to any given Cohen-Macaulay module of positive dimension). We then provide two sections of applications. In Section [5](#app){reference-type="ref" reference="app"} we show, for instance, that under appropriate conditions the Castelnuovo-Mumford regularity of Rees modules is not affected under regular hyperplane sections, and moreover, by investigating the role of postulation numbers with a view to the regularity of Rees algebras, we derive an improvement (by different arguments) of a result originally obtained in on the independence of reduction numbers. Still in Section [5](#app){reference-type="ref" reference="app"} we provide an application regarding ideals of linear type (i.e., ideals for which the Rees and symmetric algebras are isomorphic) in a standard graded polynomial ring over an infinite field. Precisely, if \(I\) is a zero-dimensional ideal with reduction number zero, then \(I\) is of linear type (the converse is easy and well-known). Finally, in Section [6](#Ulrich){reference-type="ref" reference="Ulrich"}, we establish new facts on the theory of generalized Ulrich ideals introduced in. Given such an ideal in a Cohen-Macaulay local ring (of positive dimension), we determine the Castelnuovo-Mumford regularity of its blowup algebras and describe explicitly its Hilbert-Samuel polynomial and postulation number, which as far as we know are issues that have not been considered in the literature. We also correct a result from and improve results of, and (where extra hypotheses were required), regarding a well-studied link between low reduction numbers and Hilbert functions. In addition, by detecting a normal Ulrich ideal in the ring of a rational double point, we give a negative answer to the 2-dimensional case of a question posed over fifteen years ago in. . Throughout the paper, by *ring* we mean commutative, Noetherian, unital ring. If \(R\) is a ring, then by a *finite* \(R\)-module we mean a finitely generated \(R\)-module.
# Ratliff-Rush closure
The concept of the so-called *Ratliff-Rush closure* \(\widetilde{I}\) of a given ideal \(I\) in a ring \(R\) first appeared in the investigation of reductions of ideals developed in. Precisely, \[\widetilde{I} \, = \, \bigcup_{n \geq 1}\,I^{n+1}: I^{n}.\] This is an ideal of \(R\) containing \(I\) which refines the integral closure of \(I\), so that \(\widetilde{I}=I\) whenever \(I\) is integrally closed. Now suppose \(I\) contains a regular element (i.e., a non-zero-divisor on \(R\)). Then \(\widetilde{I}\) is the largest ideal that shares with \(I\) the same sufficiently high powers. Moreover, if \(M\) is an \(R\)-module then defined the *Ratliff-Rush closure of \(I\) with respect to \(M\)* as \[\widetilde{I_{M}} \, = \, \bigcup_{n \geq 1}\,I^{n + 1}M :_{M} I^{n} \, = \, \{m \in M \, \mid \, I^{n}m \subseteq I^{n + 1}M \ \mathrm{for \ some \ n\geq 1} \}.\] Clearly, this retrieves the classical definition by letting \(M = R\). Furthermore, \(\widetilde{I_{M}}\) is an \(R\)-submodule of \(M\), and it is easy to see that \(IM\subseteq \widetilde{I}M \subseteq \widetilde{I_{M}}\). If the equality \(IM=\widetilde{I_{M}}\) holds, the ideal \(I\) is said to be *Ratliff-Rush closed with respect to \(M\)*; in case \(M=R\), we simply say that \(I\) is *Ratliff-Rush closed* (some authors also use the expression *Ratliff-Rush ideal*).
Note that, if the ideal \(I\) contains an \(M\)-regular element, Lemma [\[propnaghipour\]](#propnaghipour){reference-type="ref" reference="propnaghipour"}(b) enables us to define the number \[s^{*}(I,M) \, := \, \mathrm{min}\,\{n \in \mathbb{N} \, \mid \, \widetilde{I^{i}_{M}} = I^{i}M \ \mathrm{for \ all} \ i \geq n \},\] which we simply write \(s^{*}(I)\) if \(M=R\). Since the equality \(\widetilde{I^{i}_{M}} = I^{i}M\) holds trivially for \(i=0\), we have that \(s^{*}(I,M)\geq 0\) if and only if \(s^{*}(I,M)\geq 1\). Thus we can establish that, throughout the entire paper, \[s^{*}(I,M) \, \geq \, 1.\]
If \(R\) is local, the above concept allows for an alternative, useful expression for \(s^{*}(I,M)\). Indeed, in this case, shows that \[s^{*}(I,M) \, = \, \mathrm{min}\,\{n \in \mathbb{N} \, \mid \, I^{i + 1}M :_{M} x = I^{i}M \ \mathrm{for \ all} \ i \geq n \},\] for any given \(M\)-superficial element \(x\) of \(I\).
The condition \(\bigcap_{n} I^{n}M = 0\) is satisfied whenever \(R\) is a local ring, as guaranteed by the well-known Krull's intersection theorem, so in order to use Lemma [\[lemmaswanhun\]](#lemmaswanhun){reference-type="ref" reference="lemmaswanhun"} we just need to find an \(M\)-regular element in \(I\). Lemma [\[lemmaswanhun1\]](#lemmaswanhun1){reference-type="ref" reference="lemmaswanhun1"} and Lemma [\[lemmaswanhun\]](#lemmaswanhun){reference-type="ref" reference="lemmaswanhun"} put us in a position to prove Proposition [\[regelem\]](#regelem){reference-type="ref" reference="regelem"} below, which, according to, is known when \(M = R\). Moreover, in our general context, part (b) has been stated in. However, the proofs are omitted in both situations, so for completeness we provide them here.
As a consequence of item (a) we derive the following result, which will be a key tool in Subsection [3.2](#RTT){reference-type="ref" reference="RTT"} in the case \(M=R\). Part (b) of Proposition [\[regelem\]](#regelem){reference-type="ref" reference="regelem"} will be used in other sections (e.g., in the proof of Theorem [\[Mafigene\]](#Mafigene){reference-type="ref" reference="Mafigene"}).
# First results and a question of Rossi-Trung-Trung {#R-T-T}
Let \(S = \bigoplus_{n \geq 0}S_{n}\) be a finitely generated standard graded algebra over a ring \(S_{0}\). As usual, by *standard* we mean that \(S\) is generated by \(S_1\) as an \(S_0\)-algebra. We write \(S_{+} = \bigoplus_{n \geq 1}S_{n}\) for the ideal generated by all elements of \(S\) of positive degree. Fix a finite graded \(S\)-module \(N\neq 0\). A sequence \(y_{1}, \ldots, y_{s}\) of homogeneous elements of \(S\) is said to be an *\(N\)-filter regular sequence* if \[y_{i} \notin P, \ \ \ \ \ \forall \ P \in \textrm{Ass}_{S}\left(\frac{N}{(y_{1}, \ldots, y_{i-1})N}\right)\setminus V(S_{+}), \ \ \ \ \ i = 1, \ldots, s.\] Now, for a graded \(S\)-module \(A=\bigoplus_{n \in {\mathbb Z}}A_{n}\) satisfying \(A_n=0\) for all \(n\gg 0\), we let \(a(A) = \textrm{max}\{n\in {\mathbb Z} \ | \ A_{n} \neq 0\}\) if \(A\neq 0\), and \(a(A)=-\infty\) if \(A=0\). For an integer \(j\geq 0\), we use the notation \[a_{j}(N) \, := \, a(H_{S_{+}}^{j}(N)),\] where \(H_{S_{+}}^{j}(-)\) stands for the \(j\)-th local cohomology functor with respect to the ideal \(S_{+}\). It is known that \(H_{S_{+}}^{j}(N)\) is a graded module with \(H_{S_{+}}^{j}(N)_n=0\) for all \(n\gg 0\); see, e.g.,. Thus, \(a_{j}(N)<+\infty\). We can now invoke one of the main numerical invariants studied in this paper.
It is well-known that \(\mathrm{reg}\,N\) governs the complexity of the graded structure of \(N\) and is of great significance in commutative algebra and algebraic geometry, for instance in the study of degrees of syzygies over polynomial rings. The literature on the subject is extensive; see, e.g., , , , and.
## First results
Let \(I\) be an ideal of a ring \(R\). The *Rees ring of \(I\)* is the blowup algebra \({\mathcal R}(I) = \bigoplus_{n \geq 0}I^{n}t^n\) (as usual we put \(I^0=R\)), which can be expressed as the standard graded subalgebra \(R[It]\) of \(R[t]\), where \(t\) is an indeterminate over \(R\). Now, for a finite \(R\)-module \(M\), the *Rees module of \(I\) relative to \(M\)* is the blowup module \[{\mathcal R}(I, M) \, = \, \bigoplus_{n \geq 0}\,I^{n}M,\] which in particular is a finite graded module over \({\mathcal R}(I, R)={\mathcal R}(I)\). We will be interested in the relation between \({\rm reg}\,{\mathcal R}(I, M)\) and another key invariant that will be defined shortly. First we recall the following fact.
The usefulness of the proposition below will be made clear in Remark [\[reg\<r\]](#reg<r){reference-type="ref" reference="reg<r"} and Remark [\[cotareg\]](#cotareg){reference-type="ref" reference="cotareg"}. Notice that the case \(M=R\) recovers; in fact, the proof is essentially the same and we give it in a little more detail.
Let us now introduce another fundamental blowup module. For an ideal \(I\) in a ring \(R\) and a finite \(R\)-module \(M\), the *associated graded module of \(I\) relative to \(M\)* is the blowup module \[{\mathcal G}(I, M) \, = \, \bigoplus_{n \geq 0}\, \frac{I^{n}M}{I^{n + 1}M} \, = \, {\mathcal R}(I, M)\otimes_RR/I,\] which is a finite graded module over the *associated graded ring* \({\mathcal G}(I) = {\mathcal G}(I, R)\) of \(I\), i.e., \({\mathcal G}(I) = {\mathcal R}(I)\otimes_RR/I\). Notice that this algebra is standard graded over \({\mathcal G}(I)_0=R/I\). As usual, we set \({\mathcal G}(I)_+= \bigoplus_{n \geq 1}I^{n}/I^{n + 1}\). Here it is worth mentioning for completeness that, under a suitable set of hypotheses, there is a close connection between the Cohen-Macaulayness of \({\mathcal G}(I, M)\) and the property \({\rm r}(I, M)\leq 1\). We refer to. The following fact generalizes (also ).
Next, we record a result which provides another set of conditions under which the equality \(\mathrm{reg}\,{\mathcal R}(I, M) = {\rm r}_{J}(I,M)\) holds. It will be a key ingredient in the proof of Theorem [\[Mafigene\]](#Mafigene){reference-type="ref" reference="Mafigene"}. For completeness, we recall that the condition on \(x_1, \ldots, x_s\) present in part (a) below means that no element of the sequence lies in the ideal generated by the others and in addition there are equalities \(0 :_Mx_{1}x_j = 0 :_Mx_j\) for \(j=1, \ldots, s\) and \((x_1, \ldots, x_i)M :_Mx_{i+1}x_j = (x_1, \ldots, x_i)M :_Mx_j\) for \(i=1, \ldots, s-1\) and \(j=i+1, \ldots, s\). We observe that parts (a) and (b) are satisfied if the elements \(x_1, \ldots, x_s\) form an \(M\)-sequence.
## On a question of Rossi, Trung, and Trung {#RTT}
In this part we focus on the interesting question raised in as to whether, for a ring \(R\) and an ideal \(I\) having a minimal reduction \(J\), the formula \[{\rm reg}\,\mathcal{R}(I) \, = \, \mathrm{min}\,\{n \geq {\rm r}_{J}(I) \, \mid \, I^{n + 1}: I = I^{n}\}\] holds under the hypotheses of Lemma [\[rossidinhtrung\]](#rossidinhtrung){reference-type="ref" reference="rossidinhtrung"} below. This lemma is a crucial ingredient in the proof of Theorem [\[rossiquest\]](#rossiquest){reference-type="ref" reference="rossiquest"}, which will lead us to partial answers to the question as will be explained in Remark [\[partial\]](#partial){reference-type="ref" reference="partial"} and recorded in Corollary [\[our-answer\]](#our-answer){reference-type="ref" reference="our-answer"}.
As far as we know, under the hypotheses of Lemma [\[rossidinhtrung\]](#rossidinhtrung){reference-type="ref" reference="rossidinhtrung"} there is no example satisfying \(s^*(I)>{\rm r}_{J}(I) + 1\). It is thus natural to raise the following question, to which an affirmative answer would imply the definitive solution of the Rossi-Trung-Trung problem.
Regarding the setting relative to a module, a question is in order.
# A generalization of a result of Mafi {#gen-of-Mafi}
Our main goal in this section is to establish Theorem [\[Mafigene\]](#Mafigene){reference-type="ref" reference="Mafigene"}, which concerns the interplay between the numbers \({\rm r}_{J}(I, M)\) and \(\mathrm{reg}\,{\mathcal R}(I, M)\) in a suitable setting. As we shall explain, our theorem generalizes a result due to Mafi from (see Corollary [\[Mafigene2\]](#Mafigene2){reference-type="ref" reference="Mafigene2"}) and gives us a number of other applications, to be described in Section [5](#app){reference-type="ref" reference="app"} and also in Section [6](#Ulrich){reference-type="ref" reference="Ulrich"}. As a matter of notation, it is standard to write \({\rm grade}(I, M)\) for the maximal length of an \(M\)-sequence contained in the ideal \(I\) of the local ring \((R, \mathfrak{m})\). If \(M=R\), the notation is simplified to \({\rm grade}\,I\). Note that \({\rm grade}(I, M)\) is just the \(I\)-\({\rm depth}\) of \(M\); in particular, \({\rm grade}(\mathfrak{m}, M)={\rm depth}\,M\). First we recall two general lemmas.
Our theorem (which also holds in an appropriate graded setting) is as follows.
It seems natural to raise the following question.
The result below (the case \(s=2\)) is an immediate byproduct of Theorem [\[Mafigene\]](#Mafigene){reference-type="ref" reference="Mafigene"}.
We are now able to recover an interesting result of Mafi about zero-dimensional ideals in two-dimensional Cohen-Macaulay local rings. More precisely, by taking \(M = R\) in Corollary [\[corMafigen\]](#corMafigen){reference-type="ref" reference="corMafigen"} we readily derive the following fact.
To conclude the section, we illustrate that Corollary [\[Mafigene2\]](#Mafigene2){reference-type="ref" reference="Mafigene2"} is no longer valid if we remove the hypothesis involving the Ratliff-Rush closure.
# First applications {#app}
We now describe some applications of the results obtained in the preceding section. With the exception of Subsection [5.3](#lintype){reference-type="ref" reference="lintype"}, we shall focus on the two-dimensional case.
## Hyperplane sections of Rees modules
The first application is the corollary below, which can be of particular interest for inductive arguments in dealing with Castelnuovo-Mumford regularity of Rees modules.
Taking \(M=R\), we immediately get the following consequence.
## The role of postulation numbers
This subsection (which focuses on the classical case \(M=R\)) investigates connections of postulation numbers with the Castelnuovo-Mumford regularity of blowup algebras and reduction numbers. First recall that, if \((R, \mathfrak{m})\) is a local ring and \(I\) is an \(\mathfrak{m}\)-primary ideal of \(R\), then the corresponding *Hilbert-Samuel function* is given by \[H_{I}(n) \, = \, \displaystyle{\lambda\left(R / I^{n}\right)}\] for any integer \(n \geq 1\), and \(H_{I}(n) = 0\) if \(n \leq 0\). The symbol \(\lambda(-)\) denotes length of \(R\)-modules. It is well-known that \(H_I(n)\) coincides, for all sufficiently large integers \(n\), with a polynomial \(P_I(n)\)--the *Hilbert-Samuel polynomial of \(I\)*.
Here it is worth recalling that the functions \(H_I(n)\) and \(P_I(n)\) are defined for all integers \(n\), so \(\rho(I)\) can be--and often is--negative (as emphasized in ). In the application below we furnish a characterization of \(\mathrm{reg}\,\mathcal{R}(I)\) (which by agrees with \(\mathrm{reg}\,\mathcal{G}(I)\)), in terms, in particular, of \(\rho(I)\). Notice that our statement makes no explicit use of the concept of postulation number \(p(\mathcal{G}(I))\) for the ring \(\mathcal{G}(I)\) (see ), which we only use in the proof.
Next we provide a different proof (in fact an improvement) of. Notice that Hoa's \(c(I)\) is just \(s^*(I)-1\).
## Ideals of linear type {#lintype}
For an ideal \(I\) of a ring \(R\), there is a canonical homomorphism \(\pi \colon \mathcal{S}(I)\rightarrow \mathcal{R}(I)\) from the symmetric algebra \(\mathcal{S}(I)\) of \(I\) onto its Rees algebra \(\mathcal{R}(I)\). The ideal \(I\) is said to be of *linear type* if \(\pi\) is an isomorphism. To see what this means a bit more concretely, we can make use of some (any) \(R\)-free presentation \[R^m\stackrel{\varphi}{\longrightarrow} R^{\nu}\longrightarrow I\longrightarrow 0.\] Letting \(S=R[T_1, \ldots, T_{\nu}]\) be a standard graded polynomial ring in variables \(T_1, \ldots, T_{\nu}\) over \(R=S_0\), we can identify \(\mathcal{S}(I)=S/\mathcal{L}\), where \(\mathcal{L}\) is the ideal generated by the \(m\) linear forms in the \(T_i\)'s given by the entries of the product \((T_1 \cdots T_{\nu})\cdot \varphi\), where \(\varphi\) is regarded as a \(\nu \times m\) matrix taken with respect to the canonical bases of \(R^{\nu}\) and \(R^m\). We can also write \(\mathcal{R}(I)=S/\mathcal{J}\), for a certain ideal \(\mathcal{J}\) containing \(\mathcal{L}\). Precisely, expressing \(\mathcal{R}(I)=R[It]\) (where \(t\) is an indeterminate over \(R\)), then \(\mathcal{J}\) is the kernel of the natural epimorphism \(S\rightarrow \mathcal{R}(I)\). Now, the above-mentioned map \(\pi\) can be simply identified with the surjection \(S/\mathcal{L}\rightarrow S/\mathcal{J}\). It follows that \(I\) is of linear type if and only if \(\mathcal{J}=\mathcal{L}\). For instance, if \(I\) is generated by a regular sequence then \(I\) is of linear type (see ). Now assume \((R, \mathfrak{m})\) is either local with infinite residue field or a standard graded algebra over an infinite field. The *analytic spread* of \(I\), which we denote \(s(I)\), is defined as the Krull dimension of the special fiber ring \(\mathcal{R}(I)/\mathfrak{m} \mathcal{R}(I)\). It is a classical fact that \({\rm r}(I)=0\) if and only if \(I\) can be generated by \(s(I)\) elements (see ). So the property \({\rm r}(I)=0\) is easily seen to take place, for example, whenever \(I\) is of linear type (or if \(I\) is an \(\mathfrak{m}\)-primary parameter ideal). While, in general, there exist examples of ideals with reduction number zero that are not of linear type, the natural question remains as to under what conditions the property \({\rm r}(I)=0\) forces \(I\) to be of linear type. The corollary below is a quick application of Theorem [\[Mafigene\]](#Mafigene){reference-type="ref" reference="Mafigene"} (along with the property that the Castelnuovo-Mumford regularity controls degrees over graded polynomial rings) and contributes to this problem in a suitable setting. It is plausible to believe that the result is known to experts, but, as far as we know, it has not yet been recorded in the literature.
# More applications: Ulrich ideals and a question of Corso-Polini-Rossi {#Ulrich}
This last section provides applications concerning the theory of generalized Ulrich ideals. As we shall explain, this includes a negative answer (in dimension 2) to a question by Corso, Polini, and Rossi. Here is the central notion (and the setup) of this section.
## Regularity of blowup algebras
In order to determine the Castelnuovo-Mumford regularity of the blowup algebras of an Ulrich ideal, we first recall one of the ingredients.
This lemma immediately gives the following fact, which will be useful in the proof of Corollary [\[reg=red/ulrich\]](#reg=red/ulrich){reference-type="ref" reference="reg=red/ulrich"}.
This corollary is particularly useful to test for ideals that are not Ulrich, as illustrated in the next two examples.
The next example shows that the converse of Corollary [\[UlrRatliff\]](#UlrRatliff){reference-type="ref" reference="UlrRatliff"} is not true.
Now we are able to find the regularity of the Rees algebra--hence of the associated graded ring--of an Ulrich ideal (in arbitrary positive dimension).
## Hilbert-Samuel polynomial and postulation number
Our next goal is to determine the Hilbert-Samuel coefficients and the postulation number of an Ulrich ideal. Recall that if \((R, \mathfrak{m})\) is a Cohen-Macaulay local ring of dimension \(d\geq 1\) and \(I\) is an \(\mathfrak{m}\)-primary ideal of \(R\), then it is a well-known fact that the Hilbert-Samuel polynomial of \(I\) can be expressed as \[\label{hilbpoly} P_{I}(n) \, = \, \sum_{i = 0}^{d}(-1)^{i}{\rm e}_{i}(I)\binom{n + d-i-1}{d-i},\] where \({\rm e}_{0}(I), \ldots, {\rm e}_{d}(I)\) are the so-called *Hilbert-Samuel coefficients of \(I\)*. The number \({\rm e}_{0}(I)\) (which is always positive) is the multiplicity while \({\rm e}_{1}(I)\) is dubbed *Chern number* of \(I\). For the connection between postulation and reduction numbers, the following fact will be useful.
In the Gorenstein case, Proposition [\[HS-poly\]](#HS-poly){reference-type="ref" reference="HS-poly"} admits the following version.
Next we remark that Ulrich ideals that are not parameter ideals are the same as Ulrich ideals with non-zero Chern number.
## Further results
In this part we shall provide a correction of a proposition of Mafi as well as improvements of independent results by other authors. Let \((R, \mathfrak{m})\) be a two-dimensional Cohen-Macaulay local ring with infinite residue field, and let \(I\) be an \(\mathfrak{m}\)-primary ideal satisfying \(\widetilde{I}=I\). Let \(J\) be a minimal reduction of \(I\). Then, claims that \({\rm r}_{J}(I) = 2\) if and only if \]H_{I}(n) = P_{I}(n), \ \ \ \ \ n = 1,\, 2.\[ However, if we take \(I\) as being an Ulrich ideal, then \({\rm r}_{J}(I) \leq 1\) and we have seen in Corollary [\[UlrRatliff\]](#UlrRatliff){reference-type="ref" reference="UlrRatliff"} that \(\widetilde{I}=I\); moreover, our Proposition [\[HS-poly\]](#HS-poly){reference-type="ref" reference="HS-poly"} yields in particular \(H_{I}(n) = P_{I}(n)\) for \(n = 1, 2\). Any such \(I\) is therefore a counter-example to Mafi's claim. We shall establish the correct statement in Proposition [\[genItoh\]](#genItoh){reference-type="ref" reference="genItoh"}. First, we need a lemma.
## Negative answer to a question of Corso, Polini, and Rossi
In this last part, recall that the integral closure of an ideal \(I\) of a Noetherian ring is the set \(\overline{I}\) formed by the elements \(r\) satisfying an equation of the form \]r^m+a_1r^{m-1}+\ldots + a_{m-1}r+a_m=0, \quad a_i\in I^i, \quad i=1, \ldots, m.\(\) Clearly, \(\overline{I}\) is an ideal containing \(I\). The ideal \(I\) is integrally closed (or complete) if \(\overline{I}=I\). Moreover, \(I\) is *normal* if \(I^j\) is integrally closed for every \(j\geq 1\). Let \((R, \mathfrak{m})\) be a Gorenstein local ring (with positive dimension and infinite residue field), and let \(I\) be a normal \(\mathfrak{m}\)-primary ideal of \(R\). Then following problem appeared in: Does \({\rm e}_3(I)=0\) imply \({\rm r}(I)=2\) ? It has been also pointed out that an affirmative answer to this question is given in in the case \(I=\mathfrak{m}\) (recall that \(\mathfrak{m}\) may not be normal in general). Now note that if \({\rm dim}\,R=2\) then the condition \({\rm e}_3(I)=0\) holds trivially. So, the 2-dimensional case of the above question is: If \(I\) is a normal \(\mathfrak{m}\)-primary ideal of a 2-dimensional Gorenstein local ring, is it true that \({\rm r}(I)=2\) ? Below we answer this question negatively. In our example, the ideal \(I\) satisfies in addition the property \({\rm e}_2(I)=0\) (indeed, \(I\) is Ulrich and this vanishing was observed in the proof of Proposition [\[HS-poly\]](#HS-poly){reference-type="ref" reference="HS-poly"}). Thus we may also consider a version of the Corso-Polini-Rossi problem by adding the hypothesis \({\rm e}_2(I)\neq 0\), which we hope to pursue in a future work.
The first author was partially supported by the CNPq-Brazil grants 301029/2019-9 and 406377/2021-9. The second author was supported by a CAPES Doctoral Scholarship. | {'timestamp': '2022-12-19T02:10:50', 'yymm': '2212', 'arxiv_id': '2212.08428', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08428'} |
null | null |
null | null |
# Introduction
aerial vehicle (UAV) with some advantages, such as less susceptible to atmospheric interference, low flight altitude, high resolution and, low operating costs has been widely used in land surveying and mapping, ecological environment monitoring, resource survey and classification etc.. Compared with other aerial photography collection methods, the high resolution remote sensing images taken by UAV is more suitable to extract various important ground objects, e.g., road and house. In recent years, with the rapid development of deep learning, models based on convolutional neural networks have shown superior performance in some computer vision tasks such as image classification, detection, and segmentation, etc.. Compared with traditional machine learning algorithms with manual features extraction, deep learning can automatically extract features including color, texture, shape and spatial position relationship of the image. Fully Convolutional Network (FCN), as the first semantic segmentation of natural images only using convolutional operation, realized the pixel level classification of the images. Then, there are more semantic segmentation models significantly improved the segmentation performance. All of these algorithms predict the pixel level labels based on the semantic information represented by image pixels. Different from natural images, UAV remote sensing images are often very large in size, usually with a resolution of tens of thousands by tens of thousands. UAV remote sensing image segmentation is a challenge task due to the large variations in the size, shape, color and location of the ground object, as shown in Fig.1. Figs.[\[a1\]](#a1){reference-type="ref" reference="a1"} and [\[a2\]](#a2){reference-type="ref" reference="a2"}. show the different style buildings though they belong to the same ground object. In the other case, the surface feature elements of different ground objects may be similar. As shown in Figs.[\[b1\]](#b1){reference-type="ref" reference="b1"} and [\[b2\]](#b2){reference-type="ref" reference="b2"}, the top of a concrete building is very similar to the surface of a concrete road, leading to difficulty in feature extraction. A lot of greenery on both sides or in the center of the road, and vehicles on the road cover the road, which would affect the road surface segmentation results, as shown in Figs.[\[c1\]](#c1){reference-type="ref" reference="c1"} and [\[c2\]](#c2){reference-type="ref" reference="c2"}. In addition, buildings with different scales are likely to affect the performance of the model, which requires the model with a strong ability to extract multi-scale context features. To address these issues, a new deep neural network structure, called **L**ightweight and **E**fficient **Net**work implemented Using **D**ual **C**ontext modules (LEDCNet), is proposed for the extraction of ground objects from UAV remote sensing images. It is well known that DenseNet has a strong feature extraction ability, but its parameters and calculations are large. Moreover, Liu et al. developed a convolutional neural network (CNN) called ConvNeXt, which obtained a superior performance beyond Swin Transformer by improving ResNet. Inspired by this, a lightweight high-performance convolutional neural network, named **L**ightweight **D**ensely **C**onnected **Net**work (LDCNet), is proposed as the backbone structure of the proposed LEDCNet for semantic segmentation of remote sensing images. The proposed LEDCNet is an encoder-decoder structure which employ the LDCNet as the encoder with the excellent segmentation results. And in the decoder, there are dual context feature extraction modules, combine with the Atrous Spatial Pyramid Pooling (ASPP) module and the Object Context Representation (OCR) module. Between ASPP and OCR, the Feature Pyramid Network (FPN) module is used to fuse multi-scale features after ASPP operation. The main contributions of this paper are as follows.
1. A new network based on encoder-decoder structure for the extraction of feature elements from UAV remote sensing images is proposed. A new LDCNet that is proposed in this paper is employed as a lightweight encoder to reduce model's parameters and accelerate the computational speed. And a dual multi-scale contextual feature extraction module combines with the ASPP and OCR is used to as an effective decoder to fuse and refine the target objects and boundary features.
2. A lightweight and high-performance backbone network called LDCNet, which uses the design ideas of ConvNeXt and DenseNet is developed. The LDCNet is an efficient light-weight network for features extraction of UAV remote sensing images with fewer parameters.
3. Extensive experiments are carried out to verify the effectiveness and feasibility of the proposed method on our private database and two publicly datasets, i.e., LoveDA and CITY-OSM. Several most popular backbone networks and semantic segmentation algorithms are used to compared with the proposed method on the semantic segmentation of UAV remote sensing images. The experimental results show that the proposed method can effectively extract the road and building with higher accuracy than the other compared networks.
# Related Works
## Semantic Segmentation in Remote Sensing
Fully convolutional networks (FCN), as the first semantic segmentation network based on deep learning, can accept the input size of any size. To segment the remote sensing images with more complex, there are some improved FCN-based networks proposed to enhance the segmentation performance. Maggiori et al. designed a dual-scale neuron module based on FCN for semantic segmentation of remote sensing images, which balances the accuracy of recognition and localization. Liu et al. proposed an improved FCN to high-resolution remote sensing image segmentation. However, FCN has limited ability to extract objects of very small size or very large size. Fu et al. adopted dilated Atrous convolution to optimize the FCN model and used conditional random field (CRF) to post-process preliminary segmentation results, which result in a significant improvement over previous networks. Atrous convolution can increase the receptive field of the convolution while maintaining the spatial resolution of the feature map. U-Net is another popular semantic segmentation network, which was first used in medical image segmentation. Li et al. proposed a network to segment the land and sea of high-resolution remote sensing images based on U-Net. Cheng et al. developed a network called HDCUNet combining U-Net with Hybrid Dilated Convolution (HDC) for fast and accurate extraction of coastal farming areas. It avoids meshing and increases the receptive field. Wang et al. designed a U-Net with two decoders and introduced spatial attention and channel attention. In addition, some researches improved SegNet for semantic segmentation of remote sensing images. Weng et al. proposed an SR-SegNet using a separable residual algorithm for water extraction from remote sensing images. Kniaz et al. developed a network called GeoGAN based on the Generative Adversarial Network (GAN) to extract waters in different seasons. Recently, some Transformer-based or the combination of Transformer and CNN models promoted the semantic segmentation performance with the Transformer' advantage of global receptive field. However, the large parameters of Transformer affects the calculation speed of these proposed models.
## Lightweight Network
In order to deploy the network models on devices with limited resources, it is necessary to design lightweight and efficient networks. The MobileNet series is a classic lightweight network that applies depthwise separable convolutions, inverted residuals, linear bottlenecks, and lightweight channel attention modules. It can achieve higher accuracy with fewer parameters and lower calculation costs. SqueezeNet, another lightweight network, used the Fire module for parameter compression. EffectionNet balanced the three dimensions of depth, width, and resolution well, and scales these three dimensions uniformly through a set of fixed scaling factors. GhostNet obtains redundant information by designing cost-efficient, which ensures the model can fully understand the input data. MicroNet used Micro-Factorized convolution and Dynamic Shift-Max to reduce the amount of calculation and improve network performance.
## Context Feature Extraction
By fully considering the context information can significantly boost the semantic segmentation performance and solve the problem of receptive field scale. To aggregate multi-scale contextual features, PSPNet used a pyramid pooling module, which connects four global pooling layers of different sizes in parallel, pools the original feature maps to generate feature maps of different levels, and restores them to the original size after convolution and upsampling. DeepLabV3 introduced the Atrous spatial pyramid pooling (ASPP) module, which utilized different hole rates to construct convolution kernels of different receptive fields to obtain multi-scale context information. GCNet employed a self-attention mechanism to obtain global context information. GCN enlarged the receptive field by increasing the size of the convolution kernel. The convolution kernel up to 15\(\times\)`<!-- -->`{=html}15 in size in GCN proved that the large convolution kernel has a stronger ability to extract context features.
# Proposed Method
## Overall structure of LEDCNet
The proposed lightweight and efficient semantic segmentation network, referred to as LEDCNet, is shown in Fig.[\[LEDCNet\]](#LEDCNet){reference-type="ref" reference="LEDCNet"}. The proposed LEDCNet is an encoder-decoder structure. In the encoder part, LEDCNet takes our proposed LDCNet as a backbone, which can reduce model parameters and accelerate the computational speed. First, an image of size 512\(\times\)`<!-- -->`{=html}512 is input into the backbone network LDCNet. And after feature extraction operation by LDCNet, four different levels of feature maps are output. The first three feature maps are respectively input into the Atrous spatial pyramid pooling module (ASPP) to extract multi-scale context information. The three outputs are fed into the corresponding 1\(\times\)`<!-- -->`{=html}1 convolution operation in the Feature Pyramid Network (FPN) and upsampled to the same size. Four feature maps of the same size are then concatenated as the input of Object-Contextual Representation (OCR) module. After the feature map is processed by OCR, it will produce a rough segmentation result and a feature map with object context information. This rough segmentation result is taken as one of the outputs of the network, and the feature maps are fed into the next step for further processing. Then, the feature map is processed by a refinement classification head to refine the segmentation edges. It consists of a 3\(\times\)`<!-- -->`{=html}3 depthwise separable convolution and a 1\(\times\)`<!-- -->`{=html}1 convolution. Finally, the feature map is upsampled to the size of the input image.
## LDCNet
Inspired by ConvNeXt and DenseNet, this paper develops a backbone network, called LDCNet. The structure and densely connected blocks are shown in Fig.[\[Bottleneck\]](#Bottleneck){reference-type="ref" reference="Bottleneck"}. ConvNeXt improved the classical ResNet by introducing some of the latest ideas and technologies of Transformer network to enhance the performance of CNN. Macro design, reference of ResNeXt's design ideas, inverted bottleneck layer, large convolution kernel and micro design of various layers are the five main optimization design of ConvNeXt. In the macro design of ConvNeXt, the stacking ratio of multi-stage blocks is 1:1:3:1. The number of blocks stacked in the third stage is larger. This improves the model accuracy of ConvNeXt. Following the design, we set the stacking ratio of blocks in each stage of LDCNet to 1:1:3:1. The specific layers of each stage are 2, 2, 6 and 2 respectively. The structure of the proposed LDCNet is shown in Fig.[\[LDCNet\]](#LDCNet){reference-type="ref" reference="LDCNet"}. ConvNeXt designs the effective inverted bottlenecks block with a 3×3 depthwise separable convolution, shown in Figs.[\[Bottleneck\]](#Bottleneck){reference-type="ref" reference="Bottleneck"}(a) and (b). This structure can partially reduce the parameter scale of the model while slightly improving the accuracy rate. Considering the multi-scale feature extraction ability of Inception block in GoogLeNet, a new bottleneck layer with two branches is proposed in this paper. One of the branches is a depthwise separable convolution with a 7\(\times\)`<!-- -->`{=html}7 convolution kernel, and the other is a depthwise separable convolution with a 3\(\times\)`<!-- -->`{=html}3 convolution kernel. After adding the output feature maps of two branches, and then concatenating with the input feature map, the output feature map of the proposed bottleneck layer is produced, as shown in Fig.[\[Bottleneck\]](#Bottleneck){reference-type="ref" reference="Bottleneck"}(c). In recent years, some new studies have stated the large convolution kernels are more efficient for enlarging the receptive field. In order to obtain a higher computational efficiency, a 7\(\times\)`<!-- -->`{=html}7 convolution kernel is used at the beginning of one branch in the proposed bottleneck layer. As we all know, Batch normalization (BN) is most popular optimization process in computer vision tasks by computing the mean and variance of each minibatch and pulls it back to a standard normal distribution with mean 0 and variance 1 to make neural network training faster and stabler However, BN also has may some drawbacks detrimental to the model performance. Transformers use a simpler layer normalization (LN), which is more common in natural language processing tasks. ConvNeXt replaces all BNs with LNs to improve the model performance. Since the features of our remote sensing images depend on the statistical parameters between different samples. it is inappropriate to replaces all BNs with LNs in the task of extracting ground objects from UAV remote sensing images. Therefore, we use an LN after a 3\(\times\)`<!-- -->`{=html}3 depthwise separable convolution in the bottleneck layer, and replaced the BN in the transition layer with LN.
## Atrous Space Pyramid Pooling (ASPP)
In order to obtain a large receptive field without losing spatial resolution and increasing computation, ASPP uses multiple parallel dilated convolutional layers with different sampling rates. The features which are extracted for each sampling rate are further processed in separate branches and fused to generate the final result. Atrous convolution can expand the receptive field of the convolution kernel without loss of resolution. Using ASPP can achieve multi-scale feature extraction through different receptive fields and upsampling. In a two-dimensional convolution, for each position i on the feature y of the convolution output and the corresponding convolution kernel w, for the input x, the calculation of the dilated convolution is as follows: \[\label{dilated_convolution} y[i] = \sum_{k}{x[i+r\cdot k]w[k]}\] where r is the hole rate, which represents the sampling step size of the convolution kernel on the input x of the convolution operation. k represents the position of the convolution kernel parameter. If the convolution kernel size is 3, then k=0,1,2. The receptive field of the convolution can be adaptively modified by changing the value of the dilation rate r.
## Object Context Representation (OCR)
OCR considers the relationship between a pixel and its context pixels, and aggregates similar context pixel representations with higher weights. Unlike relational context method, OCR constructs contextual pixels into object regions and exploits the relationship between pixels and object regions. The OCR module consists of three steps: soft object region extraction, object region representation computation, and object-contextual representation computation for each position. It is mainly based on the scaled dot-product self-attention of the encoder-decoder structure. The input to attention consists of: a set of \(N_q\) queries \(Q\in{\mathbb{R}^{d\times{N_q}}}\), a set of \(N_{kv}\) keys \(K\in{\mathbb{R}^{d\times{N_{kv}}}}\), and a set of \(N_{kv}\) values \(V\in{\mathbb{R}^{d\times{N_{kv}}}}\). The attention weight \(a_j\) is computed as the Softmax normalization of the dot product between query \(q_i\) and key \(k_j\): \[\label{Zi} Z_i = \sum_{j=1}^{N_{kv}}{e^{\frac{1}{\sqrt{d}}q_i^\top k_j}}\] \[\label{aij} a_{ij} = \frac{e^{\frac{1}{\sqrt{d}}q_i^\top k_j}}{Z_i}\] The attention output for each query \(q_i\) is the aggregation of values weighted by attention weights: \[\label{attn} Attn(q_i,K,V)=\sum_{j=1}^{N_{kv}}{a_{ij}Vj}\] For object context representation, the calculation formula of the relationship between each pixel and each object area is as follows: \[\label{OCR} w_ik = \frac{e^{\kappa{(x_i,f_k)}}}{\sum_{j=1}^{K}{e^{\kappa{(x_i,f_j)}}}}\] Among them, \(\kappa{(x,f)}=\phi{(x)}^\top\psi{(f)}\) is the denormalized relationship function, \(\phi{(\cdot)}\) and \(\psi{(\cdot)}\) are two transformation functions implemented by 1\(\times\)`<!-- -->`{=html}1Conv \(\rightarrow\) BN \(\rightarrow\) ReLU.
# Experiment
## Ablation study
To evaluate the contribution of each component of the proposed LEDCNet the ablation experiments are conducted using LoveDA dataset and CITY_OSM dataset, as shown in Table [\[ablation_study\]](#ablation_study){reference-type="ref" reference="ablation_study"}. After adding the ASPP module to the baseline model, mIoU on the two public datasets increased by 0.55% and 0.42%, respectively. After adding the OCR module to the baseline model, mIoU on the two public datasets increased by 1.55% and 0.88%, respectively. After the proposed model aggregates these two modules, mIoU on two public datasets improves by 2.03% and 0.69% compared to the baseline model, respectively.
# Conclusion
In this paper, a lightweight and efficient semantic segmentation network based on encoder-decoder structure was developed for ground objects extraction from UAV remote sensing images. The encoder used a new lightweight and high-performance backbone network proposed in this paper to accelerate model calculation and reduce model parameters. The decoder employed a dual-context module combined with ASPP and OCR to effectively capture multi-scale contextual information. Extensive experiments demonstrate that the proposed model achieved high performance on both private and two public datasets, with fewer parameters and operations. In the future, we hope that our LEDCNet can be expanded to more semantic segmentation fields. We will continue to improve this model to make it more efficient. | {'timestamp': '2022-12-29T02:08:44', 'yymm': '2212', 'arxiv_id': '2212.08490', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08490'} |
null | null |
# Introduction
**Context and Objective**. Learning analytics has been fuelled through increased access to *data* and analytical studies development. These studies have often focused on factors that influence *student success* to bring insights into the learning process and establish recommendations. Student success has been proven to be influenced by many factors, such as engagement, regularity, critical thinking, metacognition, and socio-emotional well-being. This knowledge has been capitalized to build *student models* that can predict course success. Their predictions are useful for several means, e.g., personalized intervention, adaptive content provisioning, and learning process understanding. Notable instructional strategies benefiting from this support include *blended learning* (e.g., flipped courses ) and *online learning* (e.g., MOOCs ), where instructors can only partially observe student learning. However, these models also introduce risks for students and instructors. Their users may not be used to reasoning about the *forms of model uncertainty* and thus *misunderstand or mistrust* its predictions. Trust is essential for the acceptance of these models in education . Hence, our goal is to investigate *weak spots* of student success models to understand the circumstances under which these *models should (not) be trusted* by the users. **Open Problem**. Humans are often convinced that predictions should be trusted, as *they result from an accurate machine-learning model*. By accurate we mean that if the model predicted well in the past it should be trusted to do so in the future. However, there is less evidence on the *weak spots* these models can develop, i.e., examples hard to accurately classify due to low representation. Our study focuses on a weak spot never explored in student success models: *unknown unknowns*. They denote students for which the model is highly confident of its prediction although it is wrong[^1]. Syntactically, the second \"unknowns\" represents model knowledge (does not know correct labels), and the first \"unknown\" represents model awareness (not aware of being wrong; very confident of the wrong predictions). Unknown unknowns might result from unmodeled data biases (e.g., under-represented groups), data distribution shifts (e.g., course changes), or factors hidden to the model (e.g., students' prior knowledge). Efforts to analyze unknown unknowns in fields other than education have uncovered *unintended machine behaviours*. **Motivating Example**. The toy example in Figure [\[fig:toy-example\]](#fig:toy-example){reference-type="ref" reference="fig:toy-example"} shows the *unknown unknowns* problem. *Student A* is a 1st year university student. The student failed a Math course taught in a flipped format. During the course, they had access to a platform for completing pre-class activities, e.g., watching videos and doing exercises. We assume that a machine-learning model flagging students needing intervention (based on educationally relevant features) was used in this course. This model was built upon historical pre-class behavioral data from the same course and demonstrated a high accuracy. However, this data included a majority of students with high prior knowledge in Math. Due to this, the model found that a low time spent on videos (a feature) was predictive of success (e.g., content was often confirmatory knowledge). *Student A* did not spend much time on videos because they were struggling. The model, unaware of students' prior knowledge, indicated to the instructor *Student A* was very likely to pass the course. Due to the limited teaching resources, the team decided to not intervene on *Student A*, who ended up failing the course. In this case, students' prior knowledge led to unknown unknowns. **Our Contribution**. In this paper, we unveil the need of characterizing unknown unknowns in student success prediction to gain a deeper understanding of when the model fails. Bridging education, trust, and machine learning concepts, we introduce a computational framework for cost-efficient identification and characterization of unknown unknowns. We then assess its effectiveness, informativeness, and cost-efficiency on two state-of-the-art student success prediction models, based on data from six courses (three flipped classroom courses and three MOOCs). With our experiment, we investigate (i) *whether unknown unknowns exist in student success prediction and how they vary in number and type across flipped courses and MOOCs*, (i) *whether we can characterize unknown unknowns in student success prediction models*, and (iii) *the extent to which providing instructors with information about unknown unknowns impact their perception of the student success model*. Our results show that unknown unknowns are a serious issue for student success models and highlight the value of our framework for the discovery of unknown unknowns across experimental conditions.
# Methodology
In this section, we describe the educational scenario (learning activities, tracked data, behavioral indicators of interest) as well as the student success models (model creation, evaluation, students grouping, group characterization, instructors' opinions) leveraged for unknown unknown analysis. Figure [\[fig:methodology\]](#fig:methodology){reference-type="ref" reference="fig:methodology"} provides an overview of the steps of our framework.
## Learning Data Collection-Step 1 / 6
While the characterization of unknown unknowns may be important for any educational scenario supported by student success models, our study focused on two teaching strategies applied to courses within a European university: flipped classroom courses and MOOCs. In the following, we explain the teaching scenario for both course types in more detail.
**Learning through Flipped Classroom Courses**. We based our analysis on three semester-long university courses with a weekly schedule including both lectures and sessions of recitation or exercises. These courses followed the learning science principles detailed in. All courses illustrated in Table [\[tab:courses\]](#tab:courses){reference-type="ref" reference="tab:courses"} are compulsory for Computer Science and Communication Systems Bachelor's degrees at said European university. In class, activities included quizzes, short problem-solving exercises, and structured proof-type problems. In addition, students were expected to spend few hours per week on individual study as a preparation for class (referred to as pre-class activities). Instructions regarding the preparatory work were sent to students (a list of sections from a MOOC with video lectures and online quizzes) one week in advance. Quizzes, usually including multiple-choice questions, enabled students to self-assess their learning. The MOOC platform collected data about student pre-class activities for all three flipped courses. Log entries were structured to report the user, activity, and timestamp (e.g., user: 10, activity: play video 32, timestamp: 05-03-2018 12:06:01). These entries were complemented with demographic attributes about students, including their gender, geographic provenience, and high school diploma. No data about in-class activities was recorded. Student achievement was measured by their grade in the final exam (grade between 1 and 6, with 4 or higher indicating a passing grade). This study was approved by the university's ethics committee (HREC 058-2020/10.09.2020, 096-2020/09.04.2020). **Learning through Massive Open Online Courses**. We complemented our analysis with three MOOCs taught by three different instructors of the same European university in Coursera (Table [\[tab:courses\]](#tab:courses){reference-type="ref" reference="tab:courses"}, last three rows). These courses were accessible to all people worldwide. Lectures were released on a weekly basis and students were expected to spend some hours per week to complete them. Each week, courses included short video lectures (10-15 minutes) introducing key concepts and quizzes for self-assessment. Students were asked weekly to complete a graded assignment, their score was used by the instructor to judge student achievement. The final course grade ranged between 0 and 100 (with at least 60 points to pass) and was calculated based on a weighted sum of the weekly assignment scores plus a final exam. We collected over 145,640 log entries, including the user, the activity, and the timestamp, reported in the same format as for flipped course logs. Students' gender and geographic provenience were attached to the log data where students provided them voluntarily. Collected data included all the activities performed by the student in the platform, covering a large part of their learning (except, for instance, offline video watching). In the MOOC setting, our study focused on a scenario where student success models were built based on the entire activity data in the platform (thus not visible to the instructor). Student achievement was measured by using the final course grade.
## Behavioral Indicators Extraction-Step 2 / 6 {#sec:features-descr}
Prior work has found significant association with academic achievement for self-regulated learning (SRL) aspects: effort regulation (persistence in learning), time management (ability to plan study time), metacognition (awareness and control of thoughts), critical thinking (ability to carefully examine material), and help-seeking (obtaining assistance if needed). A variety of learning indicator sets have been proposed in the literature accordingly. Their power for modeling learning success has been compared by, which identified the most important ones in both flipped courses and MOOCs. Empirical evidence of their importance across MOOCs was also provided by. Both our logging policy and experimental scenario were similar to those adopted in prior work. Thus, we built our models on top of the indicators proven to be important to them (see Table 3 in for a full list and description). The granularity and comprehensiveness of the collected log data allowed the studies mentioned above (and our) to consider effort regulation (effort), time management (regularity and proactivity), and metacognition (control) dimensions. Critical thinking and help-seeking could not be measured. The considered learning indicators covered the following dimensions. **Control** (22 indicators per student per week) models in-video and cross-video behavior as a proxy of student ability to control the cognitive load of video lectures through course weeks (metacognition). For instance, in-video flow, manageable through the platform functionalities (e.g., pause button), could include regular pauses to segment learning. Among others, this feature set consists of the proportion of videos watched, re-watched, or interrupted. **Effort** (13 indicators per student per week) aims at monitoring how much and how frequently students engage with the course content (both videos and quizzes), proven to be fundamental for learning success. These features included indicators such as the total number of student clicks on weekends and on weekdays, and the total number of sessions. **Proactivity** (7 indicators per student per week) attempts to measure the extent to which students are on time or ahead of the schedule, proven to be prediction of performance especially in MOOCs. These features are related to completion of videos and quizzes, according to the week of the course they are schedule on. Example features included the number of scheduled videos watched for that week and the number of quizzes passed on the first try. **Regularity** (3 indicators per student per week) is also associated with time management. It estimates the intra-week and intra-day time management patterns (i.e., capturing whether a student regularly engages on specific weekdays or day times), which have been proved to be predictive of student success in MOOCs and flipped classrooms. Since indicator scores vary in their range, we performed a min-max normalization per feature across all students and weeks for that feature. Though the selected indicators fall into four dimensions and include several measures per dimension, we acknowledge that other relevant dimensions and measures could be beneficial for student success modeling. Our study can be easily run on a different (and larger) set of behavioral indicators.
## Student Success Modelling-Step 3 / 6
A wide range of student success models have been proposed so far in the literature. To align with prior work, since our study does not aim to propose a novel model, we considered two models reporting a high accuracy while providing a certain level of interpretability. Random Forest (RF) classifiers have achieved this in both flipped and MOOC contexts, when fed with behavioral features . Recent neural network classifiers based on BiLSTMs including attention layers, sigmoid activation, and a cross entropy loss function, have resulted in higher accuracy and good interpretability as well. Again, we based our decision on the similarity of the underlying context and logging system. We acknowledge that other models, e.g., Linear Regression and Support Vector Machines, have been used in prior works (e.g., ), but we left their analysis as a future work, using RFs as a representative of this class of models. For each course and model, we applied a nested student-stratified (i.e., dividing the folds by students) 10-fold cross validation. The same folds were used for all experiments across models, and we optimized the hyper-parameters via grid search. In each iteration, we ran an inner student-stratified 10-fold cross-validation on the training set in that iteration, and selected the combination of hyper-parameter values yielding the highest accuracy on the inner cross-validation. Finally, 200 models per course were obtained (\(2\) architectures \(\times \; 10\) outer folds \(\times \; 10\) inner folds). For each model, the balanced accuracy was computed on the training and validation to check validity and on the test set to assess model generalizability. In other words, we pulled a random set of students out of the course and demonstrate generalizability in the same context. For conciseness, we do not report other similar evaluation metrics, such as kappa score and area under the ROC curve. Figure [\[fig:ssm-bacs\]](#fig:ssm-bacs){reference-type="ref" reference="fig:ssm-bacs"} collects the averaged balanced accuracy on the three sets for each course and architecture combination. Models predicting on MOOCs provided a higher balanced accuracy than those on flipped courses. This might be justified by the fact that online pre-class activities in flipped courses are a tiny portion of the learning process. RF and BiLSTM often performed equally well, with the former reporting a slightly higher balanced accuracy on MOOCs. Each model was able to predict the probability \(p\) that the student would fail the course, as an output. Specifically, \(p\) could range between \(0\) and \(1\), with values closer to \(1\) meaning that the student would probably fail the course. We finally used a decision threshold equal to \(0.5\) to obtain the final predicted label: \(\Tilde{y}=0\) in case \(p < 0.50\) (predicted pass); \(\Tilde{y}=1\) in case \(p \geq 0.50\) (predicted failure). This threshold is a common value in general machine learning and prior education-related works. The model confidence was obtained by measuring the extent to which the probability \(p\) was close to the decision threshold. Numerically, the model confidence \(c\) for a predicted probability \(p\) is equal to \(| p-0.50 |\), ranging between \(0\) and \(0.50\). Higher values mean that the model is more confident about the prediction.
## Prediction Student Grouping-Step 4 / 6 {#sec:grouping}
Obtained a trained student success model, computing an importance score for each indicator and link it to successful patterns has been a common practice concerning both RFs and linear regression models. Recent work has also extracted these scores through explainability methods. These works however assume that the model accurately reflects the relationships between learning and students' success, which might not be true for a (non-negligible) portion of students. Since no student should be negatively impacted by a model, such portion of students must not be ignored. To illustrate an emerging issue, see Figure [\[fig:ssm-scores\]](#fig:ssm-scores){reference-type="ref" reference="fig:ssm-scores"}. Each plot collects the predicted probability distribution for students within training, validation, and test sets, considering \(200\) pass and \(200\) failing students selected randomly. On flipped courses, BiLSTM models were often very confident about their predictions, being skewed at the two horizontal extremes. However, looking at Figure [\[fig:ssm-bacs\]](#fig:ssm-bacs){reference-type="ref" reference="fig:ssm-bacs"}, their balanced accuracy on flipped courses was not high. Concerning MOOCs (bottom row), the same models were again very confident, but in this case their accuracy was very high. Interestingly, on flipped courses, RF showed a high uncertainty in predictions for unseen validation and test students. These results raise concerns on the possibility the model might be very confident but actually wrong and, in turn, open to severe consequences (e.g., deviating humans' understanding of the learning process) in case models are used in the real world. Given the above observations, investigating the interplay between the predicted label correctness and the model confidence becomes essential. Let assume that \(\delta \in (0, 0.50)\) is the hypothetical trust level the user has while using model predictions. Predictions with \(c \geq \delta\) would be considered by the user with a higher trust. Conversely, the final users would consider those having \(c < \delta\) with a lower trust. Given a student known to have a true label \(y\) and a model that predicts a label \(\Tilde{y}\) with a confidence \(c\), the student can be assigned to a group \(g \in \{0, 1, 2\}\) as follows:
- **Known Knowns** (\(g = 0\): \(c \geq \delta\) and \(\Tilde{y} = y\)) include students for whom the model is very confident and the corresponding predictions correctly reflect the true success or failure. These are the ideal cases where the model is correctly optimistic: the model reports high confidence and correct label for the examples in these region(s).
- **Known Unknowns** (\(g = 1\): \(c < \delta\) and any \(\Tilde{y}\)) include students for whom we are aware the model is not sure and the corresponding predictions would require extra care before being used to take action in the real world. This concept models cases for which errors are expected based on the probability estimates of the classification.
- **Unknown Unknowns** (\(g = 2\): \(c \geq \delta\) and \(\Tilde{y} \neq y\)) include students for whom the model is confident but their predictions are actually wrong. Intuitively, these are examples distant from the decision boundary yet labeled incorrectly. These examples may be rare to be detected, but their non-negligible prevalence makes them a risk.
For convenience, we used a trust level \(\delta = 0.25\) and left analyses under other trust levels as a future work[^2].
## Prediction Group Characterization-Step 5 / 6 {#sec:explanatory}
To reduce unknown risks, someone could think that being able to classify a future student as unknown unknown would be the solution. This might mean to train a second model to predict whether a student is an unknown unknown and do something with it. This solution unfortunately has several weaknesses, e.g., complexity, more biases, and lack of data. If another model can detect unknowns unknowns, that model could be directly used to improve the original student success model. Such weak spot however comes due to knowledge that model (and the second model) was missing. Letting stakeholders be aware of unknown unknown cases would conversely provide a better support. Typical archetypes of unknown unknowns could be provided to the instructors before or while using the model. Indeed, a human might think about the information the model was missing and propose to fix data collection and processing accordingly. In this scenario, a core question is the choice of the variables used to explain the unknown unknown cases. Two complementary approaches exist for choosing them. In confirmatory research, the potential impact of different variables is hypothesized a-priori, based on existing theories. An exploration-driven approach is used when there is a lack of theories or while generalizing across domains. Given the only recently awareness of unknown unknowns, exploration research could produce new hypotheses to be evaluated later by experts (e.g., instructors) in their courses. To characterize unknown unknowns from the model perspective, we studied the relationship between (i) model confidence and predicted label correctness according to the prediction-based student groups we identified in the previous step, and (ii) a range of variables the model already knows (i.e., the behavioral indicators leveraged for training). To characterize unknown unknowns, as a source of context, we also considered some variables pertaining to demographics. Not included in the model training (also for ethical reasons), demographics might be available and used as a set of contextual variables. Formally, the *independent variables* for our characterization were represented, for each student, with a vector \(v\) including the \(45\) behavioral indicators (averaged across course weeks) and the \(2\) demographic attributes (gender and provenience). The *dependent variable* was the group label \(g\) computed in the previous section. Intuitively, we were interested in identifying variables that have a statistically significant relationship with the dependent variable. A multiple regression model was used to analyze these relationships, fitting it with the vectors \(v\) as an input and the corresponding group labels \(g\) as an output. Formally \(y = \epsilon + \gamma_0 + \sum_{j} \gamma_{j} \cdot v_{j}\), where \(y\) is the dependent variable, and \(\gamma_j\) and \(v_{j}\) respectively represent the regression coefficient and the value of the \(j\)-th variable. The coefficients \(\gamma \in \Gamma\) indicate us the extent to which a change of a given variable, holding all the others constant, leads to a change in the group. Being a multiple regression, our analysis tested the null hypothesis that all coefficients \(\gamma \in \Gamma\) are zero, versus the alternative that at least one coefficient \(\gamma_j\) is nonzero. We conjecture that variables with a nonzero statistically significant coefficient might be deemed as important to explain the unknown unknown membership.
## Instructors' Opinion Collection-Step 6 / 6 {#sec:quest}
Finally, we were interested in understanding how the instructors' perception of student success model would change when they are made aware of unknown unknowns. To this end, we designed a questionnaire[^3], following prior work on trust in artificial intelligence for education, including two main sections (see Table [2](#tab:questionnaire){reference-type="ref" reference="tab:questionnaire"}): one on demographic information, whereas the other was focused on a use case on unknown unknowns in student success modelling. In the first part, we were interested in knowing who the experts were, including the type of their organization (e.g., academia, industry), their role in the organization (e.g., full professor, researcher), the continent and country they are based in, the age group (in a specific range), and the gender identity. In addition, we asked participants how many courses they acted as an assistant (e.g., tutor) and as an instructor (e.g., full professor, associate professor) in.
In the second part, we described a use case with student success predictions presented to the instructor while investigating which students would require assistance. This use case was accompanied by a provocative user interface which would require them to think about the influence of student success models on the instructor[^4]. First, we asked experts to rate their perceived confidence, assistance, success, trust, transparency, humanity with respect to this case. This decision let us understand the original perception of instructors towards student success models in general to better contextualize our results. We then explained the concept of unknown unknowns and envisioned a simple addition to the user interface, indicating students at risk of being unknown unknowns via alerts and explanations. This information would trigger instructors to question the provided prediction and investigate the corresponding students' behavior more thoroughly. We then asked the extent to which their perception of the student success model changed relatively to their original perception (e.g., more or less trust). In both series of questions, each participant was allowed to select among four possible answers, from \"strongly disagree\" to \"strongly agree\". No neutral question was introduced; hence, answers should be interpreted as what the instructors would answer in case they were forced to make a decision. As a final field, we asked if they identified any (dis)advantages raised by being aware of unknown unknowns. Following 's protocol, we e-mailed the questionnaire to experts with a paper accepted in a top conference in education in 2021 (AIED, EAAI, EC-TEL, EDM, ICALT, ITS, LAK, L\@S). Out of 1,721 people, 112 (6,51%) completed the questionnaire. This choice was made to include people who are both educators and experts in the field. Though we acknowledge that future work will need to focus also on the perception of a generic instructor, unknown unknowns analysis is still at early stages and our feedback can make more mature our understanding before involving them.
# Experimental Results
Although our methodology can be used to analyze several perspectives, we focused on the following research questions:
- Do unknown unknowns exist in success prediction? How do they vary across flipped courses and MOOCs?
- Can we characterize common unknown unknowns in the considered student success prediction models?
- How does providing instructors with signals about unknown unknowns impact their perception of the model?
## RQ1: (Un)known Unknown Prevalence
In a first analysis, we were interested in understanding whether (un)known unknowns exist in the developed student success models and in investigating whether any difference between courses of the same type and across courses of a different type arises. To this end, Figure [\[fig:groups\]](#fig:groups){reference-type="ref" reference="fig:groups"} depicts the average percentage of students belonging to these two groups across the considered settings. As expected from the model performance in Figure [\[fig:ssm-bacs\]](#fig:ssm-bacs){reference-type="ref" reference="fig:ssm-bacs"}, more (un)known unknowns were present in flipped courses than MOOCs. We considered courses of different type separately, given that models showed a substantially different balanced accuracy across course types. We discuss the patterns in detail in what follows. Concerning flipped courses, it can be observed that RF and BiLSTM models differently suffer from (un)known unknowns. RF models were more conservative and reported a higher percentage of known unknowns than unknown unknowns. It follows that this type of model will require the instructor more consciously inspect students the model was not confident on. On the other hand, BiLSTM models were in general more confident, but this confidence was often on the wrong prediction. It was indeed evident that BiLSTM models led to more unknown unknowns than known unknowns. In general, unknown unknowns were equally distributed between wrong failing and wrong pass predictions. Estimates were similar across the flipped courses, showing that the model impacted more than the course itself. Despite of the lower absolute values, RF and BiLSTM model patterns were confirmed on students attending MOOCs. It should however be noted that MOOCs included more students and, as an example, \(2\%\) of unknown unknowns in \(\texttt{OC-DSP}\) would involve \(300+\) students. Analogously, \(4\%\) of unknown unknowns in \(\texttt{OC-HWTS}\) would involve \(100+\) students.
## RQ2: Unknown Unknowns Characterization
In a second analysis, we aimed to characterize unknown unknowns in success prediction models. To this end, we ran the explanatory framework described in Section [2.5](#sec:explanatory){reference-type="ref" reference="sec:explanatory"} on top of each model and course combination. Figure [\[fig:la-r2scores\]](#fig:la-r2scores){reference-type="ref" reference="fig:la-r2scores"} collects the averaged R2 score ranging in \([0, 1]\), i.e., the proportion of the variance in the dependent variable predictable from the independent variable(s), under each setting. It can be observed that such variance was more predictable on RF models than BiLSTM models and on MOOCs then flipped courses. Interestingly, on `OC-FP`, none of the models was able to explain reasonably well the variance in the dependent variable. While regression coefficients might give as an initial indication of the indicators more prominent among unknown unknowns, the (reasonably low) R2 score confirmed our hypothesis that a supplementary model for unknown unknowns detection would also not be the most viable solution. Figure [\[fig:la-coefflipped\]](#fig:la-coefflipped){reference-type="ref" reference="fig:la-coefflipped"} and [\[fig:la-coefmooc\]](#fig:la-coefmooc){reference-type="ref" reference="fig:la-coefmooc"} summarize the indicators having the highest coefficient for predicting unknown unknown's membership. For conciseness, we limited our discussion in the paper to the indicators whose coefficient had on average a value higher than \(1\) across courses of the same type. Concerning flipped courses, ten behavioral indicators were found important to explain unknown unknowns. Indicators pertaining to the control dimension were more present than those related to engagement and proactivity. For instance, the frequency of play (F10), stop (F11), and speed events (F15), but also the total number of video (F29) and problem clicks (F30), showed a high positive weight for being unknown unknown. In our data set, we observed that the majority of students who passed indeed tended to perform these actions many times. We conjecture that the model was challenged to understand whether those actions were frequently performed because students were struggling or they were reflecting more on the concepts and fixing them in their mind (this is something the model was not aware of). Therefore, the model tended to classify as passing with a high confidence also those students who were struggling with the content. Other observations can be made for the frequency of stop events (F12) and the alignment with the schedule (F38), which corresponded to a very low negative weight. Being not aligned with the schedule would not always mean the student will likely fail the course. For instance, a student could have learnt consistently offline and model unawareness about this could lead to unknown unknowns. In MOOCs, three indicators, all pertaining to proactivity, were found relevant through our explanatory analysis. No overlap in the selected regression coefficients was observed between flipped courses and MOOCs. Content anticipation show often associated to a high regression coefficient. Indeed, for students eager to learn, anticipating some course content can be a positive attitude for passing. For other students, who were maybe just interested in seeing what would have come next, interacting with the content in advance with respect to the schedule might not be a signal for passing. Again, the model would need additional information, to be able to distinguish between cases like this one.
By looking at the two model types on the same course, it results that different behavioral indicators confound different models on `FC-LA`. Indeed, the BiLSTM and RF models agreed on F02, F10, F11, and F38 but strongly disagreed with the other six. The two models weighted the behavioral indicators more similarly on the other two flipped courses. Regression coefficients went into this line on the MOOCs as well. Interestingly, none of the considered demographic attributes were deemed as important. In addition to being positive from a fairness perspective, this finding emphasizes the need of additional contextual variables, often non recorded in the log data neither in the university's database.
## RQ3: Unknown Unknowns Perception by Instructors
Finally, we were interested in understanding the instructors' perception of student success models in view of unknown unknowns awareness. To this end, we distributed the questionnaire and collected answers from experts worldwide as detailed in Section [2.6](#sec:quest){reference-type="ref" reference="sec:quest"}. Figure [\[fig:teaching\]](#fig:teaching){reference-type="ref" reference="fig:teaching"} collects a summary of the results from the questionnaire's answers. In Figure [\[fig:first_demo\]](#fig:first_demo){reference-type="ref" reference="fig:first_demo"}, it can be observed that our sample mostly included participants working in an academic context (92%)[^5]. Participants were almost equally distributed across associate professors (23,2%), PhD students (25,9%), and researchers (19,6%). Our sample of participants was balanced also age-wise, despite of a slight predominance of people in the 31-40 group (35,7%). With respect to gender identity, we received more responses from man (58%)\(^5\). Given the focus of our study, we were also interested in collecting information regarding the teaching experience (Fig. [\[fig:second_demo\]](#fig:second_demo){reference-type="ref" reference="fig:second_demo"}). In our sample, we found that 19.6% of participants never acted as tutors or instructors. Conversely, 25% (46,4%) out of the participants have served in more than five courses as a teaching assistant (leading instructor). Other participants reported a teaching experience between two and five courses. Overall, our sample showed interesting patterns of diversity. In the second section of the questionnaire, we first investigated the instructors' perception about an example student success model (Q09). Figure [\[fig:res_1\]](#fig:res_1){reference-type="ref" reference="fig:res_1"} highlighted that a high percentage of participants raised concerns about the example student success model. Participants strongly disagreed or disagreed (sum of the red-like bar values) with respect to being confident with the model predictions (74%), trusting predictions while using them in the classroom (73,2%), feeling of being able to use model predictions successfully (73,2%), seeing the model predictions as transparent (81,25%), and their reliability with respect to a recommendation from a human colleague (67,86%). Overall, we observed a prevalence of strongly disagree or disagree answers, except for their feeling of assistance. It should be noted that our goal was merely to collect a baseline perception, and not to assess the perception of the instructors on student success models in general which, paired with the implications of our study, will lead to next steps in future work. In Figure [\[fig:res_2\]](#fig:res_2){reference-type="ref" reference="fig:res_2"}, we collected the instructors' perception after some signals and explanations about unknown unknowns were displayed into the user interface. To make the comparison possible, we directly asked participants to indicate the relative change in their perception across the six perspectives. To this end, we alternatively added the terms \"*less*\" and \"*more*\" to the same questions originally reported in Q09. We decided to alternate positive (more) and negative (less) statements to reduce potential noise and biases in the answers. Our results showed that a high percentage of participants felt more confident (76,79% of the participants strongly disagreed or disagreed with the negative statement). A participant for instance said that \"*It helps teachers understand the reasoning behind the prediction. It helps on increasing my confidence on the model.*\". Instructors felt more assisted (82,14% of the participants strongly agreed or agreed with the positive statement). For instance, a participant found that \"*it was possible to get more insight*\". Trustworthiness was positively impacted as well, with 78,57% of the participants (strongly) disagreeing with the negative statement. Instructors felt also more successful in using model predictions (60,71% of the participants strongly agreed or agreed with the positive statement). They also thought that model predictions were not less transparent, so this perspective was positively impacted (71,43% of the participants strongly disagreed or disagreed with the negative statement). However, instructors still did not truly rely on model predictions (61,61%) as much as human recommendation. A participant emphasized that \"*I would not fully trust the predictions but they could help me to identify which students that I should I pay attention to*\". To have a detailed picture, we then cross-referenced these values with the different demographic elements in Figure [\[fig:tutoring\]](#fig:tutoring){reference-type="ref" reference="fig:tutoring"} and [\[fig:instructor\]](#fig:instructor){reference-type="ref" reference="fig:instructor"}. Interestingly, a participant reported that \"*The major advantage with additional alerts/explanation is the added context. In absence of explanations, the model is a strict \"follow or not follow\" type, but with these additional alerts its helps instructors to assess/study the students learning behavior and will help drive usability of the model.*\"
# Conclusions and Implications
With our experiments, we showed that unknown unknowns exist and vary in number and type across courses (**RQ1**). We then characterized unknown unknowns under the specific use cases (**RQ2**). Finally, we found that making instructors aware of unknown unknowns had a positive impact (**RQ3**). Our findings led to multiple implications. **Scientific Implications**. The collected data is usually partial and does not provide the global picture of students' behavior, skills, and needs. Many other variables can be hardly collected for being included in the model reasoning. For instance, a student might work a lot offline and still get very good exam grades. Being the predictions dependent on the data, there is the high risk that at the moment no enough data for the model is collected. Our study therefore calls for a more extensive data collection aimed to bridge the gap between what the models knows and what should know. Our study has shown that, BiLSTM models were often very confident about their predictions, but completely wrong in several cases. When possible, model uncertainty would be preferred to avoid misleading instructors, as in RF models. Understanding what is the source of such model confusion and how this knowledge can be induced into the model is urging. Another implication is therefore the need of student success models that reduce unknown unknown cases. Notably, we proved that assessing model performance solely based on accuracy may introduce unknown risks. When a student success model is delivered, evidence on its unknown unknowns should be provided. This evidence could be both quantitative, by reporting for instance the percentage of unseen students resulting as (un)known unknowns, but also qualitative, by characterizing the cases where the model is less confident but incorrect. Instructors will be likely to use success models as a complementary support to their personal perception. Human-in-the-loop approaches can be used to let the instructor and the model help each other while identifying students that require assistance. Unknown unknowns represent examples the instructor should reflect on while using predictions. **Technological Implications**. Our study has proven that confidence levels are not enough to prevent undesired behavior like unknown unknowns. Signals of their presence, though helpful, would just be triggers for further analysis of certain students. Indeed, pass/fail predictions alone would not be enough and more insights about learner behavior will be needed. Behavioral patterns (e.g., late submissions) can then indicate how to counsel students. In our work, we adopted BiLSTM models, with sigmoid and cross entropy. that could tend to push predictions towards the two sides, 0 and 1. Our findings show that this practice creates more unknown unknowns (more risks). Future work should carefully consider this aspect while selecting model parameters to reduce unknown risks. Once unknown unknowns patterns are identified, it will be important to understand how information about them should be presented. Having signals and explanations has led in our study to an increase of trust in the model prediction. Rather than a signal, showing a null state when a prediction is potentially risky would be another solution. Instructors might feel that the model did not have enough data to learn the likely outcome for some of the students yet. Besides being used for learning understanding, student success models can fuel tools to recommend instructors student requiring assistance (especially in large classes). It could be also important to have grouped predictions, along with individual predictions, based on similarities. Reducing unknown risks will be even more important in these settings. **Social Implications**. In our results, we showed that raising awareness of unknown unknowns led to higher confidence, feeling of assistance, trust, and transparency. Being trustworthiness a complex challenge in this field, our study introduces another source for improvement, complementary for instance to explanations. Furthermore, such models could be harmful by creating false expectations. Being aware of unknown unknowns can help to prevent such situations. Finally, predicting whether a student is going to pass or fail a course is of practical utility if it is done early. Once all of the data is collected over the course end, the models' purpose might be just to detect at-risk students who might benefit from remedial sessions between the course end and the final exam. Nevertheless, we believe that the unknown unknowns might be even more evident in very early predictions, which we plan to investigate in future work. Scientific, technological, and social shifts in education often go hand-in-hand. Countering unknown unknown issues will be essential to further strengthen the reliability of emerging student success models in real-world education.
[^1]: On the contrary, *known unknowns* are students for which the model is unsure about the correct prediction, since it falls close to the decision boundary. This phenomenon has been studied in the active learning field, and a solution could be to solicit new labels from humans on uncertain examples.
[^2]: This threshold let us provide a more focused unknown unknowns comparison, but is not required by the final system. In any case, it might be adjusted according to the model (e.g., its accuracy). For convenience, our study considers the same threshold for both models to make the comparison fair.
[^3]: The full questionnaire is available at the following webpage: <https://shorturl.at/qvX47>.
[^4]: The user interface was on purpose limited to a rudimentary pass/fail setting, without any confidence level, to stimulate instructors' critical thinking and reduce the impact of other user interface elements on the unknown unknowns perception.
[^5]: This aspect should be taken into account while analyzing our findings. This imbalance however comes from the adopted protocol, which involved a large pool of educators yet experts in the field who have recently published in the major educational conferences. | {'timestamp': '2022-12-19T02:13:44', 'yymm': '2212', 'arxiv_id': '2212.08532', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08532'} |
# Introduction
1G in the 1980s to recently commercialized 5G, modern communications have experienced substantial technical revolutions in almost every decade. Meanwhile, tremendous new applications, such as massive machine-type connectivity with stringent latency and reliability constraints, spawn and exhibit significant differences from the classical throughput-oriented transmission. Correspondingly, it is highly demanded to re-design smarter and more efficient communications to satisfy the diversified quality of service (QoS) requirements whilst balancing the dilemma of finite communication resources and infrastructures. Basically, an intelligent communication system, which is a vital ingredient in both industry and academia endeavors, is perceived to not merely memorize data flows with rigid rules, but also understand, process, and express the underlying semantics. In that regard, apart from post-processing in downstream information carriers, compression and understanding from the very beginning would be efficient, which also emerges as a driving force to push the frontier of exploring "meanings" in a communication system. Building upon the classical information theory (CIT) as revealed by Shannon, modern transmission schemes provide cutting-edge protocols (e.g., CDMA, OFDM, MIMO, etc.) for reliable communication, securing bit-level transmission accuracy. In spite of its popularity, treating each bit with equal importance can be content-and intelligence-agnostic in real applications and inconsistent with human perception as well. Essentially, bit accuracy is not sufficient for "intelligent" and semantic transmission-intuitively, different expressions may exactly convey a similar meaning, while an expression with a high word error rate (WER) may still convey the semantics. Therefore, instead of fully transmitting every single bit, it will be desirable to develop a semantics-empowered paradigm that seeks to communicate the underlying meanings. Transmission with semantics can date back to as early as 1949, when Weaver first introduced this idea. Unfortunately, owing to the technical constraint and a more urgent need for Shannon's CIT, related researches entered a period of silence. As a matter of fact, it took the community several decades to find concrete clues to breakout the "Shannon's trap", since mastering semantics was widely beyond the scope of CIT. Benefiting from the latest advance in artificial intelligence (AI), the prerequisite techniques are becoming gradually available and ready for building a prototype model. As a piece of evidence, research on semantics-empowered communication (SemCom for short) is now experiencing an unprecedented growth. According to Google scholar, the number of papers titled "semantic" and "communication" in 2022 exceeds the summation of those in the previous two years. Notably, this surging trend is expected to continue in the near future since SemCom is a promising enabler for 6G and beyond applications (e.g., Metaverse applications).
Nevertheless, coming along with its popularity, researches on SemCom are relatively scattered and with varied research interests. The discussions on "*what*", "*why*" and "*how*" are at their infancy. We argue that a comprehensive tutorial-cum-survey on SemCom covering the basic definitions, research taxonomy, and the developing trends will be very beneficial for the researchers and practitioners.
:::
[^1]: Notably, besides AI, there are alternative means (e.g., ) to implement SemCom.
[^2]: The details of CSIT and TSSI shall be given in Section [3.2](#sec:semantic_uncertainty){reference-type="ref" reference="sec:semantic_uncertainty"}.
[^3]: In this paper, we use the terminologies "task-oriented" and "goal-oriented" interchangeably. | {'timestamp': '2022-12-22T02:14:44', 'yymm': '2212', 'arxiv_id': '2212.08487', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08487'} |
# Introduction
Constituency parsing is a fundamental task in natural language processing that involves assigning constituent labels to each span of a sentence and constructing a parse tree. Constituency parsing is useful for a variety of downstream NLP tasks, including information extraction, machine translation, and question answering. There are mainly two paradigms for constituency parsing, transition parsers and chart parsers. In the former, a constituent parse tree is constructed by a sequence of shift-reduce operations. Although transition parsers have \(O(n)\) time complexity, they usually require additional feature designs to increase the accuracy. In the latter, the decoder utilizes the CKY algorithm to obtain the optimal parse tree based on the span scores assigned by neural networks. In this paper, we concentrate on chart parsers, which have the advantage of achieving globally optimal solutions, rather than being limited to local optima as in transition parsers. In the past, the CKY algorithm constructs a parse tree using context-free grammar (CFG), which contains a large number of syntactic rules. Probabilistic context-free grammar (PCFG) assigns a probability to each rule based on CFG, and the probability of a parse tree is the product of probabilities of all the rules applied when constructing the tree, where the parse tree with the maximum probability indicates the optimal tree. In recent studies, most researchers utilize neural models for constituency parsing. The seq2seq model interprets parsing as a sequential task, and when the attention mechanism is employed, it requires \(O(n^2)\) complexity. present a basic neural structure that independently scores the label for each span. Rather than depending on grammar to construct the parse tree, this approach enables the neural network to learn implicit rules. Subsequent work has achieved outstanding performance by improving model structures or span representations. The limitation of previous decoding methods is that the highest-scoring label is first determined for each span, and then the CKY algorithm is used to locate the optimal split point of the span, starting from the bottom. However, if the labels of different spans have already been decided, the correlation between labels and spans will be ignored throughout the parse tree constructing process, as the combined spans with different labels must be consistent with syntactic rules. As shown in Figure [\[fig:te\]](#fig:te){reference-type="ref" reference="fig:te"}, although the conventional neural CKY decoding obtains a higher score, it constructs a wrong tree. To attack this problem, we propose a rule-based CKY decoding algorithm. The main idea is to ensure that the label of each node in the tree and those of its children must be combined into one of the rules that exist in the training data. The main difference between our decoding and conventional neural CKY decoding is that our decoding determines a per-label optimal split point for each span, which is constrained by the rules, while conventional decoding determines the label of each span and its split point independently. The purpose of our rule-based decoding goes beyond improving decoding performance. We implement a fast GPU accelerated version of our decoding so that we can use it in training without incurring a significantly prolonged training time. In contrast to the conventional CKY decoding written in *C* of \(O(n^3)\) complexity on sentence length \(n\), our fast rule-based algorithm is \(O(n)\) in terms of GPU operations, although its theoretical complexity is \(O(n^3|\mathtt{L}||\mathtt{R}|)\), where \(|\mathtt{L}|\) is the number of labels and \(|\mathtt{R}|\) is the number of rules. Compared with another of our implemented GPU accelerated CKY decoding, the actual training time of our GPU accelerated rule-based decoding merely increases each training epoch from 15 minutes to 17 minutes on a GeForce Titan RTX. To evaluate the proposed method, we build our parser based on a prior state-of-the-art parser, and evaluate on PTB and CTB. Furthermore, we test the generalization ability of our parser on MCTB in zero-shot settings. The experimental results show that our rule-based CKY decoding algorithm brings significant improvements to constituency parsing in both English and Chinese. In conclusion, the main contributions of our research can be summarized as follows:
- We propose a GPU accelerated rule-based constituency decoder, whose complexity is \(O(n)\) in terms of GPU operations.
- With our accelerated rule-based decoder, we investigate the joint training that jointly optimizes the losses from a rule-based parser and a conventional rule-less parser.
- Experimental results show that our method outperforms previous approaches, which achieves 95.89 and 92.52 F1 on PTB and CTB respectively, and also achieves strong cross-domain performance in zero-shot settings.
# Related Work {#relate}
**Label parsing** Recently, numerous methods have been presented to simplify constituency parsing to a sequence tagging task. Supertagging is one of these methods, which is common in CCG parsers. CCG parsers impose constraints on the valid derivation of supertags and require a complex search procedure to find a sequence with the highest score. propose a method with an infinite size label set, which completely models parsing as tagging. **Transition parsing** Transition parsers utilize a transformation system that takes a sequence of lexical inputs and produces a sequence of tree-constructing actions. They parse sentences sequentially and have linear time complexity. Using special features, or advanced decoding methods, the accuracy of transition-based parsers is improved effectively. use only four actions, further improving the efficiency.
[\[a1\]]{#a1 label="a1"}
[\[a2\]]{#a2 label="a2"}
[\[a3\]]{#a3 label="a3"}
[\[a4\]]{#a4 label="a4"}
[\[a5\]]{#a5 label="a5"}
## Cross-domain Results
In Table [\[tab:cross-domain\]](#tab:cross-domain){reference-type="ref" reference="tab:cross-domain"}, we compare the generalization ability of our method with previous approaches. All the parsers are trained on the PTB training set and evaluated on the PTB dev set, and tested their cross-domain performance on MCTB in zero-shot settings. Our rule-based parser achieves 88.09 F1 on average in cross-domain datasets with an absolute improvement of 1.41 points compared with our rule-less parser, which also outperforms other previous approaches. To further explore the role of syntactic rules in our parser, we count the number of rules in each test set compared to the rules in the training set (the PTB training set for English and the CTB training set for Chinese). As shown in Figure [\[fig:recall\]](#fig:recall){reference-type="ref" reference="fig:recall"}, our syntactic rules in both PTB and CTB test sets have over 99.9% recalls (weighted by the number of occurrences) over their training sets. In cross-domain datasets, although about 25% to 35% of the rules do not appear in the PTB training set, the weighted statistics show that the recalls of the rules in cross-domain datasets are all over 95.8%. This demonstrates that our simple rule labeling scheme extends effectively to other domain datasets, which also improves the generalization ability of our model. Although our data-driven algorithm cannot generalize to a small proportion of unseen rules without a specific generalization objective, our method can utilize known syntactic rules to correct errors, resulting in better parsing performance.
## Speed Analysis
Our method can be completely parallelized using GPU accelerations. For each epoch, it takes 15.2 minutes for rule-less training, and 17.5 minutes under a mixture of rule-based training and rule-less training, which is merely increased by 15%. It shows the efficiency of our GPU accelerated implementation. We compare our parsing speed with other parsers in Table [\[tab:speed2\]](#tab:speed2){reference-type="ref" reference="tab:speed2"}, although this comparison might be unfair since they are run on different hardware. Our rule-less parser processes 121 sentences per second, and our rule-based parser processes 54 sentences per second. Compared with which use a GeForce RTX 3090, our parser on a GeForce Titan RTX is much faster, while F1 scores of the two parsers are comparable.
# Conclusion {#sec:conclusion}
In this paper, we first implement a fast neural CKY decoding, based on which we further propose a rule-based CKY decoding, whose time complexity is \(O(n)\) in terms of GPU operations. We also investigate the joint training of a rule-based parser and a conventional rule-less parser. Our experiments show that incorporating syntactic rules into decoding and training can rectify errors in the parse tree, which achieves significant improvements on both English and Chinese benchmark treebanks. Furthermore, we demonstrate that our method is robust in zero-shot cross-domain settings. | {'timestamp': '2022-12-19T02:11:23', 'yymm': '2212', 'arxiv_id': '2212.08458', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08458'} |
# Introduction
Spectroscopy is the essential tool to investigate the structures of atoms and molecules, which has outstandingly contributed to atomic and molecular physics, analytical chemistry, and molecular biology. It also gained enormous significance in the fields of optical sensing, environmental study, and medical diagnostic. Spectroscopic measurements with high resolution, broad bandwidth and calibrated optical frequency are fundamentally important for high-precision investigations of multiple samples. Spectrometers, widely implemented in spectroscopic measurement, are facilitated by using speckle patterns . The speckle pattern generated in disordered media is distinct at each wavelength to reconstruct the input spectrum. Random interference or scattering generate speckle patterns in space during the propagation of light in multimode fiber , integrating sphere , or some other materials . Similarly, speckle patterns in time domain are generated from Rayleigh scattered lightwave in single-mode fiber, which can be detected by using a single photodetector (PD) instead of using camera or PD array . The compact speckle-based spectrometers have reached picometer resolution outperforming the state-of-art grating spectrometers . It is challenging to further reduce the spectral resolution. The determination of wavelength with higher precision may be realized by using speckle-based wavemeter. Sub-femtometer resolutions have been demonstrated respectively by using cross-correlation algorithm , principal component analysis , convolutional neutral network . The speckle-based wavemeter only determine sparse discrete wavelength rather than an optical spectrum. Besides, it is not suitable to be used in swept laser spectroscopy, since the speckle patterns may be distorted. Therefore, the current implementations of speckle-based reconstructive schemes are incapable to realize high-resolution spectroscopy. Optical frequency comb (OFC) is invented from femtosecond fiber laser for precision metrology by building a link between optical and radio frequency. Although laser spectroscopy is promoted from the assistance of OFC , the equidistant coherent lines from OFC may be more powerful tools for simultaneously measuring broadband spectra as direct comb spectroscopy. Several techniques have been proposed for resolving the comb lines in broad bandwidth, including using a virtually imaged phase array , using a scanning Fabry-Perot (F-P) cavity , Fourier transform spectrometer , dual-comb spectrometer , and speckle-based fiber spectrometer . In the last decade, more comb sources have been demonstrated with different repetition frequencies at more wavelength regions . Electro-optic frequency combs (EOFCs) are attractive tools in high-resolution spectroscopy for providing ultra-fine comb lines. EOFC is generated by electro-optic modulation of a continuous-wave laser, which provides the agility of both centre frequency and repetition frequency. The repetition rate of EOFC, practically spectral resolution in many direct comb spectroscopy, is not limited by physical cavity length. Spectroscopy with sub-MHz resolution is facilitated by digitally-generated ultra-fine EOFCs , which reaches high refresh rate using coherent detection. Relatively, the bandwidth of EOFC is limited by the electrical driving synthesizer. Although the measurement bandwidth of EOFC may be broadened from the frequency agility provided by the tunable seed laser, the frequency precision is far behind the requirement for sub-femtometer resolution spectroscopy. Here, our work leverages the advantages of speckle-based wavemeter and ultra-fine EOFC to achieve ultra-high resolution spectroscopy with calibrated absolute frequency over a broad bandwidth. A novel wavelength (or optical frequency) determination technique based on whispering-gallery-mode (WGM) barcodes reaching sub-femtometer resolution is proposed. WGM barcodes (or WGM speckle patterns) are composed of complex WGM resonances with different depths, spacings, linewidths, and resonant wavelengths, which are firstly proposed for temperature sensing . A continuous-wave (CW) laser is used to obtain measurement speckle for wavelength determination and generate ultra-fine EOFC for high-resolution spectroscopy. Compared with a reference speckle calibrated by an ultra-stable laser, the measurement speckle determine the absolute wavelength of the CW laser thanks to the ability of two-tone demodulation . High-resolution spectrum of sample and accurately determined frequency of each comb lines are simultaneously achieved in millisecond measurement time. Broadband measurement is realized by doing sequential acquisitions at different centre wavelength without the requirement of precise wavelength adjustment of the tunable CW laser. As an application of the scheme, spectroscopic measurement of a high-Q fiber F-P cavity with 0.8 fm resolution in 1 nm bandwidth and HCN gas cell with 8 fm resolution in 60 nm bandwidth are respectively demonstrated with a fiber laser and an external cavity diode laser (ECDL). The proposed method bridges the fields of speckle-based wavemeter and comb-based spectroscopy, which shows promising prospects for precise spectroscopy with broad bandwidth.
# Principle and Methods
The principle of speckle-based spectroscopy with an EOFC is shown in Fig. 1. The WGM speckle patterns (or WGM barcodes) are composed of complex WGMs excited in a silica micro-rod resonator. A tapered fiber is used to couple the light into and out of the WGM resonator. The WGM speckle patterns are unique at each wavelength (or optical frequency), since the depths, spacings, linewidths, and resonant wavelengths of WGM resonances are different. Hence, wavelength can be determined from such wavelength-dependent speckles. Cross-correlation algorithm is used to find the relative location of the speckle compared with a broadband reference speckle. The reference and measurement speckles are obtained by swept laser configuration with the setup shown in Fig. 2. The reference WGM speckle is recorded before measurement. A linearly frequency swept RF signal with a sweep range of 10 GHz is applied on the single-sideband modulator (SSBM). The CW laser is modulated and passes through the WGM resonator. A PD with 120 MHz bandwidth and an OSC with 250 MSa/s sampling rate are used for detection. The WGM resonances in 10 GHz banwidth is measured, which is regarded as an individual WGM speckle. The power fluctuation during the frequency sweeping is eliminated by introducing a channel without the resonator. Several WGM speckles over a broad bandwidth are obtained by adjusting the centre frequency of the tunable CW laser with a step of about 5 GHz or 40 pm. The spectra of adjacent acquisitions have an overlapping region, which is used for the stitching of these acquisitions to get a reference speckle covering the whole bandwidth. The stitching points are calculated by using cross-correlation algorithm. The reference speckle can be regarded as a "ruler" in optical frequency domain by using the speckle of an ultra-stable laser to calibrate the absolute frequency. Hence, the absolute frequency of the CW laser is obtained from relative location. The EOFC generated from the CW laser records the spectrum of the sample, and its comb lines are resolved in an interferometer. Extending the bandwidth of high-resolution spectroscopy is enabled due to the calibrated frequency and the tunability of the CW laser. By adjusting the centre frequency of the tunable laser, the broadband precise spectra are acquired from EOFC centred at different frequencies. Importantly, precise wavelength tuning is not required for high-resolution spectroscopy, since the absolute wavelength is determined from the speckle patterns. The bandwidth of spectroscopy is no longer limited by the bandwidth of EOFC but by the tunable range of the CW laser considering much broader range of the WGM speckle. The experimental setup is shown in Fig. 2. The WGM resonator is a silica micro-rod with a diameter of 1.5 mm fabricated by using CO\(_2\) laser. The Q-factor of the resonator reaches over 10\(^7\). Higher Q-factor contributes to increasing the resolution of frequency determination by both reducing the FWHM of the cross-correlation peak and the noise floor. The quantified simulation results are detailed in Supplementary 1, Section 2. The average insertion loss of the resonator is about 10.2 dB. The tunable laser and the ultra-stable laser are coupled to obtain their WGM speckles. A SSBM is driven by a linearly-swept radio frequency (RF) signal. The dual-wavelength lightwaves are frequency-modulated to record several WGM resonances of the WGM resonator around their centre wavelengths. The WGM speckles are detected and digitized by using a PD and an oscilloscope (OSC). The data, represented as the superposition of two WGM speckles, is compared with the reference speckle by using cross-correlation algorithm to determine the absolute frequency of the tunable laser. Another partial output of tunable laser is modulated by using an electro-optic modulator (EOM) to generate EOFC. The driven RF signal generated from an arbitrary waveform generator (AWG) is a periodic waveform with a flat spectrum, which can be regarded as an "electrical comb". An EOFC with twice of the RF bandwidth is generated. The repetition frequency set by the period of the driven signal reaches sub-MHz level for high-resolution spectroscopy. The comb lines recording the spectrum of the sample are measured by using an interferometer. An acousto-optic modulator (AOM) introduces a centre frequency shift to distinguish the positive and negative sidebands of the EOFC. The interferograms are detected by using a balanced photodetector (BPD). The comb lines are resolved after Fourier transformation in digital processing. The acquisition of WGM speckle and comb spectrum is realized within 1 ms, regarded as one acquisition at one centre frequency. Longer measurement time contributes to improving the signal-to-noise ratio (SNR). A sequence of acquisitions by changing the centre frequency of the tunable laser obtains a broad bandwidth. In following experiments, a fiber laser with low frequency noise, and an ECDL with broad tuning range, respectively serve as the tunable laser to reach 100 kHz spectral resolution and 80 nm bandwidth.
# Results
## High resolution of the system
A fiber laser serves as the tunable laser in the demonstration of high-resolution measurement. The superposition of the WGM speckles of the tunable laser and the ultra-stable laser in 0.8 ms is shown in Fig. 3(a). The range in optical domain is 8 GHz, which contains totally more than one hundred WGM resonances. These resonances have different depths and widths, as shown in the zoom-in image (Fig. 3(b)) in 50 \(\mu\)s. The WGM speckles composed of a mass of WGM resonances can be regarded an approximately stochastic curve at certain wavelength region. Cross-correlation results comparing this measurement speckle and the reference speckle are shown in Fig. 3(c). The bandwidth is limited by the tunable range of the fiber laser to be 120 GHz. The left peak, shown in Fig. 3(d) in red, represents the location of the ultra-stable laser at 193.3480000 THz, which calibrates the frequency of the whole cross-correlation results. The frequency of the fiber laser is determined to be 193.4046926 THz with a resolution of 100 kHz according to the right peak of the correlation-curve as shown in Fig. 3(e). The peak location determination is based on Lorentz fitting. The full widths at half maximum (FWHMs) of two peaks are both about 41 MHz. The resolution of frequency determination is evaluated by measuring a frequency-modulated fiber laser. An AOM is introduced to modulate the centre frequency. The readouts of the relative frequency obtained from the WGM speckles are shown in Fig. 3(f). The sinusoidal frequency modulation with an amplitude of 100 kHz is well demodulated. The refresh rate of the frequency determination is 1.25 kHz. The Fourier transformation shown in Fig. 3(g) reveals the modulation frequency at 250 Hz with a SNR of 261.
## Spectroscopic measurement of a high-Q Fabry-Perot cavity
The EOFC generated from the fiber laser by electro-optic modulation is measured in the heterodyne interferometer. The interferograms in 1 ms recording time are Fourier transformed to obtain the spectrum of EOFC as shown in Fig. 4(a). 25000 comb lines covering a bandwidth of 2.5 GHz are resolved. The centre frequency in RF domain is 80.025 MHz corresponding to the frequency shift introduced by AOM, and the centre frequency in optical domain is determined by its WGM speckle. A zoom-in figure in 2.5 MHz bandwidth is shown in Fig. 4(b), in which the linewidth of comb line is Fourier-transform-limited to be 1 kHz. The line-spacing is 100 kHz, corresponding to the resolution of spectroscopy. The sample is a fiber F-P cavity composed of two fiber Bragg gratings. Its reflectance spectrum is measured by the EOFC. One resonance of the F-P cavity is also represented in Fig. 4(b). The spectroscopy with extended bandwidth is performed by adjusting the centre frequency of the fiber laser. The whole reflectance spectra in 117 GHz bandwidth are obtained from a sequence of 47 acquisitions, as shown in Fig. 4(c). Each acquisition resolves 25000 comb lines in 1 ms recording time, and the absolute frequency is determined in 0.8 ms. Totally more than 10\(^6\) spectral points are demodulated. The cavity resonances are in the reflective band around 193.39 THz in 0.5 THz bandwidth. Two adjacent acquisitions (respectively in blue and red) containing several narrow resonances are shown in Fig. 4(d). These resonances in overlapping region are well matched thanks to the absolute frequency calibration. The SNR is 227, calculated from the standard deviation of the baseline. The ripple on the baseline introduced from the fiber F-P cavity is eliminated by using a fitting process in calculation. The free spectral range of the fiber cavity is measured to be 105.6 MHz. One of the resonance with a linewidth of 250 kHz is shown in Fig. 4(e), corresponding to a cavity Q-factor of 7.7\(\times\)`<!-- -->`{=html}10\(^8\). The ultrahigh-Q cavities, as essential tools in precise optical sensing and nonlinear optics, can be properly characterized in the proposed technique.
## Broad bandwidth of the system
The broad bandwidth of the system is performed by using an ECDL as the tunable laser. The tunable range of the laser is extended to 9.5 THz (or 76 nm). Similarly, the cross-correlation result between the measurement speckle and reference speckle is shown in Fig. 5(a). The green curve shown in Fig. 5(b) and the red curve shown in Fig. 5(c) respectively represent the centre frequencies of the ECDL and the ultra-stable laser. The frequency of the ECDL is determined to be 190.164017 THz. The resolution is 1 MHz, limited by the frequency noise of the ECDL rather than instrumental performance. To evaluate the resolution, frequency modulation with an amplitude of 1 MHz is also introduced by using AOM. The frequency readouts are shown in Fig. 5(d). The refresh rate of the frequency determination is increased to 12.5 kHz. The Fourier transformation reveals the frequency modulation with an amplitude of 1 MHz and a frequency of 2.5 kHz, as shown in Fig. 5(e). Intrinsic frequency noises of the ECDL in low frequency region with about 1-MHz amplitudes are also observed. Therefore, the readouts in Fig. 5(d) are the superposition of frequency modulation and frequency noise. The frequency noise is about 0.48 MHz in 1 ms, which is the limitation of spectral resolution for further spectroscopic measurement.
## Spectroscopic measurement of HCN gas
An EOFC with 2.35-GHz bandwidth and 1-MHz repetition rate is generated to measure the transmission spectrum of a H\(^{13}\)CN gas cell. The H\(^{13}\)CN gas cell with a length of 15 cm and a pressure of 25 Torr is in a laboratory temperature of about 297 K. The results shown in Fig.6 (b) are composed of the resolved spectra of 1890 acquisitions. Each acquisition obtained in 1 ms contains 2350 comb lines with a resolution of 1 MHz. Totally 4\(\times\)`<!-- -->`{=html}10\(^6\) spectral points are recorded in the whole spectrum covering a bandwidth of 4.74 THz or 38 nm. The measurement result and a reference database of the \(R\) and \(P\) branches of H\(^{13}\)CN are reversely shown in Fig. 6(b) and 6(a), respectively in blue and red. The residual errors between the database and the result are shown in Fig. 6(c). The standard deviation is calculated to be 0.0069, corresponding to a SNR of 144.
# Discussion
Generally speaking, speckle-based wavemeter and spectrometer suffer from the environmental perturbation, since the speckle patterns generated from multiple interference and scattering light are extremely sensible. Therefore, active stabilization techniques are required for practical measurement, otherwise, the distortion of speckle decreases the precision and accuracy. The WGM resonances are also sensible to external temperature and vibration. All characteristics of resonances, such as linewidths, coupling depths, and resonant wavelengths, are included in WGM barcodes , which resists mild nonlinear response of moderate external perturbation. Therefore, the frequency calibration, conducted by measuring the coupling of the ultra-stable laser and the measurement laser, enables our measurement without active stabilization of the resonator. The WGM speckle-based scheme proposed in this paper can also be independently used as a wavemeter, which can determine absolute frequency with a resolution of 0.8 fm and a bandwidth of 80 nm. In speckle-based spectrometer or wavemeter, optical path length is usually longer than other conventional spectrum analyzers with the same device size since multiple interference or scattering, which facilitates miniaturization and high resolution. The inverse relation between FWHM of the cross-correlation peak and Q-factor in WGM speckle-based wavemeter also demonstrates that higher resolution can be achieved with a longer optical path length (details are in Supplement 1). Compared with other structures, WGM micro-resonator is able to realize longer optical path length with same footprint due to high Q-factor. Therefore, we believe that WGM micro-resonator is a potential choice to achieve stability, miniaturization, and high resolution for speckle-based spectrometer or wavemeter. Here, we also propose and experimentally realize a novel spectroscopy method that introduces ultra-fine EOFC to exploit the high resolution and broad bandwidth of speckle-based wavemeter for spectroscopic measurement. Ultra-fine EOFC precisely samples the spectrum, whose centre frequency is absolutely determined at the same time. This scheme breaks through the bandwidth limitation of EOFC in high-resolution spectroscopy by stitching a sequence of spectral acquisitions centred at different frequencies. Such combination of EOFC and speckle-based wavemeter fully utilizes each characteristics and advantages to be a prospective way towards ultra-high resolution spectroscopic measurement over an much broader bandwidth for precise investigation in bio-chemical sensing and physics.
## Data, Materials, and Code Availability {#data-materials-and-code-availability .unnumbered}
The data that support the findings of this study are available from the corresponding author upon reasonable request. | {'timestamp': '2022-12-19T02:13:08', 'yymm': '2212', 'arxiv_id': '2212.08510', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08510'} |
# Introduction
Radio galaxies are galaxies that emit a large fraction of their electromagnetic output in the radio band. The structures visible in radio wavelengths are typically larger than the structures visible in optical wavelengths. Radio galaxies are a class of active galactic nuclei (AGN) and are powered by supermassive black holes at the centres of galaxies. The extended emission is produced by synchrotron radiation of highly relativistic particles accelerated by the AGN. Studying radio galaxies helps to understand the effects of massive black holes on their environment (see e.g. ). The jets of highly energetic particles emitted by giant radio galaxies potentially play a major role in the creation of cosmic magnetic fields. A lot of new radio sources will be discovered with the new generation of radio telescopes (e.g. LOFAR, MeerKAT, and in the future the SKA ). Processing the incoming data is one of the biggest challenges in radio astronomy. The cause is not only the enormous amount of data, but also the higher source density due to the improved sensitivity of the instruments. Novel techniques are required for this purpose. Deep learning has proven to be useful for automation of data processing, but commonly requires large amounts of labelled training data for supervised algorithms, which are not always available. For instance, the SKA data challenges have demonstrated the difficulties of source finding for SKA data. Morphological classification of radio sources can be achieved by deep learning models trained on well-understood data sets. use [cnns]{acronym-label="cnn" acronym-form="plural+short"} trained on data from the FIRST (Faint Images of the Radio Sky at Twenty-Centimeters) survey for the classification of radio galaxies. The architectures of the neural networks for classification are inspired by the Alexnet. For other approaches in radio galaxy classification, see e.g.. In other areas of astronomy, similar morphological classification problems arise, e.g. for classification of optical galaxies, and of gravitational lenses. Here supervised methods of machine learning have been applied with some success, see e.g.. However, the existing number of radio sources with morphological labels is limited (\(\approx1300\) in the MiraBest data set with two classes ). These class labels are typically extracted from catalogues created and curated manually by experts. Small data sets used in the training of deep learning models for galaxy classification can be enlarged by data augmentation, e.g. by applying random rotations and reflections to the images (classical augmentation). A different approach based on equivariance implements the symmetry constraints of the problem directly in the construction of the model. This may help classifiers to understand symmetries without relying exclusively on augmentation and may be particularly useful for problems with sparse data. In this work, we investigate a novel application of generative models to enhance the available training sets. For this augmentation technique, multiple neural networks are combined to learn the underlying distribution of a data set. We focus on the task of classifying different morphological types of radio galaxies. The morphological classification scheme by Fanaroff-Riley is fundamental for such applications. For the class FRI, the unique maximum of the radio emission resides in the centre of the source and the surface brightness decreases along the jets. For FRII sources, the two maxima of the radio emissions are located at the edges of the jets and the surface brightness in the centre is lower. As radio sources have a large variety of structures, we consider two more classes. Unresolved and point sources are contained in the Compact class. The Bent class consists of sources for which the angle between the jets differs significantly from 180 degrees. The two sub-types narrow-angle tail (NAT) and wide-angle tail (WAT) are further discriminated by the angle, but are fully subsumed in the Bent class for this study. As in, we study a four-class classification problem, including bent-tail and compact sources in addition to the classes FRI and FRII of. illustrates the considered classes (FRI, FRII, Compact, and Bent). In this study, we investigate whether different radio galaxy classifiers can be improved when training is supported by providing additional data generated with a [wgan]{acronym-label="wgan" acronym-form="singular+short"}. For similar approaches from different fields see, for example,. We extend our framework presented in to handle larger ratios between real and generated images. Additional images are only generated when they are needed during training. As before, we start with a simple model, namely a [fcn]{acronym-label="fcn" acronym-form="singular+short"}. In addition, we apply [wgan]{acronym-label="wgan" acronym-form="singular+short"}-supported augmentation to a [cnn]{acronym-label="cnn" acronym-form="singular+short"} and a [vit]{acronym-label="vit" acronym-form="singular+short"}.\
The long-term goal is to use classification models to process incoming data from new radio telescopes. For this purpose, classification models need to generalise particularly well. A common problem in astronomy is the scarcity of labelled data in the face of large amounts of new data to process. This is a very different situation as for instance in particle physics, where simulations are highly fine-tuned and experiments are constantly repeated. In particular, for forthcoming radio surveys, and even for the majority of sources in FIRST, no morphological labels are available. As a result unsupervised, semi-supervised and self-supervised methods have gained attention without reaching the performance of supervised methods. The current classification scheme of radio galaxies and our physical interpretation will be challenged by new radio surveys (see e.g ). Discoveries of rare morphologies can help to extend our understanding of radio sources, but are potentially prohibited by supervised learning techniques. This paper is organised as follows: In , we introduce the data set used for training, validation, and testing. The generative model and its implementation are described in . The training procedure and the assessment of image quality are discussed in . The results of the comparison between only classical and classical plus [wgan]{acronym-label="wgan" acronym-form="singular+short"}-supported augmentation for different classifiers are presented in before we conclude in .
# Data {#Sec:Data}
We combined different catalogues that characterise radio sources from the FRIST survey to create a data set of 2158 radio galaxy images with morphological labels. The labelling in the catalogues is typically performed by experts by considering radio images and the corresponding optical counterparts. We group radio sources in 4 classes, namely FRI, FRII, Compact and Bent. The data set is published on zenodo and on GitHub (<https://github.com/floriangriese/RadioGalaxyDataset>). The radio galaxy images of the FIRST survey are collected from the virtual observatory skyview[^1]. We downloaded the original images with a size of 300 x 300 pixels. Then we adopted the preprocessing from after cropping the images to the input size of our generative network (128 x 128 pixels). In particular, we set all pixel values below three times the local RMS noise to the value of this threshold. Subsequently, the pixel values were rescaled to the range between-1 and 1 to represent floating point greyscale images. Finally, we applied classical augmentation to each image of the training set during training. We separated 100 sources per class from the data set for the final evaluation of our models. For validation purposes during training (e.g. choosing the best model), we use a 5-fold cross-validation. Therefore, we do not need a separate validation set. As a result, we lose less training data. In particular, we split the training set into five blocks and did five separate training runs. For each of these runs one of the five blocks was used as the validation set and the remaining four blocks represented the corresponding training set. The quantities per class and per split are shown in .
# Wasserstein GAN {#sec:wGAN}
The ability to learn representations of underlying statistical distributions of data sets makes generative models a powerful tool for the creation of additional data points. In particular, sampling from those representations allows to speed up conventional simulation techniques significantly and may be useful for further subsequent treatments. Three different categories of generative models are well-established: generative adversarial networks (GANs), variational autoencoders (VAEs), and flow-based models. Diffusion models represent a relatively new development in this area. In this work, we focus on GANs. They consist of two neural networks: a generator \(G\) that generates fake images from a noise vector \(Z\) and the discriminator \(D\) that discriminates between real and fake images. This architecture was first introduced in. In a two-player minimax game, the generator learns to create fake images, which become less and less distinguishable from the real ones in the course of the training. The loss function for this setup reads : \[L=\min_G \max_D \mathbb{E}[ \log D(x) ] + \mathbb{E} [\log (1-D(G(z)))],\] where \(x\) represent real samples and \(G(z)=\tilde{x}\) generated samples. For this project, we employ a variant of the standard GAN setup called [wgan]{acronym-label="wgan" acronym-form="singular+short"} that uses the Wasserstein-1 metric, also referred to as the Earth Mover's distance, as main term in the loss function . This loss function is calculated as \[L=\sup_{f\in\text{Lip}_1}\{\mathbb{E}[f(x)-\mathbb{E}[f(\tilde{x})]\},\] where \(f\) denotes a 1-Lipschitz function that is learned during the training procedure. The discriminator of a standard GAN is transformed into a critic and is used to estimate the Wasserstein distance between real and generated images. Hence, the absolute value of the loss function is correlated with the image quality, resulting in the name change. Additionally, the training of [wgans]{acronym-label="wgan" acronym-form="plural+short"} is often more stable and more likely to converge than standard GAN setups. To approximate the Wasserstein-1 metric by use of a critic network, it has to be ensured that the 1-Lipschitz constraint is fulfilled. This is achieved by applying a gradient penalty term to the loss function as in \[L=\lambda\ \mathbb{E}[(\|\nabla_{\hat{x}}f(\hat{x})\|_2-1)^2]\] for random samples \(\hat{x}\sim\mathbb{P}_{\hat{x}}\). Since we work with image data, it has proven to be the most promising approach to construct a [wgan]{acronym-label="wgan" acronym-form="singular+short"} setup based on convolutional layers. The generator receives a noise tensor of size 100\(\times\)`<!-- -->`{=html}1 and a class label \(y\) and, through multiple layers of 2D transposed convolution operators, enlarges this to a 128\(\times\)`<!-- -->`{=html}128 tensor, consistent with the dimensions of real images. The critic is given either real or generated images, as well as the class label \(y\). The output of the critic is a single real value, which represents the belief of the critic for the image to be real. Generator and critic are trained intermittently, where the critic has five training cycles per training cycle of the generator. While training the generative model, it is necessary to apply classical augmentation such that the symmetries of the training set are also present in the generated data sets, and to avoid introducing a bias due to the limited number of training examples.\
Morphologies of radio galaxies are diverse and result in very different images. Consequently, it is reasonable to condition the networks with the class label \(y\) such that a combination of image and class label is provided to the networks. In particular, this allows applying supervised learning techniques on the output of the generator. For our setup, this is achieved for the generator by applying a 2D transposed convolution operator on a matrix of image dimensions filled with the class label. The transpose-convoluted layer is then concatenated to the first transpose-convoluted layer of the noise tensor. Batch normalisation in 2D and ReLU activation functions are used. The concatenated tensor is then passed through five additional 2D transposed convolutions, where no normalisation or activation is applied after the last layer. Instead, the individual pixel values are clipped to \[-1,1\] for conversion to grayscale. The critic is built analogously, but uses 2D convolutional layers, resulting in a single output node representing the critic score for image quality. Here, layer normalisation and leaky ReLU functions are used except for the last layer. A schematic of the wGAN setup can be found in .
# Results of image generation with a wGAN {#sec:GenImg}
## Training
For each choice of training and validation data in the cross-validation procedure a [wgan]{acronym-label="wgan" acronym-form="singular+short"} training run is launched on the corresponding training set. The training is performed with a single NVIDIA A100 GPU provided by the Maxwell cluster at DESY for 40k generator iterations. A batch size of 400 is chosen and one training run takes roughly seven hours to complete. The generator and critic weights are saved every 250 iterations, allowing to scan for the best training state later on, as described in the following section.
## Evaluation of image quality
In this section, we present images created using the generator of the [wgan]{acronym-label="wgan" acronym-form="singular+short"} and examine the quality of the generated images in several ways.
### Distribution-based comparison
We define a set of distributions to compare generated images with the training data set, in order to determine the quality of generated images and thus to find the best performing training iteration. This includes normalised histograms of pixel intensities, the number of pixels with an intensity greater than zero and of the sum of intensities. These histograms are compared for each class individually and the relative mean absolute error (RMAE) between the generated set of 10k images and the training set is computed. The RMAEs for the different distributions are summed up to yield a single figure-of-merit (FOM), where the [wgan]{acronym-label="wgan" acronym-form="singular+short"} training iteration with the lowest FOM value is used in the following. This procedure is followed for each of the four classes separately, i.e. we allow a different iteration of the generator training to yield the best model for each class.\
Arbitrarily chosen examples of these distributions are shown in , where the distribution of the real images is shown in orange and the distribution of the generated images in blue. The uncertainty for each bin is given by the square-root of entries in that bin before normalisation. The bottom panels in this figure show the per-bin divergence between the distributions, where absolute deviations larger than 1 are indicated by the corresponding value written in boxes. Here, only examples from the first cross-validation fold (of five) are shown.\
Overall, the distributions of the generated images tend to follow the distribution of the real images. Nevertheless, the generated images have difficulties in recreating very low, but non-zero, intensities. This can be seen for pixel values between 1 and 20 in , which directly translates into under-representing the number of pixels with an intensity \(I>0\) in .
### Visual comparison
In order to get a visual idea of image quality, we generated a set of 5k images per class and compared them to the full training data set over all cross-validation folds. The images are rotated so that their principal components are aligned. Subsequently, we compute the pixel-by-pixel difference for all possible pairs of real and generated images. All classes also include a few difficult to define sources with rather small spatial extension that are easy to emulate but do not show the generator's capability of reproducing the more interesting extended sources. Thus, we only consider images with an intensity sum of at least 15k (5k) for the extended (compact) radio galaxies. We show the resulting closest pairs for each class in . By eye, the generated images appear very similar to the analogue real images, indicating a good performance of the generator setup. To also get an impression of the diversity of the generated data we show a random set of generated images in App. [\[App:img\]](#App:img){reference-type="ref" reference="App:img"}.
### Classifier-based comparison
Next, we use a CNN trained solely on the data set of real images to assess the image quality further. We compare the performance of this classifier on the real test set and a set of generated images. The architecture of the CNN used for this experiment is summarised in . A comparison of the confusion matrices on both sets tests for any bias introduced by the image generation. In particular, we evaluate the conditioning on the class labels. First, we show in the top panel of the confusion matrix of the classifier on the real test. We compare this to the confusion matrix of the same classifier evaluated on the set of generated images on the bottom panel of . We find that the class conditioning of the generated images works well overall. However, confusion for images of the class FRI with the predicted classes FRII is enhanced on the generated test set. The classification performance of the class Compact is deteriorated on the generated test set and particularly the confusion for images of the class Compact with the predicted class FRII is enhanced. As some FRII-like sources resemble a combination of two compact sources this is not surprising. Confusion for images of the class Bent with the predicted classes FRI is for instance slightly reduced.
# Results of classifier training using wGAN-supported augmentation {#sec:wGANsupAug}
We assess the new approach of supplementing the training set with generated images by comparing the performance of different classifiers (each trained on different setups with increasing amount of generated data). Our benchmark is the performance of the classifier trained on the original training set. We test the performance of the classifier trained with the original training set plus simulated images by the generator of the [wgan]{acronym-label="wgan" acronym-form="singular+short"} against this benchmark. We start with a [fcn]{acronym-label="fcn" acronym-form="singular+short"} (see ). Subsequently, we increase the complexity of the classifier by training a CNN (see ). Finally, we apply our framework to a state-of-the-art classifer, namely the [vit]{acronym-label="vit" acronym-form="singular+short"}. We use the ViT-B_16 vision transformer configuration with pre-trained weights from the ImagetNet21k data set[^2] with a resetted head layer. The GAN generated images with pixel sizes 128x128 are zero-padded up to 224x224 pixels to fit the pre-trained model input size. As a loss function, a weighted cross-entropy loss is implemented. As an attention based model, the ViT splits the image into fixed-size patches processed by the transformer encoder. The ViT is able to output attention maps. These illustrate regions considered important by the classifier. Therefore, attention maps help to understand the internal processes of the model. We generate images on the fly, i.e. each time a generated image is loaded it is newly generated. These setups are not optimised to reach maximal classification accuracies. The goal of this study is only to compare classical augmentation with wGAN-supported augmentation for each classifiers. We do not compare the performance of the three classifiers in detail either.
#### Evaluation metrics {#evaluation-metrics .unnumbered}
To compare the overall performance among different training setups and to determine the best training iteration of a classifier training run, we use the multi-class Brier score . The Brier score is essentially the mean squared error of the predicted probabilities of a classifier for all classes. This has the advantage that also the certainty of the classifier's decision is considered, which winner-takes-all FOMs such as accuracy do not take into account. For each setup, i.e. for a given ratio between the number of generated \(i_g\) and real images \(i_r\), denoted \(\lambda = i_g/i_r\), we have five models due to the 5-fold cross-validation. The final evaluation is performed on an independent test set that contains real data only. We use the most commonly applied metric in radio astronomy publications: multi-class accuracy. In order to estimate statistical fluctuations, we average the performance metrics over the five best models of each cross-validation fold.
#### Accuracy {#accuracy .unnumbered}
The multi-class accuracy, i.e. number of correct classifications over number of all classifications, on the test data set is shown in for the three different classifiers investigated here. The results are shown for different training scenarios, where the number of generated images used to augment the training data set (represented by \(\lambda\)) is varied. The blue markers (uncertainty bars) represent the mean (standard deviation) of the obtained results over all cross-validation folds for the augmented training data sets and the horizontal orange line (area) show the corresponding result for the real data only case.\
presents the results for the [fcn]{acronym-label="fcn" acronym-form="singular+short"}, which yields an improvement in accuracy of \(\SI{18(5)}{\percent}\) over the baseline setup at \(\lambda=2\). All augmented training setups outperform the real data only case which reaches an accuracy of \(\SI{59+-2}{\percent}\).\
The highest obtained average for the [cnn]{acronym-label="cnn" acronym-form="singular+short"} classifier is reached for \(\lambda = 3\), as can be seen in , which is \(\SI{3(2)}{\percent}\) higher than the real data only baseline at \(\SI{79+-1}{\percent}\). The highest obtained average using [wgan]{acronym-label="wgan" acronym-form="singular+short"} augmented training data for the [vit]{acronym-label="vit" acronym-form="singular+short"} classifier is reached at \(\lambda = 2\), see , which is \(\SI{0.7+-2.0}{\percent}\) lower than the real data only baseline at \(\SI{81+-1}{\percent}\).
# Discussion and Conclusion {#Sec:Disc}
Other studies probe the use of generative models to create images of radio galaxies. These studies are based on Variational Autoencoders (VAEs), so that the approach used for the study presented here, utilising a [wgan]{acronym-label="wgan" acronym-form="singular+short"}, is, to our knowledge, novel to the field of radio astronomy.\
We are able to generate highly realistic images of radio sources of the four different radio galaxy classes investigated for this study. For this, we rely on the good agreement between the image metric distributions, such as the pixel intensity histogram, between real and generated images, as well as the good agreement between the confusion of a [cnn]{acronym-label="cnn" acronym-form="singular+short"} classifier trained only on real data obtained on a real data only test set and a generated data only test set. Particularly the latter, provides confidence for the class conditioning of the generator. Following a visual inspection, we note that the generated images tend to have sharper edges, i.e. low intensity pixels directly next to high intensity pixels. This is not the case for real images, which are smeared due to detector resolution effects. Resolving these issues would yield even more realistic generated images. However, we do not observe issues known from other state-of-the-art generative networks in radio astronomy, such as different noise levels between generated and training data or pseudo-textures and pseudo-structures. The results of this study therefore constitute a major improvement in generated image quality. This high quality of the images allows us to use them to improve the training of an external classifier, called [wgan]{acronym-label="wgan" acronym-form="singular+short"}-supported augmentation here. This represents an extension to realistic data of the studies done in, which showed that statistical information contained in a simplistic toy training data set can be augmented using generative models. Another extension of this study to more realistic data in the field of particle physics is given in. Here, we are able to show that adding generated images to the training data set does clearly improve the classifier performance on a real data only test set for the [fcn]{acronym-label="fcn" acronym-form="singular+short"} classifier, where the largest improvement of \(\SI{18(5)}{\percent}\) over the baseline setup is reached for \(\lambda=2\), meaning a training data set consisting of all real images plus twice as many generated images. Additionally, similar improvements are seen for all other \(\lambda\) values that have been tested. For the considerably more complex [cnn]{acronym-label="cnn" acronym-form="singular+short"} classifier, the improvement is not so consistent and already the baseline performance is far better than even the enhanced performance of the [fcn]{acronym-label="fcn" acronym-form="singular+short"} classifier. However, we do obtain a maximal improvement of \(\SI{3+-2}{\percent}\) for \(\lambda = 3\), which also represents the overall highest accuracy for any of the setups investigated here. Finally, for the most complex classifier architecture, the [vit]{acronym-label="vit" acronym-form="singular+short"}, we are not able to show a conclusive improvement of the classifier performance, so that we might expect a dependency of the ability of generated images to add useful information to the training data set on the baseline performance (often connected to the complexity) of the classifier in question. A naïve interpretation could be that the better performing architectures are simply more sensitive to even small differences between the real and generated images. Further, we considered a three-class classification problem with extended sources only. We found that the overall accuracy is reduced as compact sources are easier to classify. More importantly, the significance of the improvement by including generated images in the training is not enhanced as the variations in the cross-validation tend to increase as well.\
The best overall accuracy is obtained by using the [cnn]{acronym-label="cnn" acronym-form="singular+short"} and [wgan]{acronym-label="wgan" acronym-form="singular+short"} augmented training data, but only by a small margin. Yet, we have shown that [wgan]{acronym-label="wgan" acronym-form="singular+short"} augmentation works in principle (similar to the goal in, as noted above) and can significantly improve a somewhat simpler algorithm. This can be useful for applications of classification algorithms in resource-constrained environments, i.e. disk-space and inference time restrictions.\
Given that our generative model is able to generate very quickly large sets of radio galaxy images of different morphologies, [wgans]{acronym-label="wgan" acronym-form="plural+short"} can play an important role in the simulation and analysis of large radio surveys. Future work involving much larger training sets from the LOFAR telescope will explore this further. Moreover, [wgan]{acronym-label="wgan" acronym-form="singular+short"}-generated images can be used to validate new interferometric machine-learning algorithms, see e.g.. To this end, we provide the model and weights with documentation at <https://github.com/floriangriese/wGAN-supported-augmentation>. | {'timestamp': '2022-12-19T02:12:54', 'yymm': '2212', 'arxiv_id': '2212.08504', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08504'} |
null | null |
# Introduction {#sec:intro}
Monads are ubiquitous in both mathematics and computer science, and many different kinds of monads have been considered in various settings. In functional programming, monads are used to capture computational effects. Strong monads have been used to provide semantics of programming languages such as Moggi's computational \(\lambda\)-calculus and models of call-by-push-value. Monads are also used in algebra to represent algebraic theories, and in fact, the class of algebraic theories is equivalent to a class of monads. This result has been adapted to the enriched case as well in order to relate various notions of computations with enriched monads. Comonads, the dual notion of monads, have also found applications: symmetric monoidal comonads are used in the semantics of linear logic. A general setting in which all these different variations of monads can be studied, has been developed by Street. This setting, known as *the formal theory of monads*, uses the fact that the notion of monad can be defined internal to an arbitrary bicategory, including 2-categories (which were used by Street). Each of the aforementioned kinds of monads is actually an instance of this more general notion. For example, monads in the bicategory of symmetric monoidal categories are symmetric monoidal monads, and strong monads are monads in the bicategory of so-called left actegories. Comonads in a bicategory \(\cat{B}\) are the same as monads in \(\cobicat{\cat{B}}\), which is \(\cat{B}\) with the 2-cells reversed. Even several kinds of distributive laws, including mixed distributive laws and iterated distributive laws, are instances of this notion of monad. An overview of the different kinds of monads internal to various bicategories can be found in Table [1](#tab:notion-of-monad){reference-type="ref" reference="tab:notion-of-monad"}. As such, the formal theory of monads provides a good setting to study monads with the right level of abstraction.
Note that we use diagrammatic order for composition instead of compositional order. We use the notation \(\homC{\cat{B}}{x}{y}\) for the category whose objects are 1-cells \(f: x \rightarrow y\) and whose morphisms from \(f: x \rightarrow y\) to \(g: x \rightarrow y\) are 2-cells \(\tau: f \Rightarrow g\). Given a 1-cell \(f: x \rightarrow y\) and an object \(w: \cat{B}\), we have a functor \(\postcomp{f}{w}: \homC{\cat{B}}{w}{x} \rightarrow \homC{\cat{B}}{w}{y}\), which sends a 1-cell \(g: w \rightarrow x\) to \(\comp{g}{f}\) and a 2-cell \(\tau: g_1 \Rightarrow g_2\) to \(\tau \vartriangleright f\). The core example of a bicategory in this paper is \(\cat{UnivCat}\). Its objects are **univalent** categories, the 1-cells are functors, and the 2-cells are natural transformations. We also have a bicategory \(\cat{Cat}\) whose objects are (not necessarily univalent) categories, 1-cells are functors, and 2-cells are natural transformations. In this paper, we also make use of *univalent* bicategories. To define this property, we use that between every two objects \(x, y: \cat{B}\), we have a type \(\adjequiv{x}{y}\) of *adjoint equivalences* between them. In addition, for all 1-cells \(f, g: x \rightarrow y\), there is a type \(\invcell{f}{g}\) of *invertible 2-cells* between them. For the precise definition of these notions, we refer the reader to the literature.
The bicategory \(\cat{UnivCat}\) of univalent categories is both locally and globally univalent. However, \(\cat{Cat}\), whose objects objects are not required to be univalent, is neither. If we have a bicategory \(\cat{B}\), then we define the bicategory \(\opbicat{\cat{B}}\) by 'reversing the 1-cells' in \(\cat{B}\). More precisely, objects are the same as objects in \(\cat{B}\), 1-cells from \(x\) to \(y\) in \(\opbicat{\cat{B}}\) are 1-cells \(y \rightarrow x\) in \(\cat{B}\), while 2-cells from \(f: y \rightarrow x\) to \(g: y \rightarrow x\) are 2-cells \(f \Rightarrow g\) in \(\cat{B}\). In addition, from a bicategory \(\cat{B}\), we obtain \(\cobicat{\cat{B}}\) by 'reversing the 2-cells'. Objects and 1-cells in \(\cobicat{\cat{B}}\) are the same as objects and 1-cells in \(\cat{B}\) respectively, but a 2-cell from \(f\) to \(g\) in \(\cobicat{\cat{B}}\) is a 2-cell \(g \Rightarrow f\) in \(\cat{B}\). Next we define *pseudofunctors*.
In applications, we are interested in a wide variety of bicategories beside \(\cat{UnivCat}\), and among those are \(\UnivCat_{\cat{Terminal}}\) and \(\cat{SymMonUnivCat}\). These examples have something in common: their objects are categories equipped with some extra structure, the 1-cells are structure preserving functors, while the 2-cells are structure preserving natural transformations. We capture this pattern using *displayed bicategories*. This notion is an adaptation of *displayed categories* to the bicategorical setting. To get an idea of what displayed bicategories are, let us first briefly discuss displayed categories. A displayed category \(\cat{D}\) over \(\cat{C}\) represents structure and properties to be added to the objects and morphisms of \(\cat{C}\). For every object \(x: \cat{C}\), we have a type of *displayed objects* \(\cat{D}_x\) and for every morphism \(f: x \rightarrow y\) and displayed objects \(\disp{x}: \cat{D}_x\) and \(\disp{y}: \cat{D}_y\), we have a set \(\dmor{\disp{x}}{\disp{y}}{f}\) of *displayed morphisms*. For example, for \(\cat{C}\) we take the category of sets, and an example of a displayed category would be group structures. The displayed objects over a set \(X\) are group structures over \(X\), while the displayed morphisms over \(f: X \rightarrow Y\) between two group structures are proofs that \(f\) preserves the unit and multiplication. Every displayed category \(\cat{D}\) gives rise to a *total category* \(\total{\cat{D}}\) and a functor \(\proj{\cat{D}}: \total{\cat{D}} \rightarrow \cat{C}\). In the example we mentioned before, \(\total{\cat{D}}\) would be the category of groups and \(\proj{\cat{D}}\) maps a group to its underlying set. In the bicategorical setting, we use a similar approach, but a displayed bicategory should not only have displayed objects and 1-cells, but also displayed 2-cells. More precisely, we define displayed bicategories as follows.
Just like for displayed categories, every displayed bicategory \(\cat{D}\) gives rise to bicategory \(\total{\cat{D}}\) and a pseudofunctor \(\total{\cat{D}} \rightarrow \cat{B}\).
Univalent displayed bicategories are defined similarly to univalent bicategories, and the precise definition can be found elsewhere. If we have a displayed bicategory \(\cat{D}\) over \(\cat{B}\) and if both \(\cat{B}\) and \(\cat{D}\) are univalent, then \(\total{\cat{D}}\) is univalent as well. This gives modular way to construct univalent bicategories. In numerous different examples, we look at displayed bicategories whose displayed 2-cells are actually trivial, in a certain sense. More precisely, we look at two properties. One of them expresses that between all displayed 1-cells \(\disp{f}: \dmor{\disp{x}}{\disp{y}}{f}\) and \(\disp{g}: \dmor{\disp{x}}{\disp{y}}{g}\), there is at most one displayed 2-cell. That property is called *locally propositional*: it expresses that the type \(\dtwo{\disp{f}}{\disp{g}}{\tau}\) is a proposition. The other property is *locally groupoidal*, and it says that every displayed 2-cell over an invertible 2-cell is again invertible.
Note that every displayed bicategory constructed using [\[def:dispcat-to-dispbicat\]](#def:dispcat-to-dispbicat){reference-type="ref" reference="def:dispcat-to-dispbicat"} is both locally propositional and locally groupoidal. Now let us look at some examples of displayed bicategories. These were already considered in previous work.
Note that similarly, one can define a displayed bicategory \(\cat{dSymMonUnivCat}\) over \(\cat{UnivCat}\) whose total bicategory \(\cat{SymMonUnivCat}\) is the bicategory of symmetric monoidal categories. In the remainder of this paper, we use several operations on displayed bicategories.
The last notion we discuss, is the notion of a *section* of a displayed bicategory. Intuitively, a section assigns to every \(x\) a displayed object \(\disp{x}: \cat{D}_x\) in a pseudofunctorial way. As such every section \(s\) of \(\cat{D}\) induces a pseudofunctor \(\Section{s}: \cat{B} \rightarrow \total{D}\) such that every \(x: \cat{B}\) is *definitionally* equal to \(\proj{\cat{D}}(\Section{s}(x))\).
# The Bicategory of Monads {#sec:bicat-mnd}
Two concepts play a key role in the formal theory of monads: the bicategory of monads \(\mnd{\cat{B}}\) internal to \(\cat{B}\) and Eilenberg-Moore objects. In this section, we study the first concept, and we use displayed bicategories to construct the bicategory \(\mnd{\cat{B}}\) given a bicategory \(\cat{B}\). We also show that \(\mnd{\cat{B}}\) must be univalent if \(\cat{B}\) is. The main idea behind the construction is to split up monads in several independent parts. We first define \(\dEndo{\cat{B}}\) whose displayed objects over \(x\) are 1-cells \(e: x \rightarrow x\). After that, we define \(\dUnit{\cat{B}}\) and \(\dMult{\cat{B}}\) whose displayed objects are the unit and multiplication of the monad respectively. We finally take a full subcategory for the monad laws.
Next we define two displayed bicategories over \(\Endo{\cat{B}}\). For both, we use [\[def:dispcat-to-dispbicat\]](#def:dispcat-to-dispbicat){reference-type="ref" reference="def:dispcat-to-dispbicat"}.
Next we define \(\dMndData{\cat{B}}\) to be \(\sigmaD{\dUnit{\cat{B}} \times \dMult{\cat{B}}}\) and we denote its total bicategory by \(\MndData{\cat{B}}\). To obtain \(\mnd{\cat{B}}\), we take a full subbicategory.
Before we continue, let us discuss the objects, 1-cells, and 2-cells of \(\mnd{\cat{B}}\), and fix the relevant notation for the remainder of this paper. This also explains what the displayed objects, 1-cells, and 2-cells in \(\dmnd{\cat{B}}\) are. The data of a monad \(m\) in \(\cat{B}\) consists of
- an object \(\monadob{m}: \cat{B}\);
- a 1-cell \(\monadendofull{m}: \monadob{m} \rightarrow \monadob{m}\);
- a 2-cell \(\monadunit{m}: \id{\monadob{m}} \Rightarrow \monadendo{m}\);
- a 2-cell \(\monadmult{m}: \comp{\monadendo{m}}{\monadendo{m}} \Rightarrow \monadendo{m}\).
If no confusion arises, we write \(\monadendo{m}\) instead of \(\monadendofull{m}\). The data of a monad morphism \(f: m_1 \rightarrow m_2\) between monads \(m_1\) and \(m_2\) consists of a 1-cell \(\monadmorobfull{f}: \monadob{m_1} \rightarrow \monadob{m_2}\) and a 2-cell \(\monadmorendo{f}: \comp{\monadmorob{f}}{\monadendo{m_2}} \Rightarrow \comp{\monadendo{m_1}}{\monadmorob{f}}\). Lastly, the data of monad 2-cell \(\gamma: f_1 \Rightarrow f_2\) between monad morphisms is a 2-cell \(\monadcellobfull{\gamma}: \monadmorob{f_1} \Rightarrow \monadmorob{f_2}\). We write \(\monadmorob{f}\) instead of \(\monadmorobfull{f}\) and \(\monadcellob{\gamma}\) instead of \(\monadcellobfull{\gamma}\) if no confusion arises. Next we look at the univalence of \(\mnd{\cat{B}}\), and to prove it, we use displayed univalence. This way, it suffices to prove the displayed univalence of \(\dEndo{\cat{B}}\), \(\dUnit{\cat{B}}\), and \(\dMult{\cat{B}}\). We also need \(\cat{B}\) to be univalent.
We also characterize invertible 2-cells and adjoint equivalences in \(\mnd{\cat{B}}\).
Note that in [\[prop:mnd-adjequiv\]](#prop:mnd-adjequiv){reference-type="ref" reference="prop:mnd-adjequiv"}, we assume that \(\cat{B}\) is univalent. Because of univalence, it suffices to prove this proposition assuming that \(\monadmorob{f}\) is the identity 1-cell, which simplifies the involved coherencies. The same idea can be used to show that pointwise pseudonatural adjoint equivalences are adjoint equivalences in the bicategory of pseudofunctors.
# Examples of Monads {#sec:examples}
Next we look at examples of monads, and we start by characterizing monads internal to several bicategories. Let us start by observing that monads in the bicategory \(\cat{UnivCat}\) of categories correspond to monads as how they usually are defined in category theory. However, since this notion of monads is defined in *every* bicategory, we can also look at other bicategories, such as \(\cat{SymMonUnivCat}\),and \(\UnivCat_{\cat{Terminal}}\). In a wide variety of applications, one is interested in monads in a bicategory of categories with some extra structure. For example, symmetric monoidal monads are monads internal to the bicategory of symmetric monoidal categories. Strong monads are monads in the bicategory of left actegories. The bicategories in these two examples can be constructed as a total bicategory of some displayed bicategory over \(\cat{UnivCat}\). To characterize monads in total bicategories, we define *displayed monads*.
Note that we do not require any coherencies in [\[def:disp-mnd\]](#def:disp-mnd){reference-type="ref" reference="def:disp-mnd"}, because the involved displayed bicategory needs to be locally propositional.
Displayed monads have also been used by Ahrens, Matthes, and Wullaert to determine monads in the bicategory of left actegories. Analogously, we can characterize monads in \(\cat{SymMonUnivCat}\). Next we look at monads in \(\opbicat{\cat{B}}\) and \(\cobicat{\cat{B}}\).
Analogously, we can show that objects in \(\mnd{\cobicat{\mnd{\cobicat{\cat{B}}}}}\) are *mixed distributive laws* in \(\cat{B}\). We can also look at *iterated distributive laws*, which are monads in \(\mndn{\cat{B}}{n}\). Next we look at two ways to general constructions of monads. First of all, we look at the identity monad: on every object \(x: \cat{B}\), we construct the identity monad \(\idmonad{x}\).
In the remainder, we only use the pseudofunctor arising from [\[constr:sec-mnd\]](#constr:sec-mnd){reference-type="ref" reference="constr:sec-mnd"} and [\[constr:section-to-psfunctor\]](#constr:section-to-psfunctor){reference-type="ref" reference="constr:section-to-psfunctor"}, and this pseudofunctor is denoted as \(\idmonadpsfunctor: \cat{B} \rightarrow \mnd{\cat{B}}\). Second, we notice that every monad \(m: \mnd{\cat{B}}\) gives rise to a monad of categories. This example is used in [5](#sec:em-ob){reference-type="ref" reference="sec:em-ob"}.
Next we show that pseudofunctors preserve monads.
In fact, if we have a pseudofunctor \(F: \cat{B}_1 \rightarrow \cat{B}_2\), then we obtain a pseudofunctor \(\mnd{F}: \mnd{\cat{B}_1} \rightarrow \mnd{\cat{B}_2}\). Next we show that monads can be composed if we have a distributive law between them.
# Eilenberg-Moore Objects {#sec:em-ob}
The second important concept in the formal theory of monads is the notion of *Eilenberg-Moore objects*. An important property of monads in category theory is that every monad gives rise to an adjunction. One can do this in two ways: either via Eilenberg-Moore categories or via Kleisli categories. In this section, we study Eilenberg-Moore objects, which formulate Eilenberg-Moore categories in bicategorical terms. Note that the terminology in this section is slightly differently compared to what was used by Street. Whereas Street would say that a bicategory admits the construction of algebras, we would say that a bicategory has Eilenberg-Moore objects. In addition, our notions are defined slightly differently, but still equivalent to the original definition by Street. The difference comes from the fact that we use *Eilenberg-Moore cones*.
More concretely, an Eilenberg-Moore cone \(e\) for a monad \(m\) consists of
- An object \(\emobfull{e}\);
- A 1-cell \(\emmor{e}: \emobfull{e} \rightarrow \monadob{m}\);
- A 2-cell \(\emcell{e}: \comp{\emmor{e}}{\monadendo{m}} \Rightarrow \comp{\id{x}}{\emmor{e}}\)
such that the following diagrams commutes \[\begin{tikzcd} \emmor{e} & {\comp{\emmor{e}}{\id{\monadob{m}}}} &[2em] {\comp{\emmor{e}}{\monadendo{m}}} \\ && {\comp{\id{x}}{\emmor{e}}} \arrow["\emcell{e}", from=1-3, to=2-3, Rightarrow] \arrow["{\emmor{e} \vartriangleleft \monadunit{m}}", from=1-2, to=1-3, Rightarrow] \arrow["{\rinvunitor{\emmor{e}}}", from=1-1, to=1-2, Rightarrow] \arrow["{\linvunitor{\emmor{e}}}"', from=1-1, to=2-3, Rightarrow] \end{tikzcd}\] \[\begin{tikzcd}[column sep = huge] {\comp{\emmor{e}}{(\comp{\monadendo{m}}{\monadendo{m}})}} &&& {\comp{\emmor{e}}{\monadendo{m}}} \\ {\comp{(\comp{\emmor{e}}{\monadendo{m}})}{\monadendo{m}}} & {\comp{(\comp{\id{x}}{\emmor{e}})}{\monadendo{m}}} & {\comp{\emmor{e}}{\monadendo{m}}} & {\comp{\id{x}}{\emmor{e}}} \arrow["{\lassociator{\emmor{e}}{\monadendo{m}}{\monadendo{m}}}"', from=1-1, to=2-1, Rightarrow] \arrow["{\emcell{e} \vartriangleright \monadendo{m}}"', from=2-1, to=2-2, Rightarrow] \arrow["{\lunitor{\emmor{e}} \vartriangleright \monadendo{m}}"', from=2-2, to=2-3, Rightarrow] \arrow["{\emcell{e}}"', from=2-3, to=2-4, Rightarrow] \arrow["{\emmor{e} \vartriangleleft \monadmult{m}}", from=1-1, to=1-4, Rightarrow] \arrow["{\emcell{e}}", from=1-4, to=2-4, Rightarrow] \end{tikzcd}\] If no confusion arises, we write \(\emob{e}\) instead of \(\emobfull{e}\). Next we look at the universal property of Eilenberg-Moore objects. Since Eilenberg-Moore objects are examples of limits in a bicategories, there are multiple methods to express their universal property. A first possibility, which is used by Street, is to use biadjunctions, and a second option is to write out explicit mapping properties. Alternatively, one could express the universal property as an adjoint equivalence on the hom-categories. We use the last option. To write out the desired definition precisely, we first define a functor with domain \(\homC{\cat{B}}{x}{\emob{e}}\) where \(e\) is an Eilenberg-Moore cone and \(x\) is any object.
Another way to formulate the universality of an Eilenberg-Moore cone, is by using Eilenberg-Moore categories. If we have a monad \(m\) on a category \(\cat{C}\), we write \(\emlim{}{m}\) for the Eilenberg-Moore category of \(m\).
Alternatively, universality can be formulated using mapping properties. These properties are deduced from the fact that every adjoint equivalence is split essentially surjective and fully faithful. More precisely, an Eilenberg-Moore cone \(e\) is universal if and only if
Let us consider some examples of Eilenberg-Moore objects.
One can also show that \(\cobicat{\cat{SymMonUnivCat}}\) has Eilenberg-Moore objects. These are given by Eilenberg-Moore categories of comonads.
# Duality and Kleisli Objects {#sec:duality}
The goal of this section is to construct Eilenberg-Moore objects in \(\opbicat{\cat{UnivCat}}\). To do so, we start by characterizing such objects via *Kleisli objects*.
As such, to find Eilenberg-Moore objects in \(\opbicat{\cat{UnivCat}}\), we need to find Kleisli objects in \(\cat{UnivCat}\). However, before we look those, we look at Kleisli objects in \(\cat{Cat}\). These are constructed via the usual definition of *Kleisli categories*.
Note that even if \(\cat{C}\) is required to be univalent, the Kleisli category \(\kleisli{m}\) is **not** necessarily univalent. As such, to obtain Kleisli objects in \(\cat{UnivCat}\), we need to use an alternative definition for the Kleisli category. First, we define a functor \(\freealg{m}: \cat{C} \rightarrow \eilenbergmoore{m}\) which sends objects \(x: \cat{C}\) to the algebra \(\monadmult{m}(x): \monadendo{m}(\monadendo{m}(x)) \rightarrow \monadendo{m}(x)\) and morphisms \(f: x \rightarrow y\) to \(\monadendo{m}(f): \monadendo{m}(x) \rightarrow \monadendo{m}(y)\). Note that this 1-cell can actually be defined in arbitrary bicategories (see [7](#sec:adj){reference-type="ref" reference="sec:adj"}). By taking the full image of this functor, we obtain the category \(\univkleisli{m}\).
- Objects of \(\univkleisli{m}\) are pairs \(y: \cat{C}\) together with a proof of \(\trunc{\sigmatype{x}{\cat{C}}{m(x) \cong y}}.\)
- Morphisms from \(y_1: \univkleisli{m}\) to \(y_2: \univkleisli{m}\) are morphisms \(y_1 \rightarrow y_2\) in \(\cat{C}\).
If \(\cat{C}\) is univalent, then \(\emlim{\cat{UnivCat}}{m}\) is univalent, and thus \(\univkleisli{m}\) is so as well.
In univalent foundations, not every functor that is both fully faithful and essentially surjective is automatically an adjoint equivalence as well. This statement only holds if the domain is univalent. However, we can still use the functor \(\kleislifunctor{m}\) to deduce the universal property of \(\univkleisli{m}\). For that, we use Theorem 8.4 in.
# Monads and Adjunctions {#sec:adj}
The cornerstone of the theory of monads is the relation between monads and adjunctions. More specifically, every adjunction gives rise to a monad and vice versa. This was generalized by Street to 2-categories that have Eilenberg-Moore objects. In this section, we prove these theorems, and to do so, we start by recalling adjunctions in bicategories.
Our notation for adjunctions is taken from. The two coherencies in [\[def:adjunction\]](#def:adjunction){reference-type="ref" reference="def:adjunction"} are called the *triangle equalities*. As expected, adjunctions internal to \(\cat{UnivCat}\) correspond to adjunctions of categories. This is because the unitors and associators in \(\cat{UnivCat}\) are pointwise the identity, so the triangle equalities in [\[def:adjunction\]](#def:adjunction){reference-type="ref" reference="def:adjunction"} reduce to the usual ones.
[\[exa:adj-dual\]](#exa:adj-dual){reference-type="ref" reference="exa:adj-dual"} can be strengthened by using the terminology of *left adjoints* and *right adjoints*. Given a 1-cell \(f: x \rightarrow y\), the type \(\leftadj{\cat{B}}{f}\) says that we have \(r\),\(\eta\), and \(\varepsilon\) such that we have an adjunction \(\adjunction{l}{r}{\eta}{\varepsilon}\). The type \(\rightadj{\cat{B}}{f}\) is defined analogously. Now we can reformulate [\[exa:adj-dual\]](#exa:adj-dual){reference-type="ref" reference="exa:adj-dual"} as follows: we have equivalences \(\typeequiv{\leftadj{\opbicat{\cat{B}}}{f}}{\rightadj{\cat{B}}{f}}\) and \(\typeequiv{\leftadj{\cobicat{\cat{B}}}{f}}{\rightadj{\cat{B}}{f}}\) of types. Next we look at *displayed adjunctions*, which we use to obtain adjunctions in total bicategories. This notion is used to characterize adjunctions in bicategories such as \(\UnivCat_{\cat{Terminal}}\) and \(\cat{SymMonUnivCat}\).
Analogously, we characterize adjunctions in \(\cat{SymMonUnivCat}\). Now we have developed enough to state and prove the core theorems of the formal theory of monads. These theorems relate adjunctions and monads, and we first prove that every adjunction gives rise to a monad.
Since by [\[exa:adj-dual,exa:mnd-dual\]](#exa:adj-dual,exa:mnd-dual){reference-type="ref" reference="exa:adj-dual,exa:mnd-dual"} adjunctions and monads in \(\cobicat{\cat{B}}\) correspond to adjunctions and comonads in \(\cat{B}\) respectively, we get that every adjunction in \(\cat{B}\) induces a comonad by [\[constr:adj-to-mnd\]](#constr:adj-to-mnd){reference-type="ref" reference="constr:adj-to-mnd"}. Next we look at the converse: obtaining adjunctions from monads. For this, we need to work in a bicategory with Eilenberg-Moore objects. We show that every monad \(m\) gives rise to an adjunction and that the monad coming from this adjunction is equivalent to \(m\).
Note that we can instantiate [\[constr:mnd-to-adj\]](#constr:mnd-to-adj){reference-type="ref" reference="constr:mnd-to-adj"} to several concrete instances.
- Since \(\cat{UnivCat}\) has Eilenberg-Moore objects by [\[ex:em-cats\]](#ex:em-cats){reference-type="ref" reference="ex:em-cats"}, we get the usual construction of adjunctions from monads via Eilenberg-Moore categories.
- Since \(\opbicat{\cat{UnivCat}}\) has Eilenberg-Moore objects by [\[constr:univ-kleisli-cats\]](#constr:univ-kleisli-cats){reference-type="ref" reference="constr:univ-kleisli-cats"}, every monad gives rise to an adjunction via Kleisli categories.
One can also show that \(\cobicat{\cat{SymMonUnivCat}}\) has Eilenberg-Moore objects, and thus every comonad of symmetric monoidal categories gives rise to an adjunction. This is used to construct models of linear logic.
# Monadic Adjunctions {#sec:monadic}
By [\[constr:mnd-to-adj\]](#constr:mnd-to-adj){reference-type="ref" reference="constr:mnd-to-adj"} we have an equivalence \(\adjtomnd{\mndtoadj{m}} \simeq m\) for every monad \(m\). However, such a statement does not hold for adjunctions. *Monadic adjunctions* are the adjunctions for which we do have such an equivalence.
Next we look at a representable version of this definition. More specifically, we define monadic 1-cells using monadic functors in \(\cat{UnivCat}\). To do so, we first show that every adjunction gives rise to an adjunction on the hom-categories.
Note that we require the bicategory in [\[def:repr-monadic\]](#def:repr-monadic){reference-type="ref" reference="def:repr-monadic"} to be locally univalent, so that each hom-category lies in \(\cat{UnivCat}\). Now we show that these two notions of monadicity are equivalent. We first prove the following lemma.
# Conclusion {#sec:concl}
We studied Street's formal theory of monads in this paper. We saw that it provides a good setting to study monads in univalent foundations, because it allows us to prove the core theorems in arbitrary bicategories instead of only for categories. For that reason, it helps us with concrete problems, such as constructing an adjunction from a monad using the univalent version of the Kleisli category. This is because one only needs to prove a universal property instead of reproducing the whole construction of the adjunction. There are numerous ways to continue this line of research. One result that is missing, is Theorem 12 from Street's paper about comonads. In addition, the work in this setting provides a framework in which one can study numerous applications, such as models of linear logic, Moggi-style semantics, and call-by-push-value, and formalizing these applications would be a worthwhile extension. Finally, one could study extensions of the formal theory to a wider class of monads, such as graded monads. | {'timestamp': '2022-12-19T02:13:15', 'yymm': '2212', 'arxiv_id': '2212.08515', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08515'} |
null | null |
# Introduction
Bounding-box based person detection using neural networks is one of the most covered topics in computer vision literature which is a strong indication of its importance for many applications. However, most studies assume moderate to high quality sensors and powerful processors like GPUs with lots of resources to work with, which comes with bulky hardware, high power consumption and a big price tag in the order of thousands of dollars. Moreover, high quality sensors such as color cameras may reveal privacy-sensitive information about persons, leading to privacy concerns and mistrust in technology. In many real-life applications, such problems are unacceptable and therefore low-resolution sensors and embedded GPUs, low-power accelerators or even standard microcontrollers (MCUs) are preferred, costing rather hundreds of dollars, tens of dollars or even a few dollars, respectively. With an exponential decrease in price, comes an exponential decrease in computational resources, which makes deployment of neural networks increasingly complex, especially for MCUs. This requires smart solutions that work with low-resolution sensors, sparse model architectures and state-of-the-art compression technologies to squeeze out every last redundant computation, which in turn poses challenges in retaining sufficient prediction accuracy. Nonetheless, this work succeeds in building a highly accurate person detection system with a low-cost MCU, the cheapest category of processors, a low-cost sensor and a highly optimized deep-learning based detection algorithm. Due to the thermal spectrum, combined with ultra low-resolution, it must be noted that the identity of persons in front of the camera is guaranteed to be preserved. Our methodology, depicted in Figure [\[fig:overview\]](#fig:overview){reference-type="ref" reference="fig:overview"}, consists of a custom CNN-based single-stage object detector and a unique video background subtraction algorithm to help distinguish static person-like objects from real persons. Given the fact that deep learning is skilled at extracting useful context from the background of an image, the combination of deep learning and background subtraction might seem odd at first glance. However, this is not the case when working with such low-resolution (thermal) images, since background features from our sensor are often indistinguishable from a person, even for a human observer. Figure [\[fig:with_and_without_bg_sub\]](#fig:with_and_without_bg_sub){reference-type="ref" reference="fig:with_and_without_bg_sub"} depicts a few of such difficult examples, with and without our background subtraction applied. Our contributions are summarized as follows:
- This work proposes a truly low-cost and privacy preserving detection system for stationary setups, featuring a novel detection algorithm.
- We introduce a benchmarking system to compare the accuracy of our model against the baseline--the standard person detection software from Melexis, the manufacturer of the MLX90640 thermal sensor--and prove that our models outperform the former, achieving up to 15.6% higher in \(F_1\)-score.
- We are the first to prove the advantage of background-subtraction in combination with a CNN for ultra low-resolution thermal object detection.
- We present extensive compression experiments and a benchmark on inference time and memory utilization of our models deployed with Tensorflow Lite or microTVM on two MCU targets.
- We release our dataset in order to support future research[^1].
Our paper is structured as follows: Section [2](#sec:related work){reference-type="ref" reference="sec:related work"} discusses related work on person detection and neural network compression, given their direct relevance to our application, Section [3](#sec:approach){reference-type="ref" reference="sec:approach"} elaborates on our detection algorithm, the proposed compression method and how we collected our dataset, Section [4](#sec:experiments){reference-type="ref" reference="sec:experiments"} discusses our results while conclusions are made in final Section [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}.
# Related Work {#sec:related work}
## Person Detection
Since the dawn of the deep-learning area, person detection has been addressed mostly with neural network based object detectors such as (Fast(er)) R-CNN, YOLO and SSD. In a constant battle to improve the accuracy and reducing the computational budget, many new designs and improvements emerged in later years. Other attempts for improving the accuracy focus on (1) exploiting temporal queues in video object detection and (2) using different sensor modalities. Video object detection or recognition methods exploit the temporal domain to improve the detection accuracy. Liu et al. for example, propose to use a convolutional LSTM module into their SSD object detector, Simonyan et al. use an additional optical-flow input image in their two-stream approach, Li et al. experiment with 3D convolutions in their video-based vehicle detector, and Kang et al. use a tubelet tracking mechanism to improve their detection results. Different sensor modalities are incorporated in several works for more robust detection in difficult viewing circumstances. Experiments with time-of-flight cameras, thermal cameras or a combination of different sensors can be found in the literature. Although low-resolution object detection with deep-learning has been attempted down to 96x96 pixels, or even lower with the help of super-resolution, we are the first to try this on 32\(\times\)`<!-- -->`{=html}24 pixels in a direct way. A method for counting people in a detection-like way with the same sensor as ours exists, however their dataset seems less challenging (no person-like background objects) and their model is much heavier (130k params) compared to ours (10k params). Constrained by the very limited resources of regular MCUs, we base our model on tiny YOLOv2, which is small and much more scalable compared to two-stage detectors or multi-headed detectors like SSD or later YOLO versions. Even though we process video, spatio-temporal building blocks with a large memory footprint like 3D convolutions are to be avoided, together with convolutional LSTMs, given that the latter are difficult to tune and make compression extremely complex. We however included experiments with a motion image in Section [4.2](#sec:results_bg_sub){reference-type="ref" reference="sec:results_bg_sub"}, because it's cheap to calculate.
## Model Compression
Although several efficient CNN architectures like MobileNet or EfficientNet emerged from the need for lighter models, additional compression is often required to meet model size, computational budget, energy, or time constraints. The most popular compression techniques can be divided in three categories: (1) model quantization, (2) model pruning and (3) Network-Architecture-Search (NAS). Quantization approaches come in two flavors: (1) Post-Training Quantization (PTQ) and (2) Training-Aware Quantization (QAT). PTQ is the easiest and most widely used quantization approach for obtaining 8-bit CNNs and is a standard feature in most deployment frameworks like Tensorflow Lite or TensorRT, while QAT is the preferred choice for sub-8-bit precision. Pruning methods can be divided in two categories: (1) unstructured pruning and (2) structured pruning. Unstructured pruning methods aim to remove unnecessary neural connections by setting individual weight values to zero. However, this introduces weight sparsity which requires special libraries to support the acceleration of such models. Structured pruning methods like filter pruning on the other hand, remove entire convolution filters at once and avoid the need for special acceleration libraries or hardware. Different methods are proposed to identify good filter candidates for pruning, ranging from methods that remove filters with the smallest \(L_1\) or \(L_2\) filter norms, to more complex ones that for example require changes in the loss function. NAS has been adopted by several works to find small architectures that can be directly deployed on resource constrained devices. Although such techniques show remarkable performance, their complexity is not needed in our work. Instead, we prefer iterative channel pruning with simple norm-based saliency, which has been proved by Ophoff et al. to work surprisingly well for constrained object detection problems. We believe that our approach can be defined as a constrained object detection problem because: (1) we only have a single class, (2) we target fixed camera viewpoints and (3) most importantly, our background subtraction method greatly reduces the background variance in the input. These constraints allow simple, but well proven structured pruning methods, to compress our model by more than a factor \(\times\)`<!-- -->`{=html}100. In addition, we also quantize our models to 8-bit through regular PTQ.
# Approach {#sec:approach}
In order to train a CNN-based person detector, a large scale dataset is needed. Since no person detection datasets of our target sensor (MLX90640) are publicly available, we recorded and annotated our own dataset, which is described in more detail in Section [3.1](#sec:approach_dataset){reference-type="ref" reference="sec:approach_dataset"}. Section [3.2](#sec:approach_model){reference-type="ref" reference="sec:approach_model"} describes our CNN-based detector and its background subtraction technique, and Section [3.3](#sec:approach_compression){reference-type="ref" reference="sec:approach_compression"} elaborates on the proposed model compression method.
## Dataset {#sec:approach_dataset}
A large bounding-box annotated dataset is constructed of 190 lengthy video clips at 8FPS, containing 96k thermal video frames in total. Recordings are made using a static camera setup that is mounted on the ceiling in 12 different locations including offices, residential rooms and laboratory rooms. Each video clip is labeled with a *45-degree* or *90-degree* tag, indicating that the setup is pointed straight down or at an angle of approximately 45 degrees, respectively, as shown in Figure [\[fig:recording_angles\]](#fig:recording_angles){reference-type="ref" reference="fig:recording_angles"}. Video frames are stored as 16-bit.tiff images containing the raw temperature measurements in degrees Celsius, multiplied by 100. We split our dataset in train, validation and test set, where the videos of the validation and test sets are recorded at different locations compared to the videos of the training set, in order to avoid overfitting. To help automate the annotation process, existing person detection software, based on Mask-RCNN and an object tracker, is used to track persons in a high resolution color video feed, coming from a webcam that is included in the recorder setup. Since the recording angle and field-of-view of both cameras have been made equal, the generated tracking data from the webcam can be used as annotation data for the thermal video, after careful manual inspection and correction of the tracking data using the CVAT annotation tool. Figure [\[fig:dataset_recorder\]](#fig:dataset_recorder){reference-type="ref" reference="fig:dataset_recorder"} depicts our recorder and an observation example of both sensors.
## Model {#sec:approach_model}
Our proposed model, illustrated in Figure [\[fig:model\]](#fig:model){reference-type="ref" reference="fig:model"}, consists of a simple CNN with a YOLOv2 detection head and a background subtraction algorithm. First, a constructed background image \(B \in \mathbb{R}^{H\times W}\) is subtracted from the current normalized input image \(I_0 \in \mathbb{R}^{H\times W}\). Second, the result is sent through the CNN backbone, consisting of a single 3\(\times\)`<!-- -->`{=html}3 convolution followed by 7 depthwise-separable convolutions. Finally, the YOLOv2 head, a 1\(\times\)`<!-- -->`{=html}1 convolution followed by its typical post-processing steps, produces the bounding boxes. The following subsections (1) motivate the addition of a background subtraction algorithm, while discussing its implementation and (2) elaborate on the design choices of our CNN architecture.
### Background Subtraction.
This allows the model to make a clear distinction between stationary objects with a person-like heat signature and actual persons. Distinguishing a person from a computer for example on a single 32\(\times\)`<!-- -->`{=html}24 pixel thermal image, is often impossible without further context, even for a human observer. The top row images from Figure [\[fig:with_and_without_bg_sub\]](#fig:with_and_without_bg_sub){reference-type="ref" reference="fig:with_and_without_bg_sub"} depict a few of these difficult samples from our dataset to illustrate this. In such cases, the only difference between both objects is that a person moves from time to time, while a stationary object does not. Based on this assumption, a robust distinction can be made by maintaining an up-to-date background image of stationary features. Our experiments in Section [4.2](#sec:results_bg_sub){reference-type="ref" reference="sec:results_bg_sub"} prove that this method gives an additional boost of up to 28% in AP detection accuracy. Our proposed background updating algorithm, given by Algorithm [\[alg:background_update\]](#alg:background_update){reference-type="ref" reference="alg:background_update"}, uses the current input image \(I_0\), the current background image \(B\) and the detections \(D\) from the current input image to produce the updated background image \(B^*\). First, all bounding boxes in \(D\) are enlarged by one pixel in all directions. Second, a binary mask \(M\) is constructed that contains white pixels at locations that overlap with one or more bounding boxes from \(D\), and black pixels otherwise. Third, a background candidate \(\hat{B}\) is created from background pixels of \(I_0\) and foreground pixels of \(B\), and finally the new background \(B^*\) is created through an exponential-moving-average filter with decay factor \(\alpha = 0.99\). Figure [\[fig:with_and_without_bg_sub\]](#fig:with_and_without_bg_sub){reference-type="ref" reference="fig:with_and_without_bg_sub"} shows the qualitative effect of our background subtraction algorithm on a few samples from our dataset.
During initialization, the first background image is constructed by averaging three consecutive frames, in the assumption that no persons are present at startup. Additional computations for calculating the background are negligible, since this happens only once every 25 frames. During training, ground truth annotations are used to construct the background image instead of detections.
### Network Architecture.
Our network architecture, is inspired by tiny YOLOv2 because we value some of its design choices: (1) a single headed output, which simplifies the network graph and its post-processing, (2) a small amount of layers, which significantly reduces the time overhead for calling the layers, and (3) no residual connections, other branching structures or exotic layer types to simplify compression and deployment. Moreover, the number of layers and number of output channels per layer of our uncompressed model are the same as in tiny YOLOv2 (details in Figure [\[fig:model\]](#fig:model){reference-type="ref" reference="fig:model"}). We however prefer depthwise-separable convolutions in order to create a sparser model to start with, therefore reducing the burden on the compression stage. ReLU6 is preferred over leaky ReLU since it can be fused into its preceding convolution layer in most deployment frameworks. To retain sufficient spatial output resolution, our model down-samples the input resolution by a factor four, in contrast to the more commonly used factor 32 in other higher resolution object detectors. Our model has \(num\_anchors \times 5 = 5 \times 5 = 25\) output channels with anchor box sizes that are adjusted to our dataset.
## Compression {#sec:approach_compression}
We compress our model in two stages: first, iterative channel pruning is applied, followed by post-training quantization to 8-bit for both weights and activations.
### Channel Pruning.
We propose structured channel pruning, which avoids the need for libraries with sparsity support on the target MCU and compresses the activation tensors as well to reduce RAM memory. Each iteration, the 5% filters with the lowest normalized \(L_2\)-norm, as proposed by Ophoff et al., are removed, followed by a fine-tuning step. When fine-tuning is finished, the weights with the lowest validation loss are selected to initialize the model for the next iteration. To speed up the pruning process, we adopt the following early stopping criteria: given the validation loss of the current pruned model \(L_{curr}\) and the validation loss of the unpruned model \(L_{start}\), we skip or stop fine-tuning when \(L_{curr} \le 1.03 \times L_{start}\) and immediately continue with the next iteration.
### Quantization.
After channel pruning is completed, PTQ is applied to convert both the model's weights and activations to 8-bits. We use the Tensorflow Lite model converter for this task and provide a calibration set of 100 images to estimate the scales and zero-points of the activation tensors.
# Experiments {#sec:experiments}
In the first section of our experiments (Section [4.1](#sec:results_comparison){reference-type="ref" reference="sec:results_comparison"}), we compare our models accuracy-wise against the baseline, which is the proprietary in-house software from Melexis. Section [4.2](#sec:results_bg_sub){reference-type="ref" reference="sec:results_bg_sub"} studies the added value of our background subtraction method, Section [4.3](#sec:results_compress){reference-type="ref" reference="sec:results_compress"} elaborates on the compression ratios and accuracy losses and final Section [4.4](#sec:results_deploy){reference-type="ref" reference="sec:results_deploy"} compares the time performance and resource occupation of our models, deployed with two different run-time frameworks on two different MCU targets.
## Comparison {#sec:results_comparison}
In this section we compare the accuracy of our detection model against the accuracy of our baseline, the in-house developed detection software by Melexis, the manufacturer of the MLX90640. Melexis's proprietary detection software is built and perfected by an in-house development team, over the course of two years. Their algorithm uses dynamic noise suppression, contrast and edge enhancement in pre-processing, and calculates a dynamic threshold to identify local maxima and minima after background subtraction. A multi-object tracker is used to further improve the accuracy. Compared to a 90-degree or overhead camera view, detecting persons in video from a 45-degree camera angle turns out to be much more difficult. Therefore, accuracy results are reported separately for both views throughout the whole experiments section.
[\[tab:result_comparison\]]{#tab:result_comparison label="tab:result_comparison"}
Table [1](#tab:result_comparison){reference-type="ref" reference="tab:result_comparison"} presents the \(F_1\)-scores on the test set of our model in uncompressed and compressed state, and the baseline. For both camera angles, our models outperform Melexis's software by 19.2% and 2.7% on the 45-degree and 90-degree test sets, respectively. In compressed state, where our model is heavily pruned and quantized to 8-bit, this still results in a performance boost of 15.6% and 2.6% on the 45-degree and 90-degree test sets, respectively, making our approach the preferred solution. The Average-Precision (AP) metric, which is calculated from a Precision-Recall curve, is popular in most object detection papers, but requires a model to be able to produce a prediction probability for each produced bounding-box. Since Melexis's software is not capable of producing such probabilities, it does not make sense to use this metric here. We therefore prefer \(F_1\)-score instead, which is a percentile number produced from a single precision and a single recall value. Selecting a single precision-recall point for a detector like ours requires selecting a single threshold value. We find the optimal threshold by maximizing the \(F_1\)-score through adjustment of the threshold value. We use a standard 0.5 overlap threshold for matching detections to ground-truth boxes and set the NMS overlap threshold to 0.3.
## Background Subtraction {#sec:results_bg_sub}
This section presents results on the added value of our proposed background subtraction method. Figure [\[fig:results_45_degree_models\]](#fig:results_45_degree_models){reference-type="ref" reference="fig:results_45_degree_models"} and [\[fig:results_90_degree_models\]](#fig:results_90_degree_models){reference-type="ref" reference="fig:results_90_degree_models"} depict the PR-curve charts of a number of models tested on 45-degree and 90-degree labeled dataset images, respectively. Models that are tested on the 90-degree test set are trained on the 90-degree training set, while model tested on the 45-degree test set are trained on both 90-degree and 45-degree training data. Each chart presents results of four different models: (1) *reference*, which is our model without background subtraction, (2) *frame diff*, which is the reference model, but with an additional difference image on a second input channel, (3) *bg sub*, which is our proposed model with background subtraction and (4) *bg sub + frame diff*, which is our model with background subtraction and the additional difference image. Inspired by the optical-flow input from Simonyan et al., we experiment with a difference image, which is a computationally cheap alternative to optical-flow and a way to inject short-term motion information into the network. We construct the difference image by subtracting a previous image from the current image, where we set the frame stride between both images equal to five. From our results, it can be concluded that this difference image seems to help improve the accuracy a little when applied to the *reference* model on the 90-degree dataset. However, the added value of the difference image diminishes in comparison to our proposed background subtraction method, which boosts the accuracy by 28% and 21% AP on the 45-degree and 90-degree datasets, respectively. Since the background subtraction method eliminates the confusion between a person and an imposter background object, the network gets much less confusing samples during training, resulting in such a tremendous performance boost. The *bg sub + frame diff* models however did not improve on the *bg sub*, so the difference image is left out from our proposed method. All models in Figure [\[fig:results_models\]](#fig:results_models){reference-type="ref" reference="fig:results_models"} are directly trained on our dataset from a random initialization for up to 50k iterations (\(\le\)`<!-- -->`{=html}3 hours on a GTX1080), using the SGD optimizer with a high weight decay \(d=0.03\) to reduce overfitting. The first 1000 iterations, a warm-up stage exponentially increases the learning rate up to the base learning rate \(lr=0.001\). Subsequently, a reduce-on-plateau learning rate scheduler decays the learning rate with factor 10 whenever the validation loss did not drop in the previous 5000 iterations. For data augmentation, we use random contrast, brightness, horizontal and vertical flipping manipulations.
## Model Compression {#sec:results_compress}
The compression results of our 45-degree and 90-degree models are presented in Table [2](#tab:results_compression){reference-type="ref" reference="tab:results_compression"}. For each model, the AP and \(F_1\) accuracies are given for the original model, the pruned model and the pruned + quantized model. We start with a 1.26M parameter model and prune that down to around a 10k parameters model, with negligible loss in accuracy for the 90-degree model and acceptable loss in accuracy for the 45-degree model. This results in pruning rates of up to \(\div\)`<!-- -->`{=html}136 and \(\div\)`<!-- -->`{=html}52 for the parameters and MACS, respectively. Since standard PTQ to 8-bit weights and activations hardly influences the accuracy, there is no need for QAT.
[\[tab:results_compression\]]{#tab:results_compression label="tab:results_compression"}
Both models are pruned for up to 60 iterations on a GTX1080, which completes in about 1.5 days. We then select the model with the highest test accuracy that is close to 10k parameters. The 45-degree and 90-degree models that we selected are pruned for 44 and 49 iterations, respectively. Each fine-tuning step takes at most 10k iterations, which is 1/5^th^ of a regular training time. The learning rate is set to 0.0001 and lowered by a factor 10 after 5k iterations.
## Deployment {#sec:results_deploy}
After pruning and quantization, we deploy both 45-degree and 90-degree models on two popular low-cost MCUs: (1) the STM32F746 with a Cortex-M7 core and (2) the STM32F407 with a Cortex-M4 core. We compare the inference time and memory utilization between two open-source neural-network deployment frameworks for our application: Tensorflow Lite for microcontrollers (TFLite) and microTVM. We configure both frameworks to use the CMSIS-NN microkernel library as a backend, because it supports SIMD instructions that can execute four 8-bit multiply-accumulate operations in a single cycle, which significantly boosts the performance. For our benchmark, we compile the generated C-code with all optimizations for speed enabled (`GCC-O3`) and set the clock speeds of both platforms to their maximum, which is 216MHz for the STM32F746 and 168MHz for the STM32F407. Figure [\[fig:results_infer_times\]](#fig:results_infer_times){reference-type="ref" reference="fig:results_infer_times"} presents the inferences times of our models (*Model TFLite/Model TVM*), sensor acquisition time (*Read sensor*) and other processing (*Pre + post proc*), which includes pixel temperature calculation, input normalization, yolo post-processing, NMS and background calculation. All our models are capable of running in real-time (\(\ge\) 8 FPS) on both microcontrollers with sufficient headroom for doing other tasks. Figure [\[fig:results_ram_usage\]](#fig:results_ram_usage){reference-type="ref" reference="fig:results_ram_usage"} and [\[fig:results_flash_usage\]](#fig:results_flash_usage){reference-type="ref" reference="fig:results_flash_usage"} report the RAM memory and flash memory usage, respectively. The utilization of RAM memory in our microTVM experiments is noticeably lower compared to that of Tensorflow Lite, but the biggest difference can be seen in flash-memory usage, where microTVM uses less than half the size compared to Tensorflow Lite. This is because Tensorflow Lite's model interpreter engine is a fixed size component that becomes significantly large when working with very small models like ours. In contrast, microTVM generates direct function calls to the CMSIS-NN micro kernels without the need for interpreter code, which tremendously reduces the flash usage. Note that the statistics only report the used resources of the model.
# Conclusion {#sec:conclusion}
Although person detection is a wanted asset in many real-life applications, it is not always feasible due to (1) low-cost constraints and (2) privacy related issues if a regular camera is involved. This work solves both problems by proposing a low-cost detection system in the order of magnitude of tens of dollars, based on a privacy-preserving ultra low-resolution thermal imager and a low-cost microcontroller. Even though it is extremely difficult to create an accurate deep-learning model that works with such limited resources, our compressed models achieve an accuracy of 79.9% and 91.62% (F1-score) on our 45-degree and 90-degree benchmark test, respectively, outperforming the standard person detection software of Melexis by a significant margin. We achieve our goal by proposing a smart background-subtraction mechanism that eliminates confusion between person-like objects and real persons, boosting our model's performance up to 28%. Our processing pipeline with 90-degree model is running at 46ms/image and 87ms/image on an STM32F746 and STM32F407, respectively, and only requires 34kB of flash memory and 31kB of RAM. In future work, an additional spatio-temporal module could be added to potentially further improve the accuracy, or more extreme compression could be used to further down-size the model. | {'timestamp': '2022-12-19T02:10:27', 'yymm': '2212', 'arxiv_id': '2212.08415', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08415'} |
null | null |
null | null |
# Appendices {#appendices .unnumbered} | {'timestamp': '2022-12-19T02:10:50', 'yymm': '2212', 'arxiv_id': '2212.08429', 'language': 'ja', 'url': 'https://arxiv.org/abs/2212.08429'} |
null | null |
null | null |
null | null |
# Introduction
It was recently pointed out that by combining the distance conjectures of the Swampland program with the smallness of the dark energy in Planck units (\(\Lambda \sim 10^{-120} M_{\rm Pl}^4\)) and confronting these ideas to experiment lead to the prediction of a compact dark dimension with characteristic length-scale in the micron range . The dark dimension opens up at the characteristic mass scale of the Kaluza-Klein (KK) tower, \(m_{\rm KK} \sim \Lambda^{1/4}/\lambda\), where physics must be described by a 5-dimensional theory up to the "species scale", \(M_{\rm UV} \sim (8\pi)^{1/3} \ \lambda^{-1/3} \ \Lambda^{1/12} \ M_{\rm Pl}^{2/3}\), which can be regarded as the higher dimensional Planck scale , with \(10^{-1} \alt \lambda \alt 10^{-4}\). The dark dimension model carries with it a rich phenomenology . In particular, it was observed in that the universal coupling of the Standard Model (SM) fields to the massive spin-2 KK excitations of the graviton in the dark dimension gives an unavoidable dark matter candidate. Complementary to the dark gravitons, it was discussed in that primordial black holes with Schwarzschild radius smaller than a micron could also be good dark matter candidates, possibly even with an interesting close relation to the dark gravitons . In the dark dimension graviton (DDG) model the cosmic evolution of the hidden sector is primarily dominated by "dark-to-dark" decays, yielding a specific realization of the dynamical dark matter framework . The tower of KK states propagating in the incredible bulk provide a rich playground for novel cosmological signals. Herein, we confront DDG predictions to experiment and demonstrate it may represent a viable alternative to the \(\Lambda\) cold-dark-matter (CDM) model. The layout of the paper is as follows. In Sec. [2](#sec:2){reference-type="ref" reference="sec:2"} we show that the modulation of redshifted 21-cm lines driven by \({\rm KK} \to \gamma \gamma\) is within the reach of next generation experiments. After that, we confront predictions of the DDG model with null results of indirect dark matter searches targeting dwarf spheroidal and cluster of galaxies. These astrophysical environments are the best promising targets to search for signals of decaying dark matter in the local universe, because they have a low content of gas and dust, as well as a high mass-to-light ratio, and they are free of astrophysical photon-emitting-sources. Besides, dwarf spheroidals are also relatively nearby and many lie far enough from the Galactic plane to have low Galactic foreground. In Sec. [3](#sec:3){reference-type="ref" reference="sec:3"} we explore the global structure of the inflationary phase within the context of the dark dimension. We demonstrate that the model parameters required for a successful uniform inflation driven by a 5-dimensional cosmological constant (corresponding to a flat region of the 5-dimensional potential) are natural. The paper wraps up with some conclusions presented in Sec. [4](#sec:4){reference-type="ref" reference="sec:4"}. The article also contains an Appendix in which we briefly discuss neutrino masses, mixing, and oscillations.
# Probing the dark dimension with 21-cm cosmology {#sec:2}
During the emission of the cosmic microwave background (CMB) at redshift \(z_{\rm CMB} \sim 1100\), matter dominates the energy density of the Universe. The number density of baryons is mostly composed of neutral hydrogen atoms (\({\rm H}_{\rm I}\)), together with a smaller Helium (He) component, \(x_{\rm He} = n_{\rm He}/n_{{\rm H}_{\rm I}} \simeq 1/13\), and a small percentage of free protons and electrons, \(x_e = n_e /n_{{\rm H}_{\rm I}} = n_p /n_{{\rm H}_{\rm I}}\), which varies from about 20% at \(z_{\rm CMB}\) to roughly \(2 \times 10^{-4}\) at \(z \sim 20\) . After the gas thermally decouples from the photon temperature at \(z_{\rm dec} \sim 150\), most of the hydrogen gas is in its ground state, whose degeneracy is only broken by the hyperfine splitting of the spin-0 singlet and spin-1 triplet. In the rest frame of the gas, the energy gap between these two spin states is \(E_{21} = 5.9 \times 10^{-6} {\rm eV} \simeq 2\pi/21~{\rm cm}\), corresponding to a \(\nu_{21} = 1.4~{\rm GHz}\) spectral line. The relative number density of the two spin levels, \(n_1/n_0 = 3 e^{-E_{21}/T_s}\), defines the spin temperature \(T_s\), where the factor of 3 comes from the degeneracy of the triplet excited state. Because of the 21-cm transitions, for a given redshift \(z < z_{\rm dec}\), a shell of \(H_{\rm I}\) can act as a detector of the background photons . The observable is the 21-cm brightness temperature relative to the photon background \[T_{21} (z) \simeq 27~{\rm mK} \ x_{{\rm H}_{\rm I}} (z) \ \left(\frac{0.15}{\Omega_m h^2} \right)^{1/2} \ \left(\frac{1+z}{10}\right)^{1/2} \ \left(\frac{\Omega_b h^2}{0.02} \right) \ \left(1-\frac{T_\gamma(z)}{T_s (z)} \right) \,, \label{T21}\] where \(\Omega_m\) and \(\Omega_b\) are the the present day values of the non-relativistic matter density and baryon energy density as a fraction of the critical density, and where \(h\) is the Hubble constant constant in units of \(100~{\rm km/s/Mpc}\) . Note that if the photon temperature exceeds the spin temperature (\(T_{21} < 0\)) there will be net absorption and in the opposite case (\(T_{21} > 0\)) net emission. For \(\Lambda\)CDM, \(T_\gamma\) is given by the CMB thermal radiation, with temperature \(T_{\rm CMB}(z) = 2.725~{\rm K} \ (1+z)\). From the Dark Ages (\(1100 \alt z \alt 30\)) to Cosmic Dawn (\(30 \alt z \alt 15\)) and the subsequent Epoch of Reionization, the evolution of \(T_{21}\) can be schematically described by five distinctive regimes: *(i)* For \(z_{\rm dec} < z < z_{\rm CMB}\), the non negligible amount of free electrons couples the gas to radiation, \(T_\gamma = T_{\rm gas} = T_s\), and there is no 21-cm signal because \(T_{21} = 0\). For \(z< z_{\rm dec}\), the gas cools down more rapidly than CMB radiation, but the gas temperature has a minimum value set by its primordial heating due to Thomson scattering and so \(T_{\rm gas} (z) = T_{\rm CMB} (z) (1 + z)/(1 + z_{\rm dec})\). *(ii)* For \(30 \alt z < z_{\rm dec}\), gas collisions are efficient enough to couple the spin and gas temperatures, i.e. \(T_s \simeq T_{\rm gas}\). This means \(T_{\rm 21} < 0\) and therefore an absorption signal is expected during the Dark Ages. *(iii)* At \(z \simeq 30\) the gas becomes so rarified that the collision rate becomes too low to enforce \(T_s \simeq T_{\rm gas}\) and the 21-cm signal is again suppressed \(T_{\rm 21} \simeq 0\) because \(T_s \simeq T_\gamma\). *(iv)* At Cosmic Dawn \(z \alt 30\), the first stars fired up, making the Lyman-\(\alpha\) coupling strong and thereby the spin temperature is expected to be coupled to the gas kinetic temperature, i.e., \(T_{\rm gas} \approx T_s\). Again \(T_{21} < 0\), implying an absorption signal. *(v)* During the epoch of reionization the gas gets reheated by astrophysical radiation yielding \(T_s \simeq T_{\rm gas} > T_\gamma\), so that \(T_{21}\) turns positive and one has an emission signal from the regions that are not fully ionized. Eventually all gas gets ionized until the fraction of neutral hydrogen vanishes and the signal switches off again. The Experiment to Detect the Global Epoch of Reionization Signature (EDGES) recorded the first measurement of the global 21-cm spectrum. The data are consistent with an absorption profile at \(z \sim 15-20\), with a minimum at \(z_E =17.2\) where \(T_{21} (z_E ) =-500^{+200}_{-500}~{\rm mK}\) at 99% C.L., including estimates of systematic uncertainties . Using ([\[T21\]](#T21){reference-type="ref" reference="T21"}) with \(x_{H_I} \simeq 1\) the EDGES observation implies \(T_{\rm gas} \simeq T_s (z_E) < 3.3~{\rm K}\), wherefore in tension with \(\Lambda\)CDM that predicts the minimum gas temperature to be \(T_{\rm gas}^{\Lambda{\rm CDM}} (z_E) \simeq 6~{\rm K}\). The EDGES 21-cm signal severely constrains new physics processes which are capable of heating up the intergalactic medium prior to the reionization time, e.g., via energy injection from annihilation or decaying dark matter. For sub-MeV dark matter, EDGES bounds on the partial decay width into two photons are stronger than CMB-based limits by one to two orders of magnitude. We now turn to investigate how the EDGES bound impacts the allowed parameter space of the DDG model. We begin by considering a tower of equally spaced dark gravitons, indexed by an integer \(l\), and with mass \(m_l = l \ m_{\rm KK}\). The partial decay width of KK graviton \(l\) to SM fields is found to be, \[\Gamma^l_{{\rm SM}} = \frac{\tilde{\lambda}^2 \ m^3_{\rm KK} \ l^3}{80 \pi M_{\rm Pl}^2} \,, \label{Ggamma}\] where \(\tilde \lambda\) takes into account all the available decay channels and is a function of time . The cosmic evolution of the dark sector is mostly driven by "dark-to-dark" decay processes, which regulate the decay of KK gravitons within the dark tower . In the absence of isometries in the dark dimension, which is the common expectation, the KK momentum of the dark tower is not conserved. This means that a dark graviton of KK quantum \(n\) can decay to two other ones, with quantum numbers \(n_1\) and \(n_2\). If the KK quantum violation can go up to \(\delta n\), the number of available channels is roughly \(l \, \delta n\). In addition, because the decay is almost at threshold, the phase space factor is roughly the velocity of decay products, \(v_{\rm r.m.s.} \sim \sqrt{m_{\rm KK} \ \delta n /m_l}\). Putting all this together we obtain the total decay width, \[\Gamma^l_{\rm tot} \sim \sum_{l'<l} \ \ \sum_{0<l''<l-l'} \Gamma^l_{l' l''} \sim \beta^2 \frac{m_l^3}{M_{\rm Pl}^2} \times \frac{m_l}{m_{\rm KK}} \ \delta n \times \sqrt{\frac{m_{\rm KK} \delta_n}{m_l} }\sim \beta^2 \ \delta n^{3/2} \frac{m_l^{7/2}}{M_{\rm Pl}^2 m_{\rm KK}^{1/2}} \,,\] where \(\beta\) parametrizes our ignorance of decays in the dark dimension . We further follow to estimate the time evolution of the dark matter mass and assume that for times larger than \(1/\Gamma^l_{\rm tot}\) dark matter which is heavier than the corresponding \(m_l\) has already decayed, yielding \[m_l \sim \left(\frac{M_{\rm Pl}^4 \ m_{\rm KK}}{\beta^4 \ \delta n^3}\right)^{1/7} t^{-2/7} \,, \label{mt}\] where \(t\) indicates the time elapsed since the big bang. All in all, the recombination history of the Universe would be modified by dark gravitons decaying into SM fields, as these visible fields inject energy into the (pre-recombination) photon-baryon plasma and (post-recombination) gas and background radiation. The energy injection would increase the ionization of the gas, the atomic excitation of the gas, and the plasma/gas heating. These three processes would therefore increase the residual ionization fraction (\(x_e\)) and baryon temperature after recombination. Consistency with CMB anisotropies requires \(\Gamma^l_{\gamma \gamma} < 5 \times 10^{-25}~{\rm s}^{-1}\) between the last scattering surface and reionization . In our calculations we set \(\lambda \sim 10^{-3}\) to accommodate neutrino masses with 3 generations of massless bulk fermions, see Appendix. Taking \(\tilde{\lambda} = 1\) to set out the decay into photons we can use ([\[Ggamma\]](#Ggamma){reference-type="ref" reference="Ggamma"}) to find that the CMB requirement is satisfied for \(l \alt 10^8\) at the time \(t_{\rm MR} \sim 6 \times 10^4~{\rm yr}\) of matter-radiation equality. This corresponds to \(m_l (t_{\rm MR}) \alt 1~{\rm MeV}\). In Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} we show \(\Gamma^l_{\gamma\gamma} (t)\) normalized to \(m_l (t_{\rm MR})= 1~{\rm MeV}\). As can be seen in Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}, DDG predictions from KK decay into SM fields saturate current upper limits placed by EDGES data on \(\Gamma^l_{\gamma \gamma}\) . This means that DDG predictions are within reach of next generation experiments. In particular, interferometric experiments like the upcoming Square Kilometre Array (SKA) will have the sensitivity to make high-resolution spectra of high-redshift radio sources to probe the 21-cm signal at Cosmic Dawn . The photon emission from KK graviton decay could be distinguished from simple X-ray heating by: *(i)* occurring before the epoch of galaxy formation and *(ii)* by depositing energy more uniformly than would be expected from galaxy clustering . The lunar FARSIDE array will measure the Dark Ages global 21-cm signal at redshifts \(35< z <200\), extending the sensitivity down two orders of magnitude below frequency bands accessible to ground-based radio astronomy . Thus, FARSIDE will provide an important test both of KK decays into SM fields and of the ideas discussed in this paper. Once the free parameters of the model have been adjusted to accommodate 21-cm data, we can use ([\[Ggamma\]](#Ggamma){reference-type="ref" reference="Ggamma"}) and ([\[mt\]](#mt){reference-type="ref" reference="mt"}) to make predictions that can be confronted with observations in the local universe. For example, for \(m_l(t_{\rm MR}) \sim 1~{\rm MeV}\), ([\[mt\]](#mt){reference-type="ref" reference="mt"}) leads to \(m_l(t_{\rm today}) \sim 30~{\rm keV}\). Now, we have seen that dark matter decay gives the daughter particles a velocity kick. Self-gravitating dark-matter halos that have a virial velocity smaller than this velocity kick may be disrupted by these particle decays. Combined cosmological zoom-in simulations of decaying dark matter with a model of the Milky Way satellite population rule out non-relativistic kick speeds \(\agt 10^{-4}\) for a dark matter lifetime \(\tau_{\rm DM} \alt 29~{\rm Gyr}\) at 95%CL . However, N-body simulations of isolated dark-matter halos seem to indicate that if \(\tau_{\rm DM} \agt 60~{\rm Gyr}\) and the kick speed \(\alt 10^{-2}\) then the halos are essentially unchanged . Setting \(\delta n \sim 1\) and taking \(\beta \sim 635\) to match our normalization (\(l \sim 10^8\) at \(t_{\rm MR}\); see ([\[mt\]](#mt){reference-type="ref" reference="mt"})), we obtain \(\tau_{_{m_l (\rm today)}} \sim 68~{\rm Gyr}\) and so the kick velocity of the decay products, \(v_{\rm r.m.s.} \sim 5 \times 10^{-3}\), is consistent with the mass concentration of self-gravitating dark matter halos.[^1] There are also restrictive bounds on \(\Gamma^l_{\gamma \gamma}\) from null results of indirect dark matter searches targeting dwarf spheroidals and cluster of galaxies. In particular, null results from searches by the NuSTAR Collaboration in the direction of the bullet cluster can be translated into an upper limit \(\Gamma^{\rm CDM}_{\gamma \gamma} < 10^{-28}~{\rm s}^{-1}\) for dark matter masses in the 10 to 50 keV range . As shown in Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}, DDG predictions also saturate this bound, suggesting that indirect detection dark matter experiments are extremely competitive to detect the KK decay products in the local universe.
# Higuchi bound and the shape of Inflation {#sec:3}
In de Sitter (dS) spacetime (of radius \(1/H\)) there exists an absolute minimum for a field of spin-\(2\) and mass \(m_2\) set by the Higuchi bound \[m_2^2 \geq 2 H^2 \,,\] where \(H = \Lambda^2/M_{\rm Pl} \sim 10^{-34}~{\rm eV}\) is the Hubble parameter . If the bound is violated the massive spin-2 field contains helicity modes with negative norm, which are in conflict with unitarity. Hence, if we define the "size" of the extra dimensions by the inverse mass of the lightest KK excitation of the graviton, the Higuchi bound forbids any compactification in which the extra dimensions are larger than and \({\cal O}(1)\) factor times \(1/H\) . Now, since \(H \ll m_{\rm KK} \sim \Lambda^{1/4}/\lambda\), the Higuchi bound is inordinately satisfied by the DDG construct today. However, the Higuchi bound forbids the presence of the DDG tower over a dS background within the mass range \(0 \leq m_l \leq 2 H_*^2\), which would imply that KK excitations could not be excited during inflation. Here, \(H_* = 10^{-6} \ \sqrt{r} \ M_{\rm Pl} \sim r^{1/2} 10^{14}\) is the Hubble factor during inflation, with \(r\) the tensor to scalar ratio of primordial gravitational waves .[^2] To confront this obstacle we adopt the working assumption that the Universe undergoes a period of inflation in which the radius of the dark dimension expanded exponentially fast, from the species length up to the micron-scale.[^3] The core idea behind the inflationary phase takes after the procedure adopted in string theories with large internal dimensions and TeV-scale gravity . The higher-dimensional metric is given by \[ds^2 =-dt^2 + a^2(t) \ (d\vec x^2 + r_0 \ dy^2) \,, \label{metric}\] where \(a(t) = e^{Ht}\) is the cosmic scale factor, \(\vec x\) denotes the 3 uncompactified dimensions, and \(r_0 \sim 1/M_{\rm UV}\) is the radius of the dark dimension \(y\) at the beginning of the inflationary phase. We want \(r_0\) to grow fast up to the micron scale. This requires 45 \(e\)-folds. ([\[metric\]](#metric){reference-type="ref" reference="metric"}) appears at first sight to be a simple generalization of the scale factor which describes the growth of \(n\) large extra dimensions. However, there are fundamental differences between the two scenarios. For example, if we ask the question what is the string scale so that an expansion of 50-60 \(e\)-folds of \(n\) dimensions of string size leads to a size compatible with the observed value of the 4-dimensional Planck mass, one finds that only \(n=1\) works with a string scale around intermediate energies. As a consequence, when the string scale is of order TeV, 4-dimensional inflation is not compatible with uniform inflation of the internal dimensions and one needs to introduce 2 different scale factors that have different time dependence at different epochs of the early universe . To find such solutions, one has to tune the radion and inflaton potentials to take ad hoc forms. For \(n=1\), this is not needed and uniform inflation driven by a 5-dimensional cosmological constant (corresponding to a flat region of the 5-dimensional potential) does the job. More precisely, \[M_{\rm Pl} = M_{4+n} \ R^{n/2} \,,\] where \(M_{4+n}\) is the \((4+n)\)-dimensional gravity scale and \(R\) the radius of \(n\) extra dimensions in string units. For D-branes, \(M_{4+n} =M_s/g_s\) where the \(g_s\) is the string coupling. After inflation of \(N\) \(e\)-folds, where the scale factor was expanded by \(a = e^N\), the radius becomes \(R=e^N\) with \(N=50\)-60 depending on the scale of inflation .[^4] This implies that \(a\) is about \(10^{21}\)-\(10^{26}\), and it follows that \[M_{4+n} = \frac{M_{\rm Pl}}{a^{n/2}} \,.\] For \(n\geq 2\), then \(a^{n/2}>10^{21}\), implying \(M_{4+n} <10~{\rm MeV}\), which is excluded by experiment. The only possibility is therefore \(n=1\) leading to \(a^{1/2}=10^{10}-10^{13}\) and \[10^6 \alt M_{4 + n}/{\rm GeV} \equiv M_{\rm UV}/{\rm GeV} \alt 10^9 \,. \label{range}\] To comply with the bound on deviations from Newton's gravitational inverse-square law we restrict the range in ([\[range\]](#range){reference-type="ref" reference="range"}) to \(M_{\rm UV} > 10^8~{\rm GeV}\). The inflaton mass \(m_\phi\) is model dependent, and we will take it as a free parameter. To produce matter on the brane it is sufficient to introduce a Yukawa-like coupling \(y\) of the inflaton to brane fermions, and so the decay width is \(\Gamma^\phi_{f\bar f} \sim y \ m_\phi\). The decay width into gravitons is Planck suppressed and we have to carry out only one sum up to \(l\) because \(\delta n \sim 1\). By dimensional analysis, \[\Gamma^\phi_{\rm grav} \sim \frac{m_\phi^3}{M_{\rm Pl}^2} \ \times \ \frac{m_\phi}{m_{\rm KK}} \ \times \ \delta n \,.\] The suppressed decay into gravitons ensures that the upper limit on the number of "equivalent" light neutrino species (\(\Delta N_{\rm eff} < 0.214\) at 95% CL ) present in the era before recombination is satisfied . The KK decomposition of the 5-dimensional metric gives rise to a tower of massive physical states of spin-2, while the radion is reduced only to the 4-dimensional zero mode. This is to be contrasted with the TeV-scale model advocated in , in which the inflaton is a field localized on the brane. However, in both cases the inflaton decays re-heat predominantly brane-states while not producing significant numbers of gravitons.
# Conclusions {#sec:4}
In the first part of this paper we proposed a method to test a particular realization of the incredible bulk of dynamical dark matter in wich the ensemble of KK modes originates in a compact (dark) dimension with characteristic length-scale in the micron range . Since the cosmic evolution of the dark sector is driven by dark-to-dark intra-ensemble decays, the model is particularly challenging to probe. We developed a two-step process test to unambiguously distinguish predictions of the dark dimension gravitons from those of the \(\Lambda\)CDM model. Measurements of redshifted 21-cm lines allow us to perform a tomographic study of the Dark Ages and Cosmic Dawn to search for the photon emission from KK graviton decay. We showed that the changes induced by \({\rm KK} \to \gamma \gamma\) on the 21-cm signal are within the reach of next generation experiments and can be used to make the model fully predictable. In particular, we can estimate the evolution of the dark graviton mass over the age of the Universe to pin down its value today. We demonstrated that the sensitivity of indirect dark matter searches targeting dwarf spheroidals and cluster of galaxies are extremely competitive to detect the KK decay products in the local universe. In the second part of this paper we examined the global structure of the inflationary phase within the context of the dark dimension. We have shown that the model parameters required for a successful uniform inflation driven by a 5-dimensional cosmological constant (corresponding to a flat region of the 5-dimensional potential) are natural. In the Appendix we generalized studies of large-extra-dimension models originally set up to interpret the smallness of neutrino masses by postulating that right-handed neutrinos, unlike SM fields, do propagate in the bulk .[^5] We showed that when interpreted in the context of the dark dimension, neutrino oscillation data constrain \(m_{\rm KK} > 2.5~{\rm eV}\) at the 99% CL.
# Appendix
We adopt the working assumption that neutrino masses originate in three 5-dimensional fermion fields \(\Psi_\alpha \equiv (\psi_{\alpha L},\psi_{\alpha R})\), which are singlets under the SM gauge symmetries and interact on our brane with the three active left-handed neutrinos \(\nu_{\alpha L}\) in a way that conserves lepton number, where the indices \(\alpha = e, \mu,\tau\) denote the generation . From a 4-dimensional perspective, each of the singlet fermion fields can be decomposed as an infinite tower of KK states, \(\psi^\kappa_{L(R)}\), with \(\kappa = 0,\pm 1, \cdots, \pm \infty\). The right-handed bulk states \(\psi_R^\kappa\) combine with the left-handed bulk components \(\psi_L^\kappa\) to form Dirac mass terms, which derive from the quantized internal momenta in the dark dimension. The bulk states also mix with the active left-handed neutrinos through Dirac-like mass terms. Redefining the bulk fields as \(\nu^{(0)}_{\alpha R} \equiv\psi^{(0)}_{\alpha R}\) and \(\nu^{(K)}_{\alpha L(R)}\equiv \Big(\psi^{(K)}_{\alpha L(R)}+\psi^{(-K)}_{\alpha L(R)}\Big)/\sqrt 2\), after electroweak symmetry breaking the mass terms of the Lagrangian take the form \[\begin{aligned}
\mathscr{L_{\text{mass}}} &= & \displaystyle \sum_{\alpha,\beta}m_{\alpha\beta}^{D}\left[\overline{\nu}_{\alpha L}^{\left(0\right)}\,\nu_{\beta R}^{\left(0\right)}+\sqrt{2}\, \sum_{K=1}^{\infty}\overline{\nu}_{\alpha L}^{\left(0\right)}\,\nu_{\beta R}^{\left(K\right)}\right] + \sum_{\alpha}\sum_{K=1}^{\infty}\displaystyle m_K \, \overline{\nu}_{\alpha L}^{\left(K\right)} \, \nu_{\alpha R}^{\left(K\right)} \, + {\rm h.c.} \nonumber \\ & = & \sum_{i=1}^3 \bar{\mathbb{N}}_{iR} \ \mathbb M_i \ \mathbb{N}_{iL}+ {\rm h.c.} \,, \label{calL} \end{aligned}\] where \(m_{\alpha \beta}^{D}\) is a Dirac mass matrix, \(m_K = K/R = K m_{\rm KK}\), \[\begin{aligned}
{\mathbb{N}_{i L(R)}}=\Big(\nu_i^{(0)},\nu_i^{(1)},\nu_i^{(2)},\cdots\Big)^T_{L(R)}, ~~~~~{\rm{and}}~~~~~ \mathbb M_i= \begin{pmatrix} m_i^D&0&0&0&\ldots\\ \sqrt{2}m_i^D&1/R&0&0&\ldots\\ \sqrt{2}m_i^D&0&2/R&0&\ldots\\ \vdots&\vdots&\vdots&\vdots&\ddots \end{pmatrix}, \end{aligned}\] and where \(m_i^D\) are the elements of the diagonalized Dirac mass matrix \(=\mathrm{diag}(m^D_1,m^D_2,m^D_3)\). Greek indices from the beginning of the alphabet run over the 3 active flavors (\(\alpha,\beta = e,\mu,\tau\)), Roman lower case indices over the 3 SM families \((i = 1,2,3)\), and capital Roman indices over the KK modes (\(K =1,2,3..., +\infty\)). Note that \(\psi_{\alpha L}^{(0)}\) decouples from the system. For the configuration at hand, \[m_i^D \approx \frac{y_i \ \langle H\rangle}{\sqrt{R M_{\rm UV}}} \,,\] where \(y_i\) are the Yukawa couplings and \(\langle H \rangle\) the Higgs vacuum expectation value. A quantity of pivotal interest is \(P(\nu_\alpha \to \nu_\beta)\), which defines the probability of finding a neutrino of flavor \(\beta\) in a beam that was born with flavor \(\alpha\) and has travelled a distance \(L\). Armed with the Lagrangian ([\[calL\]](#calL){reference-type="ref" reference="calL"}) it is straightforward to parametrize \(P(\nu_\alpha \to \nu_\beta)\) in terms of three mixing angles (\(\theta_{12}, \theta_{13}, \theta_{23}\)), a Dirac CP-violating phase (\(\delta_{13}\)), and the solar (\(\Delta m^2_{21}\)) and atmospheric (\(\Delta m^2_{31}\)) mass differences . Short- and long-baseline experiments constrain the standard 6 oscillation parameters and two extra parameters taken to be \(R\) and \(m_0 \equiv m_{1(3)}^D\) for the normal hierarchy \(m_3>m_2 > m_1 = m_0\) and (inverted hierarchy \(m^D_2>m^D_1> m^D_3 = m_0\)) .[^6] The most recent analysis of neutrino oscillations, combining data from MINOS/MINOS+, Daya Bay, and KATRIN, gives \(R < 0.4~\mu {\rm m}\), for the normal hierarchy, and \(R < 0.2~\mu{\rm m}\), for the inverted hierarchy, both upper limits at 99%CL . This implies that \(m_{\rm KK} \agt 2.5~{\rm eV}\), and therefore \(\lambda \alt 10^{-3}\).
[^1]: Since the model is a realization of dynamical dark matter , a dedicated simulation should be done to ensure full compatibility with the evolution of structure formation.
[^2]: The latest CMB observations of BICEP/Keck, combined with those of WMAP and the Planck mission, place a strong upper bound \(r \leq 0.036\) (at 95% CL).
[^3]: Other aspects of inflation and higher spin states in relation with the Higuchi bound were discussed in .
[^4]: If inflation occurs above about the TeV scale, then \(N \sim 40\) .
[^5]: Neutrino masses in relation to the swampland program were discussed in .
[^6]: A point worth noting at this juncture is that in the presence of bulk masses, sterile neutrinos propagating in the dark dimension can induce electron-neutrino appearance effects, and hence have the potential to address the MiniBooNE and LSND appearance results . | {'timestamp': '2023-01-05T02:13:04', 'yymm': '2212', 'arxiv_id': '2212.08527', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08527'} |
# Introduction
The potential of a static quark-anti-quark pair \(V_{0}(r)\) has always played an important role in Quantum Chromodynamics (QCD). It can be computed via Wilson loops and established an understanding of confinement and its interplay with asymptotic freedom, a central problem of particle physics, via the formation of a flux tube between quark-anti-quark static charges . Confinement manifests itself in the linear rise of \(V_{0}(r)\) at large \(r\); the corresponding slope is known as the string tension. The static potential can be used in the Born-Oppenheimer approximation to compute the spectrum of quarkonium . It is also an important observable in setting the scale in lattice QCD. In quenched calculations, the scale has been set using the string tension, but in full QCD the string breaks at the pair-production threshold, making a precise definition difficult. The static energy allows determination of the strong coupling, \(\alpha_s\), or, equivalently, \(\Lambda_{\scriptstyle \overline{\textrm{MS}}}\); see Refs. for recent reviews. Instead of the static energy, one can also use the force \(F(r) \equiv dV_{0}(r)/dr\), which is free of the self-energy linear divergence. The dimensionless product \(r^{2}F(r)\) can be used to set the scale at distances where statistical and systematic uncertainties are under good control, *e.g.*, \(r_{0}\) or \(r_{1}\), defined by \(r_{i}^{2} F(r_{i})= c_{i}\), with \(c_{0}=1.65\) , \(c_{1}=1\) In this paper, we investigate a method for computing the static quark-anti-quark potential in lattice QCD not based on Wilson loops, but where trial states are formed from components of eigenvectors of the covariant lattice Laplace operator . In this construction, the spatial Wilson lines in the Wilson loop are replaced by outer products of Laplacian eigenvectors. This idea was proposed in the context of adjoint string breaking and of Polyakov loops and the static potential at finite temperature. The main advantage is we can not only form straight lines (on-axis), but also off-axis paths very easily. These correspond to very complicated stair-like constructions of spatial link variables. It is important to compute the static potential for many off-axis separations whenever a fine resolution is required, *e.g.*, for a detailed investigation of string breaking or to determine the scale \(\Lambda_{\scriptstyle \overline{\textrm{MS}}}\) via matching the perturbative and the lattice QCD static potential . It is even mandatory to compute all possible on-and off-axis separations to determine the static potential in momentum space representation . The implementation of which uses only the eigenvector corresponding to the lowest eigenvalue can be significantly improved by summing over several eigenvectors, weighted by Gaussian profile functions of their corresponding eigenvalues. A similar method was successfully applied to hadronic correlation functions in where an optimal smearing profile was introduced in the distillation framework , which can be equivalently expressed as an optimal creation operator for a meson. In the case of the static potential we get an improvement for the static energies, which reach their plateau values at earlier temporal distances, to be quantified below. The improved implementation can also be adapted to measure multi-quark potentials, hybrid static potentials of exotic mesons, where the gluonic string excitations can be realized by applying covariant derivatives to the Laplacian eigenvectors, as well as static-light potentials with insertions of light quark propagators. Further, we present a simple way to illustrate the flux tube between a static quark and antiquark pair using a Laplacian eigenvector pair as a 'test charge' scanning the chromo-electromagnetic field. The article is organized as follows: First, we reintroduce the notation of *Laplace trial states* in . Next we reformulate the standard Wilson loop in terms of *Laplace trial state correlators* and discuss their improvement via Gaussian profile functions in , allowing us to formulate a generalized eigenvalue problem (GEVP) for the *Laplace trial state correlation basis matrix* of the static potential, resulting in optimal profile functions for ground and excited states. We test the new improved method on a dynamical fermion ensemble in , presenting results for effective energies, static potentials as well as excited states. In we look at the spatial distribution of the optimal Laplace trial states which probe the ground and excited static potentials of a quark-anti-quark pair. We draw our conclusions and give a short outlook in .
# Laplace trial states {#sec:lts}
Let \({\bar Q}^a(\vec x)\) denote a static color source with \(a=1,2,3\) at spatial position \(\vec x\). Wilson loops arise from correlations in time of trial states \(\bar{Q}(\vec x) U_s(\vec x, \vec y) Q(\vec y)\) for a static color anti-color source pair located at spatial positions \(\vec x\) and \(\vec y\) respectively[^1]. Note the same Wilson loops are obtained when the static color sources are replaced by static quarks since the heavy quark spins decouple in the static limit and the trace over spin yields a constant, see. The spatial Wilson line \(U_s(\vec x, \vec y)=\exp(i\int_{\vec x}^{\vec y}A_\mu dx^\mu)=\prod U_\mu\) is a path-ordered product of link variables from \(\vec x\) to \(\vec y\). We want to replace the spatial part of trial states in each time-slice with an alternative operator which respects the gauge transformation behavior of the spatial Wilson line, given by \[\begin{aligned}
U_s'(\vec x, \vec y) = G(\vec x)U_s(\vec x, \vec y)G^\dagger(\vec y), \end{aligned}\] to ensure gauge invariance of the trial state. The three-dimensional gauge-covariant lattice Laplace operator \(\Delta\), acting on a field \(\psi(\vec x)\) on a single time-slice of the four-dimensional lattice gives \[\begin{aligned}
\Delta \psi(\vec x)&=&\frac{1}{a^2}\sum_{k=1}^3[U_k^\dagger(\vec x-a\hat k)\psi(\vec x-a\hat k)\\ &&\qquad\qquad-2\psi(\vec x)+U_k(\vec x)\psi(\vec x+a\hat k)]\nonumber \end{aligned}\] and has the required transformation behavior \(\Delta'(\vec x,\vec y)=G(\vec x)\Delta(\vec x,\vec y)G^\dagger(\vec y)\). Eigenvalues \(\lambda\) of \(\Delta\) are gauge invariant, while eigenvectors \(v'(\vec x)=G(\vec x)v(\vec x)\) transform co-variantly . It follows, that we can write down a combination of eigenvector components for a given eigenvalue \(\lambda\), namely \(v(\vec x)v^\dagger(\vec y)\), which has the same behavior under gauge transformations as the spatial Wilson line \(U_s(\vec x,\vec y)\): \[v'(\vec x)v'^\dagger(\vec y)=G(\vec x)v(\vec x)v^\dagger(\vec y)G^\dagger(\vec y).\] At this point, inspired by the distillation operator \[\begin{aligned}
\square^{ab}(\vec z, \vec x) & = & \sum_{i=1}^{N_v}v_i^a(\vec z)v_i^{\dagger\,b}(\vec x) \,,\label{eq:box} \end{aligned}\] we introduce the more general operator \[\begin{aligned}
\tilde\square^{ab}(\vec z, \vec x) & = & \sum_{i=1}^{N_v}\rho_i v_i^a(\vec z)v_i^{\dagger\,b}(\vec x) \,,\label{eq:mbox} \end{aligned}\] by including a quark profile \(\rho_i\), which modulates contribution from different eigenmodes. Note \(\square\) is a projection matrix, \(\square^2=\square\) onto \(V\), the vector space spanned by \(\{v_i\}\), while \(\tilde\square\) is no longer idempotent, it still has an image given by the span of \(v_i\). Next, we define the auxiliary field on each time-slice \[\begin{aligned}
\chi^a(\vec z | \vec x) & = &\tilde\square^{ab}(\vec z, \vec x) Q^b(\vec x)\quad(\text{no sum over }\vec x)\nonumber \\ &=& \sum_{i=1}^{N_v} \rho_i v_i^a(\vec z)v_i^{\dagger\,b}(\vec x) Q^b(\vec x). \label{chifield} \end{aligned}\] \(\chi^a(\vec z | \vec x)\) can be interpreted as an effective smeared color-electromagnetic field over the whole time-slice induced by the static source at \(\vec x\). At first this seems contradictory to a 'static' color source, but it follows the notation of distillation. We stress the role of the 'smearing parameter' \(N_v\), the number of eigenvectors to be summed over in , behaves opposite to intuition. \(N_v=1\) corresponds to the maximal smearing and in the limit where all eigenvectors are included \(N_v\rightarrow3N_s^3\) with \(N_s^3\) the spatial lattice volume of a time slice, the smearing operator becomes the identity. This we have to keep in mind when constructing gauge invariant trial states for a color anti-color source pair located at spatial positions \(\vec x\) and \(\vec y\), respectively, via \[\begin{aligned}
\Phi(\vec x, \vec y) & = & \sum_{\vec z} \bar{\chi}(\vec z | \vec x) \chi(\vec z | \vec y)\nonumber \\ & = &\bar{Q}(\vec x)\sum_{i,j=1}^{N_v}\rho_i\rho_jv_i(\vec x)\nonumber \\ &&\quad\quad\;\;\underbrace{\sum_{\vec z}v_i^\dagger(\vec z)v_j(\vec z)}_{=\delta_{ij}}v_j^\dagger(\vec y) Q(\vec y)\nonumber \\ & = &\bar{Q}(\vec x)\sum_{i=1}^{N_v}\rho_i^2v_i(\vec x)v_i^\dagger(\vec y) Q(\vec y) \,, \label{eq:lts} \end{aligned}\] where we used the orthonormality of the Laplacian eigenvectors, which ensures that the standard distillation operator is idempotent, *i.e.*, \(\square^2 = \square\). We denote as a *Laplace trial state*, the positions \(\vec x\) and \(\vec y\) label the sector of the Hilbert space in which the static energies will be determined. Notice the sum over eigenvectors in must be truncated at finite \(N_v\) or a non-trivial profile \(\rho_i\) must be applied to avoid the collapse of the *Laplace trial state*, or the annihilation of the quark-anti-quark pair. For example \(\rho_i=\delta_{ik}\) corresponds to the choice of a single eigenvector \(v_k\). A simple truncation of the sum at some finite \(N_v=k\) could be formulated via \(\rho_i=\Theta(k-i)\) and we can of course introduce multiple profile functions to define an operator basis \(\Psi^{(k)}\) via different profiles \(\rho_i^{(k)}\) For example, \(\Psi^{(k)}\) with \(\rho_i^{(k)}=e^{-\lambda_i^2/4\sigma_k^2}\) corresponds to a sum over eigenvectors weighted with Gaussian profiles in eigenvalue space with different Gaussian widths \(\sigma_k\), which turned out to be very efficient for meson operators in. In the following section we will reformulate the usual Wilson loops in terms of *Laplace trial state correlators* and follow the same strategy as in by introducing a set of Gaussian profile functions into the the correlators and solving a generalized eigenvalue problem (GEVP) for the *Laplace trial state correlation matrix* to extract optimal trial state profiles \(\tilde\rho_i^{(n)}\) for ground and excited states of the static potential \(V_n(R)\), (\(n=0,1,2\ldots\)).
# The static quark-anti-quark potential from Laplace trial state correlators {#sec:imp}
The standard Wilson loop \(W(R,T)\) of size \((R=|\vec r|=|\vec y-\vec x|)\times(T=|t_1-t_0|)\) can be rewritten using Laplace trial state correlators by replacing the spatial Wilson lines \(U_s(\vec x,\vec y,t)\) with Laplacian eigenvector pairs \(v_i(\vec x,t)v_i^\dagger(\vec y,t)\), as depicted in . The temporal Wilson line \(U_t(\vec y,t_0,t_1)\), representing static time-like propagation for a color source at space point \(\vec y\) from time \(t_0\) to \(t_1\) is sandwiched between eigenvectors at the corresponding start-and end-times \(v_i^\dagger(\vec y,t_0)\) and \(v_j(\vec y,t_1)\). Distinct eigenvector indices appear at the source and sink times, so this can be interpreted as the static perambulator \[\begin{aligned}
\tau_{ij}(\vec y,t_0,t_1)=v_i^\dagger(\vec y,t_0)U_t(\vec y,t_0,t_1)v_j(\vec y,t_1),\label{eq:tau} \end{aligned}\] at \(\vec y\) of time extent \(T=|t_1-t_0|\). Its expectation value \(\langle \tau_{ij}(\vec y,t_0,t_1)\rangle\) vanishes of course. When combined with another static perambulator \(\tau_{ji}(\vec x,t_1,t_0)\) at \(\vec x\), it gives the Laplace trial state correlator \[\begin{aligned}
\hspace{-8mm}L(R,T)=\bigg\langle\sum_{i,j}^{N_v}\rho_i^2(t_0)\rho_j^2(t_1)\tau_{ij}(\vec y,t_0,t_1)\tau_{ji}(\vec x,t_1,t_0)\bigg\rangle\label{eq:Lkl} \end{aligned}\] for \(R=|\vec y-\vec x|\) (in our measurements we average over all \(\vec r\) of the same \(R\)). To test the method, the correlation function of is computed on a \(N_t\times N_s^3\) lattice ensemble with \(N_t=48, N_s=24\) and compared with standard Wilson loops. The Wilson loops are determined on 4646 gauge configurations while the Laplace trial-state correlators are computed on every fourth configuration only to give 1160 measurements. We extract the static potential via \(aV_0(R)=\lim_{T\rightarrow\infty}\log[L(R,T)/L(R,T+a)]\). First, we analyze the effect of increasing the number of eigenmodes \(N_v\) for trivial quark profiles. In we plot the effective energies for the static quark-anti-quark pair for \(R/a=2, 3\) and 4, and clearly see an increasing number \(N_v\) of Laplacian eigenvector pairs improves the overlap with the ground state drastically. Already \(N_v=8\) eigenvector pairs reach the plateau values faster than the original Wilson loops. The improvement seems to saturate at about \(N_v\approx100\), we do not see a difference between \(N_v=100\) and \(N_v=200\). The ground state overlaps can also be quantified by taking the \(t\)-average over the mass-plateau region of the fractional overlap \[\begin{aligned}
A_{\rm eff}=\frac{L(R,t)}{L(R,t_S)}\frac{\cosh \left( \left( \frac{aN_t}{2}-t_S \right) aV_0(R) \right)}{\cosh \left( \left( \frac{aN_t}{2}-t \right) aV_0(R) \right)},\label{eq:ovl} \end{aligned}\] using the same \(t_S=3a\) for all \(R/a\) and corresponding ground state energies \(aV_0(R)\) from a \(\cosh\)-fit, for more details see . These fractional overlaps are listed in and demonstrate that a large number \(N_v\) of eigenvector pairs gives better overlaps for small distances \(R/a\), but with decreasing importance for large distances, where already \(N_v<100\) shows better overlaps.
Next, instead of trivial quark profiles \(\rho_{i,j}\), we use Gaussian quark profile functions \(\rho_i^{(k)}=e^{-\lambda_i^2/4\sigma_k^2}\) and \(\rho_j^{(l)}=e^{-\lambda_j^2/4\sigma_l^2}\) for the Laplace trial states at \(t_0\) and \(t_1\) with corresponding eigenvalues \(\lambda_{i,j}\) and Gaussian widths \(\sigma_{k,l}\in[0.05,0.0894,0.1289,0.1683,0.2078,0.2472,0.2867]\). We define the \(7\times7\) Laplace trial state correlation matrix \(\mathcal L_{kl}(R,T)\) and solve a generalized eigenvalue problem (GEVP) to identify the optimal trial state profiles \(\tilde\rho_R^{(n)}(\lambda)\) for various energy levels \(V_n(R)\) (\(n=0,1,2,\ldots\)). First, we apply the strategy presented in and prune \(\mathcal L_{kl}\) using the three most significant singular vectors \(u_i\) from a singular value decomposition[^2] (SVD) at a specific \(t_G=4\) via \(\tilde{\mathcal L}_{mn}=u_{m}^\dagger\mathcal L_{kl}u_{n}\), which keeps a smaller set of distinct profiles which improves the stability of the GEVP. We perform the latter at the same \(t_G\), separately for all spatial distances \(R\): \[\begin{aligned}
\tilde{\mathcal L}(t)\nu^{(n)}(t,t_G)=\mu^{(n)}(t,t_G)\tilde{\mathcal L}(t_G)\nu^{(n)}(t,t_G). \end{aligned}\] From the eigenvalues or so-called principal correlators \(\lim_{t\rightarrow\infty}\mu^{(n)}(t,t_G)=e^{-E_n(t-t_G)}\) we get the effective energies for a fixed \(t_G\), by performing a \(\cosh\)-fit in practice, due to periodic boundary conditions. From the generalized eigenvectors \(\nu_k^{(n)}\) we can construct the optimal trial state profiles \(\tilde\rho^{(n)}_R\) for the energy states provided by the GEVP, which also depend on the quark separation \(R\), obviously. First, we use the singular vectors \(u_l\) to get the pruned (or most significant) profiles \(\bar\rho_R^{(k)}(\lambda_i)=\sum_lu_{k,l}e^{-\lambda_i^2/2\sigma_l^2}\). Then we form the linear combination of pruned profiles using the generalized eigenvectors \(\nu_k\) to give the optimal trial state profiles \[\begin{aligned}
\tilde\rho^{(n)}_R(\lambda_i)=\sum_k\nu_k^{(n)}\bar\rho_R^{(k)}=\sum_{k,l}\nu_k^{(n)}u_{k,l}e^{-\lambda_i^2/2\sigma_l^2}\,,\label{eq:propti} \end{aligned}\] depicted in for the ground and excited states at \(R=4a\). The optimal profiles suggest a number \(N_v<100\) of significant/important eigenvectors in the correlator, because each trial state comes with a profile and the combination falls off about twice as fast compared to . The fractional overlaps with the ground state in also favor the Laplace trial states from a GEVP with optimal profiles in the 6th column, which are even better than standard Wilson loop results from a GEVP with different HYP smearing levels (col. 7).
# Results from optimal Laplace trial states {#sec:res}
We performed all our measurements on \(48\times 24^3\) lattices with periodic boundary conditions except for anti-periodic boundary conditions for the fermions in the temporal direction. They were produced with the openQCD package using the plaquette gauge action and two dynamical non-perturbatively \(O(a)\) improved Wilson quarks with a mass equal to half of the physical charm quark mass. The bare gauge coupling is \(g_0^2=6/5.3\) and the hopping parameter is \(\kappa=0.13270\). The scale \(r_0/a=4.2866(24)\) and the flow scale is \(t_0/a^2 = 1.8477(3)\). The corresponding lattice spacing is \(a=0.0658(10)\)fm. All measurements were performed by our C+MPI based library that facilitates massively parallel QCD calculations. A total of \(N_v = 200\) eigenvectors of the 3D covariant Laplacian were calculated on each time-slice of the lattices as described in . A total of 20 3D APE smearing steps with \(\alpha_{APE} = 0.5\) were applied on each gauge field before the eigenvector calculation so as to smooth the link variables that enter the Laplacian operator. When forming the correlations of the Laplace trial states, we apply one HYP2 smearing step to the temporal links . Standard Wilson loops were measured using the wloop package , also applying one HYP2 step to all gauge links, and 4 levels (0 10 20 30 steps) of spatial HYP smearing to form a variational basis. Wilson loops were measured on 4646 gauge configurations, while Laplace trial states were measured on every fourth configuration only (1160 measurements). The error analysis in this work was done using the \(\Gamma\) method with a recent python implementation (pyerror) with automatic differentiation . We compare the effective energies using the improved Laplacian eigenvector approach with Gaussian profiles after solving the GEVP together with smeared Wilson loop results in . Results from Laplacian modes show better ground state overlaps and higher accuracy than those from Wilson loops with only a quarter of the statistics. In we present the static potentials \(V_n\) for the ground (\(n=0\)) and excited (\(n=1,2\)) states using the Laplace trial states with optimal quark profiles after solving the GEVP. The excited states are just included to show the potential of the method, we want to stress here, that we only have tha Laplace trial states in the operator basis, which just like Wilson loops may not have a good overlap with multi-particle states. For comparison we plot the radially excited string states \(V_0+(n+1)\pi/R\), as well as the lowest \(0^{++}\) iso-scalar meson (possible glueball) state \(V_0+m_{\rm eff}(\text{iso-scalar }0^{++})\) from and two times the static-charm quark mass. The latter was also evaluated using the new method, by combining our static perambulators \(\tau_{ij}(\vec x,t_0,t_1)\) with a projector \(P_+=(1+\gamma_0)/2\) and charm-quark perambulators \(\tau_{ji}^{\alpha\beta}(t_1,t_0)=v_j^\dagger(t_1)[D^{-1}]_{t_1t_0}^{\alpha\beta}v_i(t_0)\) from , where the quark propagator \(D^{-1}\) includes the dependence on the mass of the quark. The computational effort of this new method is less than the standard Wilson loop calculation, especially for off-axis separations. In fact, for our test ensemble on a \(24^3\times48\) lattice the computation of on-axis Wilson loops using 4 spatial smearing levels (0, 10, 20, 30 HYP steps) is equally expensive as the calculation of 100 Laplacian eigenvectors and Laplace trial states with 4 Gaussian profiles including off-axis distances. The computational advantage of new method can be explained by the fact that the static perambulators can be computed first at each position, resulting in complex numbers, which then can easily be multiplied for arbitrary on-and off-axis separations without the need to compute spatial Wilson lines. In we present the optimal static potential \(V_0(R)\) for all on-and off-axis separations \(R/a\) from \(N_v=100\) Laplacian eigenvectors compared to on-axis Wilson loop results, which agree well within errors. We also include a measurement of un-smeared Laplace trial state correlators for \(R/a\leq3\) (no HYP smearing), showing the Coulomb behavior of the potential at small \(R\). The green points in the plot are shifted vertically such that the un-smeared potential matches the potential with HYP2 smeared temporal links at \(R/a=2\), which corresponds to removing the free energy difference. Further, we want to note that contrary to Wilson loops, Laplace trial states have an exact symmetry of the potential around half the lattice extension (in a specific direction \(\vec r\)), where in fact the force between \(Q\bar Q\) must vanish due to the periodic boundary conditions, *i.e.*, the static potential should be flat.
# The spatial distribution of optimal Laplace trial states {#sec:flux}
If we do not evaluate the spatial sum in the third line of the Laplace trial state in , we are left with an eigenvector pair \(v^\dagger(\vec z)v(\vec z)\) which acts as a 'test-charge' in the original Laplace trial state \[\begin{aligned}
\hspace{-8mm}\psi^{(n)}(\vec z,R)=\hfill\\&&\hspace{-18mm}\bigg\langle\big|\big|\sum_{ij}^{N_v}\tilde\rho_R^{(n)}(\lambda_i,\lambda_j)v_i(\vec x)v_i^\dagger(\vec z)v_j(\vec z)v_j^\dagger(\vec x+R)\big|\big|_2\bigg\rangle\,,\nonumber \end{aligned}\] which allows the scanning of individual contributions of the quark-anti-quark operator in a 3D time-slice via the free coordinate \(\vec z\). We average over the whole lattice (\(\vec x,t\)), which already gives a very smooth signal on a single configuration. Note that we include the optimal trial state profiles \[\begin{aligned}
\tilde\rho_{R}^{(n)}(\lambda_i,\lambda_j)=\sum_{k,l}\nu_k^{(n)}u_{k,l}e^{-\lambda_i^2/4\sigma_l^2}e^{-\lambda_j^2/4\sigma_l^2}\,, \end{aligned}\] which in this case still depend on the two eigenvalues \(\lambda_i\) and \(\lambda_j\), since we did not perform the sum over \(\vec z\) in and therefore did not get a \(\delta_{ij}\). The singular vectors \(u_k\) and generalized eigenvectors \(\nu^{(n)}\) come from the SVD and GEVP in the static potential calculations for specific quark separation distances \(R\) and allow us to look at the flux tube profiles for various energy states of \(V_n(R)\). In figures [\[fig:r10n0\]](#fig:r10n0){reference-type="ref" reference="fig:r10n0"} and [\[fig:r10n1\]](#fig:r10n1){reference-type="ref" reference="fig:r10n1"} we present the spatial distributions of the optimal Laplace trial states to measure the ground resp. first excited state potentials of a static quark-anti-quark pair at spatial distance \(R=10a\). The first excitation shows additional nodes in the spatial distribution along and perpendicular to the quark separation axis. The physical interpretation of these distributions in terms of the chromo-electromagnetic flux tube is not clear yet, the optimal profiles certainly contain some information of the ground and excited states of the static potential, the 'test-charge' \(v(\vec z)v^\dagger(\vec z)\) however does not measure a specific color field component.
# Conclusions & Outlook {#sec:co}
Alternative creation operators for static-quark-anti-quark states based on Laplacian eigenmodes are investigated. The use of a large number of eigenvectors weighted with Gaussian profiles is found to improve performance. An operator basis can be defined via different Gaussian profiles which can be analyzed with the GEVP formalism to extract optimal profiles and Laplace trial states. Temporal correlations of the new operators are used to compute static quark-anti-quark ground and excited state potentials. We observe earlier plateaus in the effective masses compared to standard Wilson loops. One significant advantage of the approach is its efficiency for computing the static potential not only for on-axis, but also for many off-axis quark-anti-quark separations. Indeed the new method requires far less computing time in particular for the latter case, since the eigenvector components of the covariant lattice Laplace operator have to be computed only once and can then be used for arbitrary on-axis and off-axis separations without the need to compute stair-like gauge-link connections. Finally, we visualize the spatial distribution of the optimal Laplace trial states for ground and excited state creation operators of the quark-anti-quark pair. We are currently working on an adaptation of the method to compute hybrid static potentials of exotic mesons, where gluonic string excitations requiring gluonic handles in the standard Wilson loop approach can be realized with covariant derivatives acting on the Laplacian eigenvectors, and to static-light mesons, cf.. First results were presented at the ConfinementXV , Lattice 2022 and ExcitedQCD conferences. | {'timestamp': '2022-12-19T02:12:16', 'yymm': '2212', 'arxiv_id': '2212.08485', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08485'} |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction
Over the past decades, the hierarchical structure formation scenario based on the standard \(\Lambda\)CDM (cold dark matter) model has been established by the observations such as the cosmic microwave background (CMB) and the galaxies which trace the large-scale structure of the universe. Especially, the matter power spectrum measured by the cosmological observations is consistent with the \(\Lambda\)CDM model at large scales. On the other hand, the current Lyman alpha forest observations investigate the fluctuations of the hydrogen atoms in the intergalactic medium (IGM) and constrain the matter power spectrum at small scales up to \(k\sim 2 h\mathrm{Mpc}^{-1}\). The scales smaller than \(\sim 1\) Mpc which the Lyman alpha forest cannot probe have not been well explored observationally, and the 21cm absorption line system called the 21cm forest, in analogy to the Lyman forest, is a promising tool to explore those smaller scales. The 21cm absorption line is caused by neutral hydrogen atoms due to the hyperfine structure. They appear as absorption lines in the spectra of high redshift radio-loud sources due to neutral hydrogen atoms in the intervening cold neutral IGM and collapsed objects. In particular, the smallest collapsed objects called "minihalos" can be the promising sources to lead to the 21cm forest for the scales corresponding to \(k\gtrsim 10\) \[\(\mathrm{Mpc^{-1}}\)\]. The minihalos can form when the virial temperature is below the threshold where the atomic cooling becomes ineffective (\(T_{\mathrm{vir}} \lesssim 10^4 \mathrm{K}\)). Thus, they cannot cool effectively and cannot collapse to form protogalaxies. In such a condition, the masses of the minihalos are of order \(M\lesssim 10^8 M_{\odot}\) corresponding to the scales \(k\gtrsim 10\) \[\(\mathrm{Mpc^{-1}}\)\]. Those small scales are of great interest for the further exploration of the \(\Lambda\)CDM cosmology and its extension. For instance, there are some challenges at the small scales for \(\Lambda\)CDM cosmology such as the "missing satellite problem", "core-cusp problem" and "too big to fail problem". The small-scale structures probed by the 21cm forest can also give a clue to the nature of dark matter such as the warm dark matter, ultra-light axion-like particles and primordial blackholes. We aim to demonstrate in this paper that the effects of the relative velocity between the dark matter and baryon can also be imprinted in the 21cm forest signals. Such a relative velocity arises due to the different evolution between the dark matter and baryons before the photon-baryon decoupling epoch. The photons and baryons are tightly coupled through Thomson scattering until the recombination epoch when the number density of free electrons becomes low, while the CDM fluctuations keep growing under DM's own gravity. Such a different evolution results in supersonic relative velocity between baryon and DM at the recombination epoch with the root-mean-square (RMS) velocity \(\sigma_{\mathrm{bc}} \sim 30\mathrm{km/s}\) which is coherent over the scales of order a few comoving Mpc, while the sound speed of the baryon fluid drops dramatically to about 6 km/s (and the Jeans length also drops dramatically to \(\sim 30\) kpc). Such a supersonic stream of baryons through a dark matter without falling into the potential well of DM consequently can result in the structure formation suppression. The relevant scales sensitive to the relative velocity are expected to be around the baryon Jeans scales at which the fluctuations are sensitive to both baryonic pressure and gravitational infall. The larger scale growth is more affected by the gravitational attraction to the potential well and the smaller scale growth is more affected by the baryonic pressure which exists even without the supersonic relative velocity. The large suppression of the matter power spectrum due to the streaming velocity hence is expected to show up around the Jeans scale (\(k\) of order \({\cal O}(10^2)\mathrm{Mpc}^{-1}\) at the redshift \(z\gtrsim 10\)), and it is intriguing that the 21cm forest observations can provide the unique and promising probes on such small scales. This paper focuses on the effects of the halo abundance suppression and leaves more detailed numerical calculations including the change of the gas profiles in the presence of the streaming velocity to the future work. In this paper, we estimate how the relative velocity impacts the 21cm forest signals. Section [2](#Sec2){reference-type="ref" reference="Sec2"} gives a brief review of the effects of the DM-baryon relative velocity on the matter power spectrum and the halo mass function. We then outline the formalism to estimate the 21cm absorption line abundance in Section [3](#Sec3){reference-type="ref" reference="Sec3"}. Section [4](#Sec4){reference-type="ref" reference="Sec4"} gives our results, followed by the discussions/conclusions in Section [5](#Sec5){reference-type="ref" reference="Sec5"}.
# The impact of relative velocity on structure formation {#Sec2}
## Relative velocity
After the recombination, both baryon and CDM perturbations grow as the pressureless fluids above the Jeans scale with a characteristic relative velocity \(v_{bc}\sim 30 (1+z)/1000\) \[km/s\] and the baryon sound speed \(c_s =\sqrt{\gamma k_{\mathrm{B}}T_b/\mu m_{\mathrm{H}}}\). Here, \(\gamma=5/3\) for an ideal monoatomic gas, and \(\mu=1.22\) is the mean molecular weight including a helium mass fraction of 0.24. \(m_{\mathrm{H}}\) is the mass of the hydrogen. The density fluctuation evolutions are governed by the following equations +a\^-1 **v**\_c \_c&=&-a\^-1(1+\_c) \_c [\[eq:1\]]{#eq:1 label="eq:1"}\
+a\^-1 **v**\_b \_b&=&-a\^-1(1+\_b) \_b\
+a\^-1 ( **v**\_c ) **v**\_c&=&\--H **v**\_c\
+a\^-1 ( **v**\_b ) **v**\_b&=&\--H **v**\_b-a\^-1c_s\^2 \_b\
a\^-2\^2 &=& 4G \|\_m \_m [\[eq:5\]]{#eq:5 label="eq:5"} which represent the continuity equations for the CDM and baryon density fluctuation (\(\delta_c,\delta_b\)), the Navier-Stokes equations for the CDM and baryon velocity (\({\bf v}_c,{\bf v}_b\)) and the Poisson equation for the gravitational potential \(\Phi\). The different evolutions of baryon and CDM fluctuations before the decoupling can source the relative velocity between them, \(\vbc\), which can be supersonic just after the decoupling epoch, and the effects on the cosmological observables such as the baryon acoustic oscillations and 21cm fluctuations have been discussed. We in this paper discuss how the 21cm forest signals can be modified in the presence of such a streaming velocity, with a particular emphasis on the suppression of the minihalo formation, compared to those estimated without taking account of the non-negligible \(\vbc\). Note this is the non-linear effect arising from the quadratic terms in the perturbation evolution equations and does now show up in the conventional linear theory. The resultant advection of baryons out of the dark matter potential well caused by the large-scale velocity flow (associated with the baryon-photon fluid acoustic oscillations up to the sound horizon scale \(\sim 150\) comoving Mpc) leads to the modulated small-scale perturbation suppression.
## Matter power spectrum and Mass function
The presence of relative velocity changes the structure formation of the Universe. To understand the impact of relative velocity, we first see the impact of relative velocity on the matter power spectrum by solving Eqs ([\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"})-([\[eq:5\]](#eq:5){reference-type="ref" reference="eq:5"}). In Fig.[\[fig:ps\]](#fig:ps){reference-type="ref" reference="fig:ps"}, we show the matter power spectrum including the relative velocity at \(z=40,10\). At \(z=40\), the matter power spectrum is suppressed by \(\sim10\%\) at the scales of 50-500 \(\mathrm{Mpc^{-1}}\), and maximally suppressed at the Jeans scale \(k_{\mathrm{J}}\sim 200 \mathrm{Mpc}^{-1}\). The Jeans scale is the scale of balance between gravitational attraction and baryonic perturbation, thus additional pressure due to the relative velocity impacts the growth of structure formation at these scales. On the other hand, perturbations on much larger scales are not affected by pressure and are gravitationally collapsed, whereas the pressure at smaller scales prevents perturbations from collapsing. This qualitative argument also applies to the matter power spectrum suppression at \(z=10\). We next see the impact of relative velocity on the mass function. For the halo mass function, we adopt the Press-Schechter mass function. We mention that the Sheth-Tormen mass function is more precise at least for a low redshift. The halo mass function at a high redshift of our interest is however less certain, and the difference between Press-Schechter formalism and Sheth-Tormen formalism is small and does not affect our discussions. Hence the Press-Schechter mass function suffices for our purpose of demonstrating the potential effects of the relative velocity on the 21cm forest signals. Fig.[\[fig:ps\]](#fig:ps){reference-type="ref" reference="fig:ps"} shows the anticipated suppression in matter power spectrum in the presence of the relative velocity at the scales of \(k\)=20-3000 \(\mathrm{Mpc}^{-1}\) at \(z\)=10. The corresponding halo mass scale enclosed in the sphere \(r= 2\pi /k\) is around \(4.7 \times 10^3 M_{\odot}\lesssim M\lesssim 5.3 \times 10^9 M_{\odot}\), and the suppression of the abundance of those minihalos is illustrated in Fig. [\[fig:MF\]](#fig:MF){reference-type="ref" reference="fig:MF"}.
# 21cm forest {#Sec3}
The 21cm forest is a system of 21cm absorption lines that appears in the continuum spectrum of a radio background source. The 21cm absorption lines are generated by neutral hydrogen atoms in the diffuse IGM or minihalos. The column density of minihalos is larger than that of diffuse IGM and hence minihalos can produce a larger optical depth. In this work, we focus on the 21cm forest by the neutral hydrogen gas in the minihalos based on.
## Halo and gas profiles
We first specify the dark matter halo and gas profiles analytically for the estimation of the 21cm absorption lines. We assume the Navarro-Frenk-White (NFW) profile for the dark matter density distribution characterized by the concentration parameter \(y=r_{\rm vir}/r_{s}\),where \(r_{s}\) is the scale radius and \(r_{\rm vir}\) is the virial radius given by \[\begin{aligned} r_{\mathrm{vir}}=& 0.784\left(\frac{M}{10^8 h^{-1} M_{\odot}}\right)^{1 / 3}\left[\frac{\Omega_m}{\Omega_m^z} \frac{\Delta_c}{18 \pi^2}\right]^{-1 / 3} \\ & \times\left(\frac{1+z}{10}\right)^{-1} h^{-1}[\mathrm{kpc}]. \label{eq:virial} \end{aligned}\] \(\Delta_c=18 \pi^2+82 d-39 d^2\) is the over-density of halos collapsing at redshift \(z\) with \(d=\Omega_m^z-1\) and \(\Omega_m^z=\Omega_m(1+z)^3 /\left(\Omega_m(1+z)^3+\Omega_{\Lambda}\right)\). The concentration parameter \(y\) depends on the halo mass and redshift. conducted the high-resolution N-body simulations and showed that the redshift dependence of the concentration parameter is inversely proportional to \((1+z)\). We assume that the concentration parameter \(y\) is given by the fitting formula obtained by. Given the dark matter density profile, we can obtain the gas density profile analytically with the assumption that gas is in isothermal and hydrostatic equilibrium states. In that case, the gas density profile is given by \[\ln \rho_g(r)=\ln \rho_{g 0}-\frac{\mu m_p}{2 k_{\mathrm{B}} T_{\mathrm{vir}}}\left[v_{\mathrm{esc}}^2(0)-v_{\mathrm{esc}}^2(r)\right], \label{eq:gas_profile}\] where \(\mu=1.22\) is the mean molecular weight of the gas and \(m_p\) is the proton mass. \(T_{\mathrm{vir}}\) is the virial temperature given by \[\begin{aligned} T_{\mathrm{vir}}=& 1.98 \times 10^4\left(\frac{\mu}{0.6}\right)\left(\frac{M}{10^8 h^{-1} M_{\odot}}\right)^{2 / 3}\left[\frac{\Omega_m}{\Omega_m^z} \frac{\Delta_c}{18 \pi^2}\right]^{1 / 3} \\ & \times\left(\frac{1+z}{10}\right)[\mathrm{K}]. \end{aligned}\] \(\rho_{g0}\) is the central gas density given by \[\rho_{g 0}(z)=\frac{\left(\Delta_c / 3\right) y^3 e^A}{\int_0^y(1+t)^{A / t} t^2 d t}\left(\frac{\Omega_b}{\Omega_m}\right) \bar{\rho}_m(z),\] where \(A=3 y / F(y)\) and \(F(y)=\ln (1+y)-y /(1+y)\).\(\bar{\rho}_m(z)\) is the mean total matter density at \(z\). \(v_{\mathrm{esc}}\) is the escape velocity given by \[v_{\mathrm{esc}}^2(r)=2 \int_r^{\infty} \frac{G M\left(r^{\prime}\right)}{r^{\prime 2}} d r^{\prime}=2 V_c^2 \frac{F(y x)+y x /(1+y x)}{x F(y)},\] where \(x\equiv r/r_{\mathrm{vir}}\). \(V_{\mathrm{c}}\) is the circular velocity given by \[\begin{aligned} V_c^2=\frac{G M}{r_{\mathrm{vir}}}=& 23.4\left(\frac{M}{10^8 h^{-1} M_{\odot}}\right)^{1 / 3}\left[\frac{\Omega_m}{\Omega_m^z} \frac{\Delta_c}{18 \pi^2}\right]^{1 / 6} \\ & \times\left(\frac{1+z}{10}\right)^{1 / 2}[\mathrm{~km} / \mathrm{s}]. \end{aligned}\] We leave more detailed numerical calculations including the change of the temperature/gas profiles and the star formation rate in the presence of streaming velocity to the future work.
## Spin temperature
The spin temperature is a key quantity in estimating the 21cm line spectrum. The spin temperature \(T_{\mathrm{S}}\) is defined by the ratio between the number density \(n_i\) of the neutral hydrogen atom in the two hyper-fine levels (singlet \(n_0\) and triplet \(n_1\)) \[\frac{n_1}{n_0}=3\exp \left(\frac{-h \nu_{21}}{k_{\mathrm{B}} T_{\mathrm{S}}}\right),\] where \(h\) is Planck constant, \(k_{\mathrm{B}}\) is Boltzmann constant and \(\nu_{21}=1.4\)GHz. The spin temperature is determined by the interaction of neutral hydrogen with CMB and Lyman-\(\alpha\) photons, and by collisions between hydrogen atoms by the following equation, \[T_{\mathrm{S}}^{-1}=\frac{T_\gamma^{-1}+x_c T_{\mathrm{K}}^{-1}+x_\alpha T_{\mathrm{C}}^{-1}}{1+x_c+x_\alpha}.\] Here, \(T_{\gamma}=2.73 \mathrm{K}\) is CMB temperature at redshift \(z\), \(T_{\mathrm{K}}\) is the gas kinetic temperature, and \(T_{\mathrm{C}}\) is color temperature of Lyman-\(\alpha\) photon. \(x_{\alpha}\) and \(x_{c}\) are coupling coefficients for the interaction with Lyman-\(\alpha\) photon collision with neutral hydrogen atom, respectively. To understand how the relative velocity impacts the 21cm forest, we ignore any UV radiation field and radiative feedback and thus we set \(x_{\alpha}=0\). We also set \(T_{K}=T_{\mathrm{vir}}\). This is a good approximation for a minihalo because the minihalos are considered the collapsed objects with \(T_{\mathrm{vir}}=T_{\mathrm{K}} < 10^4 \mathrm{K}\) and the gas cooling is inefficient within the minihalos. Although the main contributions to collisional coupling are H-H and H-\(e^{-}\), we only consider H-H collision for computation of \(x_{c}\) because the fraction of free electron is small in a minihalo and we can ignore its effect. The spin temperature approaches the virial temperature in the inner regions of a minihalo(and thus a larger minihalo has a larger spin temperature) and approaches the CMB temperature in the outer regions of a minihalo because the collisional coupling becomes ineffective due to the small gas density.
## Optical depth
The optical depth to 21cm absorption lines by neutral hydrogen gas in a minihalo with mass \(M\) (at a frequency \(\nu\) and at impact parameter \(\alpha\)) is given by \[\begin{aligned} \tau(\nu, M, \alpha)=& \frac{3 h_{\mathrm{p}} c^3 A_{10}}{32 \pi k_{\mathrm{B}} \nu_{21}^2} \int_{-R_{\max }(\alpha)}^{R_{\max }(\alpha)} d R \frac{n_{\mathrm{HI}}(r)}{T_{\mathrm{S}}(r) \sqrt{\pi} b} \\ & \times \exp \left(-\frac{v^2(\nu)}{b^2}\right), \end{aligned}\] where \(r^2=\alpha^2+R^2\) and \(R_{\mathrm{max}}\) is the maximum radius of halo at \(\alpha\). The exponential factor represents the Doppler broadening with \(v(\nu)=c(\nu-\nu_{21}))/\nu_{21}\) (\(\nu_{21}\)=1.4GHz) and \(b=\sqrt{2 k_B T_{\mathrm{vir}} / m_p}\) is the velocity dispersion. \(n_{\mathrm{HI}}\) is the number density of neutral hydrogen atoms in the minihalo. A smaller impact parameter results in a larger optical depth due to a larger column density despite a larger spin temperature. As we mentioned above, the spin temperature approaches the virial temperature in the inner regions of a minihalo, thus a smaller minihalo has a smaller spin temperature and this results in larger optical depth at a fixed impact parameter.
## Abundance of the 21cm absorption lines
In order to evaluate the number of 21cm absorption lines from minihalos in the observed spectrum per redshift, we calculate the abundance of the 21cm line absorption as \[\frac{d N(>\tau)}{d z}=\frac{d r}{d z} \int_{M_{\min }}^{M_{\max }} d M \frac{d N}{d M} \pi r_\tau^2(M, \tau), \label{eq:abundance}\] where \(dN/dM\) is the halo mass function which represents the comoving number density of collapsed dark matter halos with a mass between \(M\) and \(M+dM\). \(dr/dz\) is the comoving line element. \(r_\tau(M, \tau)\) is the maximum impact parameter in the comoving unit that gives the optical depth larger than \(\tau\). In our study, we focus on 21cm forest generated by neutral hydrogen atoms in minihalos. The upper bound of the minihalo mass range of our interest corresponds to the minimal halo mass for which the stars can be formed and the lower bound is the Jeans scale (or a 'filtering scale' which is essentially the time-averaged Jeans scale by taking into account of the time dependence of the Jeans scale ). Our maximum minihalo mass scale can be bigger in the presence of the relative velocity because the relative velocity can cause the baryons to overshoot the dark matter halo retarding their collapse. The numerical simulations show that the relative velocity would not change the minimum mass required for the star formation so significantly (typically by less than a factor 2) . We hence conservatively use the typical upper bound minihalo mass corresponding to the virial temperature of \(T_{\rm vir} > 10^4\) K which is often adopted in the standard discussions to estimate the minihalo contribution to the 21cm forest without the streaming velocity \[M_{\max }(z)=3.95 \times 10^{7}\left(\frac{\Omega_{m} h^{2}}{0.15}\right)^{-1 / 2}\left(\frac{1+z}{10}\right)^{-3 / 2} M_{\odot}. \label{eq:maximum}\] Our lower bound minihalo mass scale can also increase due to the anisotropic pressure caused by the streaming velocity. We set the minimum mass by the baryonic Jeans mass that takes into account the effect of the relative velocity. The Jeans scale \(k_{J}\propto 1/c_s(z)\) which is inversely proportional to the gas sound speed can be influenced by the stream velocity. We accordingly consider the effective sound speed taking account of the relative velocity \(v_{\mathrm{eff}}=\sqrt{c_s^2(z)+v_{bc}^2(z)}\), so that the effective Jeans scale \(k_{\mathrm{J,eff}}\propto 1/v_{\mathrm{eff}}\). The corresponding Jeans mass can be estimated as \[M_J=\frac{4 \pi \bar{\rho}}{3}\left(\frac{5 \pi k_B T_{\mathrm{IGM}}}{3 G \bar{\rho} m_p \mu}\right)^{3 / 2}\left( \frac{\sqrt{c_\mathrm{s}^2+v_{\mathrm{bc}}^2}}{c_\mathrm{s}}\right)^3 \label{eq:minimum}\] where \(\bar{\rho}\) is the total mass density including dark matter and \(T_{\mathrm{IGM}}\) is the temperature of the IGM. We set \(T_{\mathrm{IGM}}=2\) K at \(z=10\) which satisfies the current lower bound of the IGM temperature obtained by HERA experiment. We use Eqs ([\[eq:maximum\]](#eq:maximum){reference-type="ref" reference="eq:maximum"}) and ([\[eq:minimum\]](#eq:minimum){reference-type="ref" reference="eq:minimum"}) as the upper and lower minihalo mass scales \(M_{\mathrm{max}}\) and \(M_{\mathrm{min}}\), respectively, in estimating the 21cm signals.
# Results {#Sec4}
In Fig.[\[fig:21cm_forest_z10\]](#fig:21cm_forest_z10){reference-type="ref" reference="fig:21cm_forest_z10"}, we show the abundance of the 21cm absorption lines at \(z=10\) as a function of the optical depth per redshift interval along a line of sight and per optical depth. We show the abundance of the 21cm absorption lines with and without the relative velocity for illustration. The abundance of the 21cm absorption lines in the case that includes the relative velocity is smaller than in the case that does not include the relative velocity by around one order of magnitude. In the absence of the relative velocity, the abundance of the 21cm line absorption lines per line of sight direction is around \(O(10)\) at \(\tau \lesssim 0.1\), whereas, in the presence of relative velocity, it reduces to around \(O(1)\). The impacts of relative velocity on the 21cm forest appear via the suppression of the mass function (or matter power spectrum) and the enhancement of Jeans mass. As shown by Fig.[\[fig:MF\]](#fig:MF){reference-type="ref" reference="fig:MF"}, the presence of relative velocity suppresses the halo mass function and, as shown in eq([\[eq:abundance\]](#eq:abundance){reference-type="ref" reference="eq:abundance"}), the suppression of the mass function reduces the number of 21cm absorption lines. As shown in eq([\[eq:minimum\]](#eq:minimum){reference-type="ref" reference="eq:minimum"}), the Jeans mass is roughly 20 times larger due to the relative velocity. The Jeans mass affects the minimum minihalo mass which contributes to the abundance of 21cm absorption lines, and such a change substantially can reduce the 21cm signals because the abundance of the 21cm absorption lines is more significantly affected by the smaller minihalos with a smaller spin temperature and a bigger optical depth rather than by the bigger minihalos. In Fig.[\[fig:21cm_forest_zevolution\]](#fig:21cm_forest_zevolution){reference-type="ref" reference="fig:21cm_forest_zevolution"}, we show the abundance of 21cm absorption lines at different redshifts. At a higher redshift, the relative suppression of the 21cm absorption line abundance is more prominent. One reason is the halo mass function. At a higher redshift, the relative velocity \(\propto (1+z)\) is bigger and a smaller halo has a less deep gravitational potential well. Another reason is the Jeans mass. In the absence of X-ray heating, the IGM gas temperature scales as \(\propto (1+z)^2\). Thus, the IGM gas temperature is higher and the Jeans mass is larger at a higher redshift (eq([\[eq:minimum\]](#eq:minimum){reference-type="ref" reference="eq:minimum"})). At a lower redshift, the number of halo increases, and the gas temperature (Jeans mass) decreases, thus the number of 21cm absorption lines increases. At \(z=11\) and \(z=15\), the abundance of the 21cm absorption lines is \(O(1)-O(10)\) in the absence of the relative velocity. Taking the relative velocity into account, we can see from the figures that the abundance of the 21cm absorption lines decreases while keeping \(O(1)\) at a lower \(\tau\) (up to \(\sim 0.03)\) at \(z=11\). At \(z=20\), even in the presence of the relative velocity, the number of 21cm absorption lines is less than \(O(0.1)\) for any optical depth and it becomes less than \(O(10^{-4})\) if the effect of relative velocity is included. For illustration, we have shown the 21cm absorption abundance in the case that minimum minihalo mass is determined by Jeans mass with \(T_{\mathrm{IGM}}=2\mathrm{K}\) at \(z=10\) and the kinetic temperature evolves as \(\propto (1+z)^2\). We next show the temperature dependence of the number of 21cm absorption lines. Fig.[\[fig:21cm_forest_temperature\]](#fig:21cm_forest_temperature){reference-type="ref" reference="fig:21cm_forest_temperature"} shows the abundance of the 21cm line absorption lines for the case that the IGM temperature is equal to CMB temperature at \(z=10\) (\(T_{\mathrm{IGM}}=30\mathrm{K}\)). When the kinetic temperature becomes larger, the Jeans mass becomes larger. In our formalism, the minimum minihalo mass is determined by Jeans mass. The contributions to 21cm absorption lines mainly come from smaller minihalos (which possess a larger optical depth than the larger minihalos ) and a larger kinetic temperature of the IGM reduces the abundance of 21cm absorption lines. In Fig.[\[fig:21cm_forest_temperature\]](#fig:21cm_forest_temperature){reference-type="ref" reference="fig:21cm_forest_temperature"}, for \(T_{\mathrm{IGM}}=T_{\mathrm{CMB}}\), we can see that the abundance of the 21cm forest reduces to less than \(\mathcal{O}(10)\) in both cases with and without the relative velocity. In particular, the abundance of the 21cm absorption lines is less than 1 if we take the relative velocity into account. We also find that the abundance of the 21cm absorption lines including the relative velocity with \(T_{\mathrm{IGM}}=2\mathrm{K}\) is very similar to that of the 21cm absorption lines which does not include the relative velocity with \(T_{\mathrm{IGM}}=T_{\mathrm{CMB}}\). Although this behavior is just a coincidence in this example, it illustrates that raising the IGM temperature and including the relative velocity generate the degeneracy for the abundance of the 21cm absorption lines.
# Summary & Discussion {#Sec5}
In this paper, we investigated the possible impact of dark matter-baryon relative velocity on the 21cm forest. The relative velocity generates the additional anisotropic pressure which suppresses the structure formation and consequently reduces the number of 21cm absorption lines produced by neutral hydrogen atoms in the minihalos. Compared with the 21cm emission lines from diffuse IGM, the advantage of the 21cm forest observation is to avoid the diffuse foregrounds such as the synchrotron emission from our galaxy because the 21cm forest observation utilizes the 21cm absorption lines in the spectra from bright radio sources. On the other hand, the use of radio background sources has a disadvantage because the 21cm forest strongly relies on the existence of the radio bright sources at a high redshift. According to, we can estimate the required minimum brightness of a radio background source for the detection of the 21cm forest by \[\begin{aligned} S_{\min }=& 17.2 \mathrm{mJy}\left(\frac{0.01}{1-F_{21, \mathrm{th}}}\right)\left(\frac{\mathrm{S} / \mathrm{N}}{5}\right)\left(\frac{5 \mathrm{kHz}}{\Delta \nu}\right)^{1 / 2}\\ &\times \left(\frac{1000 \mathrm{hr}}{t_{\mathrm{int}}}\right)^{1 / 2}\left(\frac{600 \mathrm{~m}^2 \mathrm{~K}^{-1}}{A_{\mathrm{eff}} / T_{\mathrm{sys}}}\right) \label{eq:Smin} \end{aligned}\] where \(F_{21,\mathrm{th}}=e^{-\tau}\) is 21cm transmission for 21cm optical depth \(\tau\), \(\Delta \nu\) is a frequency resolution, \(A_{{\rm eff}}/T_{{\rm sys}}\) is the ratio between an effective collecting area and a system temperature and \(t_{{\rm int}}\) is the observation time. In eq([\[eq:Smin\]](#eq:Smin){reference-type="ref" reference="eq:Smin"}), we normalize each quantity by the SKA-like specifications. Recently, over 280 quasars have been reported at \(z>5.7\). In these quasars, some of them are radio-loud active galactic nuclei (AGNs) such as the blazar PSO J030947.49+271757.31 with a flux density \(S_{147\mathrm{MHz}}=64.2\pm 6.2 \mathrm{mJy}\). The estimate of the number of radio quasars based on the extrapolations of the observed radio luminosity functions to a higher redshift indicates that there could be as many as \(\sim 10^{4}-10^{5}\) radio-loud quasars with a sufficient brightness at \(z\)=10. Several experiments, such as Low-Frequency Array (LOFAR) Two-metre Sky Survey (LoTSS), the Giant Metrewave Radio Telescope (GMRT) all-sky radio survey at 150MHz, and the Galactic and Extragalactic All-sky Murchison Widefield Array survey (GLEAM), are projected to detect hundreds of bright radio sources at \(z>\)`<!-- -->`{=html}6. These observation surveys and theoretical predictions support future 21cm forest studies. As seen in Fig.[\[fig:21cm_forest_temperature\]](#fig:21cm_forest_temperature){reference-type="ref" reference="fig:21cm_forest_temperature"}, the 21cm forest generated by neutral hydrogen atoms in the minihalos strongly depends on the thermal state of the IGM because the IGM temperature regulates the minimum mass of minihalos sensitive to the estimated number of 21cm absorption lines. Although the HERA experiment currently puts the lower limits on the IGM temperature at \(z\sim 8\), the uncertainty of the thermal state of the IGM at a high redshift is still large. Nevertheless, it would be worth mentioning that the higher IGM temperature can further suppress the number of 21cm absorption lines. So far, we assumed the conventional \(\Lambda\)CDM cosmology. The 21cm observations can be useful to probe different cosmological models. For instance, if we consider the warm dark matter or the axion-like ultralight particles, the number of 21cm absorption lines can be reduced due to the free streaming or \"quantum pressure\". On the other hand, if the axion-like particles are generated due to the spontaneous symmetry breaking after the inflation (so-called the post inflation PQ symmetry breaking scenario), the isocurvature perturbations can be generated which can dominate at the small scales and enhance the 21cm forest signals. The application of the 21cm forest to the study of dark matter such as the axion-like cold dark matter and warm dark matter in the presence of the dark matter-baryon relative velocity is also left for future work. While our paper focused on the suppression of small-scale structures due to the relative velocity, we also mention that the relative velocity can potentially help the formation of massive stars. For instance, the recent cosmological simulation has shown that relative velocity can play an important role to form massive stars and those massive stars become the seeds for supermassive black holes (SMBH). Supersonic relative velocity prevents the gas cloud formation until the rapid gas condensation is triggered in a protogalactic halo. When the gas cloud becomes gravitationally unstable, it is directly collapsed and the supper massive star, which is \(M \sim\)`<!-- -->`{=html}34,000 \(M_{\odot}\), can form. Such supermassive stars can become a seed of SMBHs. The phenomenology relevant to the relative velocity is rich and further study is warranted. The 21cm forest can be a useful probe on the small-scale structures which can go even beyond \(k\gtrsim 10 \mathrm{Mpc}^{-1}\), and we illustrated its usefulness by demonstrating that it is indeed sensitive to the dark matter-baryon relative velocity. The further application of 21cm forest observations to the small-scale structures such as its usefulness to study the subhalo structures would be also worth exploring and left for future work. | {'timestamp': '2022-12-19T02:10:23', 'yymm': '2212', 'arxiv_id': '2212.08409', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08409'} |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction
As a result of ongoing attempts to decrease the dependency on fossil fuels, the public interest in electric vehicles (EVs) has been steadily growing for the past decade. Consequently, the need for efficient batteries has also increased. In this work, we examine the Lithium Iron Phosphate (LiFePO\(_{4}\)) battery used in the Tesla Model 3. LiFePO\(_{4}\) batteries are a category of Lithium-ion (Li-ion) batteries which are widely used due to their high energy density, small size, and wide operating temperature range, ,. However, they are very sensitive to temperature variations. At high temperatures, Li-ion batteries degrade rapidly, while the low temperatures cause a reduction in power and energy output. Therefore, monitoring temperature evolution is crucial for EV manufacturers. Theoretically, heat transfer inside Li-ion batteries can be described by Equation [\[eq:34\]](#eq:34){reference-type="ref" reference="eq:34"}. \[\label{eq:34} \begin{aligned} mc_{cell}\dfrac{\mathrm{d}T_{cell}}{\mathrm{d}t} = I^2 R + T_{cell} \Delta S \dfrac{I}{nF} + Ah(T_{cell}-T_{amb}), \end{aligned}\] The first part represents the heat generated by the battery cell's internal resistance. The second term denotes the entropy change during discharge. Finally, the last term is the heat transferred to the ambience by convection. Traditional numerical methods for solving such problems require a new simulation for every new instance of the ODE and hence are computationally expensive. Therefore, there is a potential for alternative data-driven models that are fast, flexible, and robust in predicting battery temperature. Motivated by this potential, we make following contributions in this work:
1. We propose a data-driven approach for learning time and environmental condition-dependent neural operator that predict the temperature evolution of a LiFePO\(_4\) EV battery.
2. We introduce regularisation strategies to improve the robustness of the neural operator.
3. The connection of neural operators with a surrogate model allows for in-situ prediction of charging statistics.
4. Good generalisation is verified in terms of unseen driving data.
# Methodology
The experimental setup for acquiring the battery temperature data is shown in Fig.[\[fig:sensorinst\]](#fig:sensorinst){reference-type="ref" reference="fig:sensorinst"} in Appendix A. The temperature probe of an Inkbird IBS-TH1 was installed on the floor of the cabin of a Tesla Model 3 with a CATL LiFePo\(_{4}\) battery under the driver's seat. This setup also offers insights into the electrochemical system such as maximum charging power. We also collected different driving diagnostics from Tesla API, including the car's speed, power consumption, outside temperature, and battery level.
## Neural Operator
We model the baseline model of predicting battery temperature evolution in the form of Ordinary Differential Equations (ODEs) as shown in Equation [\[eqn:ode\]](#eqn:ode){reference-type="ref" reference="eqn:ode"}: \[\begin{aligned} \dfrac{\mathrm{d}u(t,\textbf{E})}{\mathrm{d}t}= f(t, u(t,\textbf{E})) \approx NN_{\theta}(t, \textbf{E}), \qquad t \in [0, T] \\ u(t_{0}, \textbf{E}_{0} ) = u_{0} \qquad \qquad \qquad \qquad \qquad \end{aligned} \label{eqn:ode}\] where u(t, **E**) is our hidden solution of temperature, t is relative time (duration of each drive session starting from \(t_{0}\) = 0 to final time \(T\)), **E** is the vector of environmental parameters, while \(u_{0}\) and **E\(_{0}\)** are the initial temperature and initial environmental parameters, respectively. We represent the unknown right-hand side \(f(t, u(t,\textbf{E})\) with a neural network \(NN_{\theta}\) (multilayer perceptron (MLP)) as shown in Equation [\[eqn:ode\]](#eqn:ode){reference-type="ref" reference="eqn:ode"}. It takes relative time and environmental parameters (power, speed, battery level, outside temperature, current battery temperature) as inputs and outputs the estimation of the neural operator at discrete time steps. The neural network is then trained by minimizing the mean squared error (MSE) between the prediction of our network and the ground-truth values for the differential operator calculated by forward differences (FD) as shown in Equation [\[eqn:mseloss\]](#eqn:mseloss){reference-type="ref" reference="eqn:mseloss"}. Thus, we learn the neural operator that describes the underlying physics of time-dependent temperature evolution in a supervised manner. \[\begin{aligned} \mathcal{L}_{NO} = \sum_{i=0}^{N-1} \ \norm{ \dfrac{u(t_{i+1},\textbf{E\(_{i+1}\)})-u(t_{i},\textbf{E\(_{i}\)})}{t_{i+1}-t_{i}}-NN_{\theta}(t_{i}, \textbf{E}_{i}) }_2^2 \\[2ex] \end{aligned} \label{eqn:mseloss}\] The trained neural operator is then used in Euler Forward numerical scheme to iteratively estimate the complete temperature evolution during inference with the time-step \(h\) as shown in Equation [\[eqn:euler\]](#eqn:euler){reference-type="ref" reference="eqn:euler"}: \[\begin{aligned} u(t_{i+1}, \textbf{E\(_{i+1}\)}) = u(t_{i}, \textbf{E\(_{i}\)}) + h* NN_{\theta}(t_{i}, \textbf{E}_{i}) \qquad for \quad i = 0, \dots, N-1 \\[2ex] \end{aligned} \label{eqn:euler}\]
## Regularisation
To explore further improvements to the baseline model, we introduce a regularizer \(\mathcal{L}_{smooth}\) minimizing the norm of the partial derivative of \(NN_{\theta}\) with respect to its input parameters to improve the smoothness of the prediction (see Equation [\[eqn:reg\]](#eqn:reg){reference-type="ref" reference="eqn:reg"}). \[\begin{aligned} \mathcal{L}_{smooth} = \dfrac{1}{M} \sum_{j=1}^{M} \norm{ \dfrac{\mathrm{d}NN_{\theta}(t, \textbf{E})}{\mathrm{d}E_{j}}}_2^2 \\[2ex] \end{aligned} \label{eqn:reg}\] where M is the total number of environmental parameters, \(E_{j}\) is one of the environmental parameters considered. Consequently, the total loss function is \(\mathcal{L}_{reg} = \mathcal{L}_{NO} + \lambda \mathcal{L}_{smooth}\) where \(\lambda\) is the weighting factor for the regularisation term. This model will be denoted as regularised LiFe-net.
## Neural Operator with Time Stability Loss
In this subsection we incorporate the temperature evolution into the training objective in terms of a time-stability loss \(\mathcal{L}_{TS}\), by introducing the Euler forward method into the training loop. It allows us to recover the complete time evolution of the temperature for each drive during training and do a multi-time-step optimization between the predicted and the ground-truth temperatures at once (see Equation [\[eqn:nostb\]](#eqn:nostb){reference-type="ref" reference="eqn:nostb"}). This enforces local smoothness like \(\mathcal{L}_{reg}\) as well as consistency with our training data. \[\begin{aligned} \mathcal{L}_{TS} &= \sum_{i=1}^{N-1} \norm{ u_{gt}^i-(NN_{\theta}(t_{i-1}, \textbf{E}_{i-1})*h+u_{pred}^{i-1}) }_2^2 \end{aligned} \label{eqn:nostb}\] where \(u_{gt}^{i}\) and \(u_{pred}^{i}\) are the ground-truth and predicted temperatures at time step \(i\), respectively. N is the number of time steps for the particular drive session. This model will be denoted as LiFe-net (time-stability) in this article. The training pipeline of this model is illustrated in Fig.[\[fig:training_pipeline\]](#fig:training_pipeline){reference-type="ref" reference="fig:training_pipeline"}
## Surrogate Model of Charging Statistics
The correlation of battery temperature to charging speed \(c_p\) and charging time \(t_{charge}\) is modeled by a linear regression model, viz. \[\label{eq:surrogate} \begin{split} c_p &= c_{soc} C_{bat} + c_{T} u_i(t_i, E_i) + o \\ t_{charge} &= c_{soc}^s C_{bat}^s + c_{soc}^e C_{bat}^e + c_{T}^{t} u_i(t_i, E_i) + o^{t} \\ \end{split}\] with \(c_{soc}\) explaining the contribution of state-of-charge \(C_{bat}\) and \(c_{T}\) scaling the contribution of battery temperature \(u_i(t_i, E_i)\). Charging time is also modeled by coefficients \(c_{soc}^s, c_{soc}^e\) corresponding to state-of-charge at the beginning \(C_{bat}^s\) and end \(C_{bat}^e\) of the charging session. \(c_{T}^{t}\) models the contribution of battery temperature while \(o, o^{t}\) introduce offsets to the predicted charging time and power. The connection of this surrogate model with our neural operator now allows for advanced driver's assistance in deciding whether to heat the battery or not (see Fig. [\[fig:pwr\]](#fig:pwr){reference-type="ref" reference="fig:pwr"} in Appendix A).
# Results and Discussion
The full dataset consists of 445 000 data measurements collected during 99 drives under different environmental conditions. The data was split with respect to individual drives leaving 92 drives for training and 7 for testing. Hence, the split for the training and test dataset is 80% and 20%, respectively. Table [\[table:Data\]](#table:Data){reference-type="ref" reference="table:Data"} summarizes the range for parameters used. All models were trained on the Taurus HPC system at TU Dresden using Tesla V100-SXM2-32GB GPU. Furthermore, we validate the performance of the models on the test data using MSE, MAE and relative error between predicted and ground-truth temperature values. The corresponding dataset and the implementation code are available at the following GitHub repository: <https://github.com/Jeyhun1/LiFe-net.git>
[\[table:mlparch\]]{#table:mlparch label="table:mlparch"}
## Prediction accuracy
We performed hyperparameter optimization of all discussed models using Weights and Biases library. The results summarised in Table [\[table:lambda\]](#table:lambda){reference-type="ref" reference="table:lambda"} show that regularised LiFe-net with \(\lambda\) parameter value of 0.1 outperforms the baseline model where \(\lambda\) = 0. Furthermore, Tables [\[table:lifenetreg\]](#table:lifenetreg){reference-type="ref" reference="table:lifenetreg"} and [\[table:lifenettstb\]](#table:lifenettstb){reference-type="ref" reference="table:lifenettstb"} (See Appendix A) display MSE for the different number of hidden layers and different numbers of neurons per layer for both regularised LiFe-net and time-stability LiFe-net, respectively. Finally, Table [2](#table:testdata){reference-type="ref" reference="table:testdata"} summarises the prediction accuracy of best performing models for each method considered in this work averaged over seven test data. Furthermore, comparison plots of all three models on individual test datasets are shown in Fig. [\[fig:tstb\]](#fig:tstb){reference-type="ref" reference="fig:tstb"}. The results show that the LiFe-net model trained with the time stability loss function performs the best, especially on data at the interface of the training data range (see Figure [\[fig:first\]](#fig:first){reference-type="ref" reference="fig:first"}). This can be explained by the multi-time step optimisation during training which the other two methods lack. However, this comes at the cost of considerably slower training time (see Table [2](#table:testdata){reference-type="ref" reference="table:testdata"}).
## Learning Charging Statistics from Temperature Data
The coefficients of our surrogate model explaining charging statistics as of Equation [\[eq:surrogate\]](#eq:surrogate){reference-type="ref" reference="eq:surrogate"} have been estimated by solving the corresponding least-squares objective function by normal equations. The regression model of \(c_p\) got a \(R^2 = 0.726\) while \(t_{charge}\) was explained even better since \(R^2 = 0.884\). The coefficients can be found in Table [\[tbl:surrogateModel\]](#tbl:surrogateModel){reference-type="ref" reference="tbl:surrogateModel"}. It can be observed that the state-of-charge at start of charging session is important for predicting both peak charging power (\(c_{soc}=-0.9557\), \(p=0.015\)) as well as-time (\(c_{soc}^s =-0.6213\), \(p=0.018\)). The reciprocal relationship implies that lower state-of-charge results in higher maximum charging power and therefore fast charging times since electrons can be transferred easier into the battery. Electrochemical processes also react faster in warmer environmental conditions which are reflected by the significant contribution of \(c_{T}\) and \(c_{T}^{\Delta_c}\).
[\[table:testdata\]]{#table:testdata label="table:testdata"}
[\[table:modelstime\]]{#table:modelstime label="table:modelstime"}
# Conclusion and broader impact
In this work, we introduced three variations of a Neural Operator that is learning the temperature evolution of a temperature probe attached to the surface of an EV battery depending on environmental conditions. We demonstrated that the LiFe-net trained with time-stability loss where we optimized the solution operator in a multi-timestep fashion is the most accurate. This information is then passed into a surrogate model that relates the measured temperature to charging statistics such as charging time and maximum charging power. Hence, this model enables EV manufacturers to get real-time never-before accessible data regarding battery sensitivities to numerous environmental conditions, which will give insights into designing more efficient battery systems in the future. | {'timestamp': '2022-12-19T02:10:07', 'yymm': '2212', 'arxiv_id': '2212.08403', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08403'} |
# Author Contributions {#author-contributions .unnumbered}
All authors conceived of the project. CR and AY worked on the design of the project. CR, AY and APM worked on the implementation. AY and CR produced the results reported in the manuscript. All authors reviewed, contributed and approved the final version of the manuscript
# Discussion and Outlook {#sec:outlook}
In this work we discussed an implementation of a virtual environment where the behaviour of swarms of agents controlled by SNNs could be optimized using evolutionary algorithms. The implementation combines NEST simulations for the SNNs, an environment described with NetLogo and an optimization with genetic algorithms using the L2L framework. This implementation is open source (see Supplementary Material) and can be used or extended to explore a variety of scientific questions regarding swarm optimization as well as emergent communication and organization. In particular, it could be of high interest to study the coordinated emergent behaviour and communication between agents in neurorobotics and cobotics applications, where it is essential to ensure safety, optimal resource usage, and adaptation to dynamic environments. We demonstrated that we could use genetic algorithms to successfully evolve spiking neural networks to control agents (here, ants) to solve a multi-agent food foraging task. Our approach shows emerging self coordination via pheromone as a result of the network optimization, without defining action rules or probabilistic state models. This work serves as a proof of concept to illustrate multi-agent interactions with emergent communication and self-coordination utilizing SNNs as underlying architectures. We provide a short summary of our methods in the following:
- Our SNN architecture encodes the inputs from the environment and steers the behaviour of the ant when foraging for food.
- Based on the fitness function, which evaluates the foraging performance, we utilize a genetic algorithm to optimize the synaptic weights and delays. This allows the network to achieve an efficient foraging behaviour.
- Our analysis investigates the emergent relationship established by the SNN between the inputs and outputs to trigger certain actions during the evolutionary process.
By applying the methods mentioned above, we observed an emergence of stigmergy, i.e. the communication and collaboration via pheromones. In our evolved network we proved that communication became essential for the efficient performance of the ants during the foraging task. Impairing the pheromone sensing capabilities of the ants drastically hindered the performance of the collective. This highlights the importance of stigmergic communication, which emerged through the optimization process, for the effective collaboration between the agents. In our experiments we showed that agents using pheromones were able to quickly forage for food even when the the position of the nest was changed. It is important to note that the release of pheromone does not obey any pre-defined rule or manual synaptic pre-configuration in the network. Instead, the stigmergy is an emergent behaviour that is triggered under certain conditions established during the evolutionary process. Furthermore, we would like to highlight that the sensing of the pheromone is not associated with any kind of pre-defined behavioural response in the network. There is no hard-coded synaptic communication that explicitly maps the sensory input of pheromone with a certain action. Instead, the observed behaviour is completely based on emergent communication resulting from the collective evolution of the colony. What we observe is that the optimization operates directly on the structure (synapse weights) of the brain networks of the ants, but there is also a higher level optimization in order to create a functional network created by the collaborative behaviour of the agents. This behaviour is achieved by the development of communication strategies between the agents using the pheromone signals. In essence, the ants form a new network where pheromone signals are used to coordinate the movement of the agents as a whole to efficiently solve the task. We can see this emergent functional effect as the correlations and anti-correlations between directed movement and the sensing of pheromone increase with each generation (see ). Based on it is possible to arrive at the conclusion that the pheromone dropping is uncorrelated to actions and sensing of the environment. However, optimized movement based on pheromone signals makes it possible to have higher concentrations of pheromone around the areas of interest and thus enhance the ability of the swarm to collectively solve the task. We also confirmed the emergence of this functional network when we disabled the pheromone receptor and the colony was not able to efficiently solve the task anymore (see ). Interestingly, the emergent communication via a simulated chemical signal resembles the use of positive (attractor) pheromones observed in real ant colonies. Although attractive trails of pheromone are more commonly found in nature, it has been observed that negative (repelling) pheromones are also used in certain species of ants . This negative pheromone also known as a "no-entry\" signal is used by the ants to mark unrewarding paths to prevent other cohorts from using them. It opens the question of whether further experimentation simulating different evolutionary trajectories could lead to the emergence of negative repelling pheromones among other sophisticated coordination mechanisms in self-organized simulated swarms. Another related question would be to study which environmental and evolutionary requirements are necessary for different signals to emerge within the swarm's communication repertoire. The usage of the L2L framework allows the modification of the outer loop optimization algorithm, opening the possibility to explore the impact of different evolutionary optimization features in the resulting communication patterns. At the beginning of this work we formulated two research questions. The results presented in the previous sections answer these. First, the usage of pheromone as a mean of communication and self-organized coordination between agents can emerge in the absence of predefined rules. Neither the mechanism to trigger the release of pheromone nor the mechanism of interpreting these signals was engineered as part of the behaviour of the agents. Through this we can infer that the simulated embedded physiological properties of the agents and the characteristics of the environment are exploited by the optimization algorithm. To answer the second research question, the results in illustrate the difference in the performance of the colony in terms of foraging efficiency. The evolved colony with stigmergic communication achieved a higher performance than the multi-agent rule-based colony. We would also like to address some of the main shortcomings of our work. First, the genetic algorithm requires several hundreds of generations until it provides a parameter setting which leads to a suitable task performance. We want to investigate other optimization techniques which may decrease the number of generations and result in a similar performance. Second, in this work we use a manually defined network architecture for all individuals. This limits the optimization process and the potential functional capabilities of the ants. Future work optimizing the network architecture could help find new strategies of the ants to use their physiology to efficiently perform the foraging task. Third, even if we did not define any rule about the relationships between input and output actions of the ants, the fitness function used to optimize the network is static and manually defined. This imposes evolutionary pressure on the ants towards specific solutions. Exploring other combinations of the fitness metrics will help us, in future experiments, understand better the strength of such pressure in the emergence of specific behaviour and solutions by the colony. Finally, as discussed in the results section, our analysis of the correlations between input and output is also limited and further work needs to be done to fully understand the relationships within the ant networks and the colony as a whole through time. As future work we want to analyze further the emergence of the communication features in the functional network created by the collective of ants and how this evolves through time guided by the genetic algorithm. In particular, we want to measure the effect of pheromone concentration and evaporation rate in the communication strategy. A way to do this would be to represent the whole colony as functional network where connectivity is defined by spatial coincidence in the environment between one or more ants, signals are represented by pheromone concentration and delays by the pheromone evaporation rate. We also want to explore the usage of plasticity within the ant networks and optimize the learning capabilities of the ants in the inner loop of L2L and see if this capability could further enhance the performance of the swarm. For example, we want to learn the foraging task utilizing synaptic plasticity in the inner loop and optimize the model's hyper-parameters in the outer loop. In particular, the emergence of connectivity motifs, as emerging from the usage of STDP in, could be further studied and correlated to specific communication functions in the colony. During pilot experiments (not shown in this manuscript) we observed that the emergence of stigmergy could be influenced by the number of neurons in the network of each agent. As future work we want to explore the minimal structural requirements (number of neurons, topology, types of synapses, etc) for the SNNs to enable the emergence of collective communication and its effect on functional efficiency of the swarm as a whole. The emergent foraging behaviour that has been described in this work resembles to some extent the self-organisation observed in social insects. Self-organisation is a process, where high order structure emerges. The emergent behaviour allows the swarm to collectively adapt to changing environments where the colonies become robust systems even if some individuals fail to perform their tasks. Robustness and adaptability are desired characteristics in the application of robotics, specially for the deployment in changing environments with harsh or extreme conditions where single agents are more prone to failure (e.g. space exploration, undersea navigation, disaster areas).
# Introduction {#sec:introduction}
Communication using chemical signals is an efficient way to interact and collaborate within swarm colonies. For example, social insects such as ants and termites deposit pheromones to guide their cohorts while foraging for food. The coordination does not follow a top-down hierarchy, i.e., there is no directing leader. Instead, the coordination is self-organized and enables the swarm to solve complex tasks, such as finding the shortest path between the nest and the food source. This behaviour is shaped through evolution over millions of years . From a computational point of view the self-organized behaviour in swarms described in the literature has been implemented by using (simple) action rules which allow the agents to interact with other cohorts and the environment by depositing chemical signals. The self-coordination of the swarm emerges as a result of these interactions, allowing the agents to solve problems, e.g. finding the shortest path in an environment. Defining explicit rules may help the system to solve specific tasks yielding in a foreseen behaviour, i.e. the system is guided to find a certain solution. Although this approach may converge into a solution it may not be the most efficient one. Moreover, defining rules, which lead to emergent interactions and self-coordination at the level of the swarm, may not be straightforward for complex tasks. The pre-defined rules are biologically implausible, since they include probabilistic actions, which may not reflect the natural behaviour of the swarm. In this work we investigate the emergence of self-coordination and communication in evolved swarms without defining any explicit rules which describes the behaviour of each agent. Specifically, our research questions are:
- Does **self-organized coordination emerge** within the swarm, based on the simulated physiological properties of the agents and the characteristics of the environment?
- Could the presence of a self-organized coordination between the agents **enhance the collaborative approach** to solve specific problems?
To tackle these questions, we define a task and evolve an entire swarm of agents searching for food and bringing it back to the nest, resembling the ant colony foraging behaviour. In particular, we are interested in exploring the role of stigmergy in the evolutionary emergence of communication within the swarm. Stigmergy refers to indirect interactions which take place via the environment for the purpose of navigation, communication or collaboration. Ant colony foraging is a self-organized behaviour which is well investigated within the literature. In these works several computational models replicate to some extent the dynamics observed in biological colonies. For instance, the authors of and similarly of compare the foraging behaviour of the Argentine ants Linepithema with an implementation of a probabilistic model. Their models are able to find the shortest paths in a maze to a food source relying on self-organization based on pheromone trails. In, the authors show the result of implementing an ant swarm algorithm to control simulated and real robots in environments, where the distribution of food changes. They control each robot using a rule-based probabilistic model mimicking the foraging behavior of seed-harvester ants and optimize the performance of the collective behaviour using a genetic algorithm. They analyze how the collective strategies and communication are evolved. The model presented in illustrates self-coordination and collaboration between ants through the use of pheromone. This model is utilizing simple, pre-defined rules to control the local decisions of the ants. In an alternative approach, propose a framework to soft control agents, called shills, which are able to follow designed update rules. Their task is based on an iterated prisoner's dilemma game, where shills can update their strategies using particle swarm optimization (PSO) mechanisms. In their setting, shills can choose better performing strategies and provide those to other individuals. The frequency of the communication between the shills can be steered by different parameter configuration. Swarms have also been combined with spiking neural networks (SNNs) in the past. In , the authors implement a spatio-temporal model called \"SpikeAnts\" and depict the emergence of synchronized activities in an ant colony. They describe a sparsely-connected network of spiking neurons, where each ant is represented by two interconnected neurons. The network does not receive any external stimuli and does not incorporate learning rules. The emergent behaviour shows a distributed learning process in the population of agents and it is concluded that in their setting the swarms is able to self-organize. The aforementioned works describe and analyze emergent communication due to chemical signals, however, these approaches are probabilistic state models or rule based systems, which enforce a collaborative behaviour. If we want a swarm of robots to optimally use their physiology to achieve certain tasks in a collaborative and adaptive way, trying to formalize all the rules and interactions which address not only individual but also collective behaviours in a dynamic realistic environment becomes complex. This approach may also not scale well if the complexity of the task or the number of agents increases. Leveraging emergent behaviour could provide a reasonable and adaptable approach to deal with this scenario. For this reason, we implemented a virtual environment where the behaviour of swarms of agents can be optimized using evolutionary algorithms. With these tools we study and analyze the benefit of evolved emergent behaviour against rule based approaches in solving a task where coordination plays a key role in the performance. In our approach agents do not follow pre-defined rules to guide their behaviour. This is a critical difference to previous work, as with this approach we do not steer the learning process towards specific solutions derived from our own understanding of how these complex systems work. Instead, we give the optimization process full control over the solutions it will explore in order to maximize the fitness measures. In our setting, each agent is controlled by an SNN composed of several neurons. The agents perceive their environment through several neurons which receive sensory information and implement a mapping from this input to output behaviours. In general, spiking neural networks have been proposed as alternative methods to solve machine learning tasks which offer biologically realistic configurations in comparison to artificial neural networks or state models. SNNs have the capability to control agents and robots in dynamical environments. We optimize the synaptic weights and spike time delays of the entire network using genetic algorithms, so that the network is able to control the agents to navigate in the virtual environment while searching and collecting food and bringing it back to the nest. At the beginning of the evolutionary process, the SNN is agnostic of the foraging task and there is no pre-defined communication or coordination mechanism encoded into the architecture of the SNN. There is no explicit mapping between the sensory input and the actions taken by the ants, e.g. visual information and pheromone depositing. The virtual environment, in which the simulations are run, is based on the ant-colony model by . In this model, the diffusion and evaporation of pheromone is simulated in a 2-dimensional grid world. For an example of a possible distribution of pheromone for an active swarm, see . The ants colored in red are exploring the world for food (green leaves) and depositing pheromones near the food patches and in the vicinity of their nest. While some ants are following the pheromone trail leading to the food patches or to the nest, others are exploring the environment. Every ant is steered by a spiking neural network. To optimize the network parameters and create the environment, we connect the Learning to Learn framework with the multi-agent simulator NetLogo and the spiking neural network simulator NEST . Our manuscript is structured as follows. First, in Section [3](#sec:methods){reference-type="ref" reference="sec:methods"}, we introduce our toolchain and setup of the environment. We describe the structure of the agent network and how we employ a genetic algorithm to optimize the synaptic weights and delays over multiple generations. Then, in Section [4](#sec:results){reference-type="ref" reference="sec:results"} we analyze the emergent communication via pheromones and compare our network structure to the base model implemented in in terms of foraging performance. We examine how the pheromone usage and concentration is shaped by the evolutionary process which enhances the self-coordination and communication between the ants. Moreover, we investigate the correlation between the input and output spike activity produced by an agent's network. This analysis shows us that certain input actions lead to specific output reactions. We measure the significance and relate these actions to each other. Finally, we discuss the results of the swarm simulations, with special interest on the emergent communication, and we mention some avenues for future work. As a summary, in this manuscript we show that agents without pre-defined rules can self-organize through evolutionary optimization which allows them to solve complex tasks as a swarm, leveraging their physiological capabilities. Interestingly, in our specific use case of an ant colony, we observe that the pheromone acts as an attractor rather than an event-based signalling mechanism, which resembles the behaviour in biological colonies. This work provides a proof of concept environment which can be used to study learning and emergence of complex behaviour in multi-agent settings.
# Material and Methods {#sec:methods}
We carry out optimization of a spiking neural network to steer an ant colony foraging for food using a genetic algorithm (GA). The optimization workflow is depicted in . There are \(32\) individuals per generation. Each individual corresponds to an instance of an SNN (see ) It is important to note, that an individual refers to the SNN instance and not to a simulated ant. There are \(15\) ants per simulation forming the colony and every ant is controlled by an SNN, which is a copy of the same individual (network). The GA optimizes the connection weights and spike time delays of each network. Since there are \(32\) individuals we are running \(32\) simulations in parallel. Thus, we execute \(15\) identical networks during one simulation per individual and \(480\) networks in total per generation. Each colony is simulated within a 2D environment to perform the foraging task. After the simulation finishes, a fitness value is calculated to evaluate the performance of the run.
## Toolchain {#subsec:toolchain}
Our main toolchain[^1] consists of three applications: The meta-learning framework Learning to Learn , the multi-agent simulator NetLogo and the spiking neural network simulator NEST . **L2L** (Version 1.0.0-beta) is the orchestrator of the workflow (see Supplementary Material for L2L optimization details). The framework requires the definition of the optimization algorithm, including its (hyper-)parameters, and the simulation whose parameters are going to be optimized. The software architecture consists of a two loop structure, the inner loop (optimizee) and the outer loop (optimizer). In the inner loop, the optimizee--for example a neural network with or without learning capabilities--is trained on a family of tasks. In our setting, the SNNs in the inner loop do not learn and only the performance on the foraging task is evaluated, the optimization is conducted in the outer loop. The performance of the training or simulation run is evaluated by a fitness function, yielding a fitness which can be a scalar or a vector value. In the outer loop, the (hyper-) parameters of the optimizee are optimized. For example, if the optimizee is a network, the parameters could be the learning rate and the weights. The aim is to reach an overall good performance and to enable faster learning on new, unknown tasks. In this work we use the L2L framework's genetic algorithm as optimizer in the outer loop, which wraps the implementation from the DEAP library. **NetLogo** (Version 6.2.0) is a multi-agent simulator. Every object in a NetLogo simulation is an agent and can be used as means of communication with other agents. With NetLogo it is easy to replicate the same simulation and agent. In our settings the simulations are spiking neural networks and the agents are ants. Each ant is controlled by one instantiation of a common spiking neural network. To communicate with each other the ants can deposit pheromones. Patches are used to represent the pheromone which is dropped by the ants on the environment. Each ant is also capable of sensing the presence of pheromones. An important difference to notice in this implementation as compared to the works discussed in Section [2](#sec:introduction){reference-type="ref" reference="sec:introduction"}, is that the dynamics of the pheromone depositing and sensing are not hard-wired in the ant network model. NetLogo allows the simulation of the pheromone dynamics through its embedded language primitives. The simulations uses the spiking neural simulator **NEST** (Version 3.0) as the back-end. NEST is a spiking neuronal simulator for biologically plausible, large-scale neural network simulations. Simple neuron models with biological realistic connection structures can be simulated in an efficient way in a variety of hardware from local machines up to High Performance Computing (HPC) systems . NEST supports multi-threading and parallel execution with the message passage interface (MPI).
## Ant colony foraging for food {#subsec:ant_colony}
In this section we describe the experimental setup for the ant colony. The agents have to search for food and bring the food back to the nest, shown as a brown/black patch in . Each ant can only collect one food patch at a time and must transport it back to the nest before looking for food again. There is no pre-programmed behaviour to signal the ants that this is the activity pattern they should follow. This pattern is enforced by not allowing ants to carry more than one piece of food and rewarding ants when they bring the food to the nest, at which point they also regain the capability to pick up another piece of food. In every generation we change the position of the nest randomly, in order to avoid overfitting the network and make the colony more adaptable to the location of the food piles. As depicted in , every agent can deposit pheromone which is represented by a white patch. The dropped pheromone spreads from the point (patch) of release to a certain amount of neighboring patches according to the diffusion rate, defined in the simulation parameters. Pheromone concentration also decreases (evaporates) exponentially over time. There is no predefined behaviour that binds depositing pheromone to any action or sensor. Other agents can smell the pheromone and react to it. The response elicited in an agent by the pheromone emerges from the evolutionary process carried out by the genetic algorithm. For example, the sensing of pheromone at a certain location could modify the trajectory of an agent looking for a food source or the nest. Each agent is controlled by its own network, a copy of the individual SNN. An example of such a network is shown in . The network consists of \(36\) neurons and \(720\) connections. The agents perceive their environment through sensory stimuli from various modalities such as visual, mechanoreceptory or chemical (pheromone) sensors while navigating. The input neurons that receive the sensory information are connected to a second layer of neurons which are all-to-all connected. By having an all-to-all connectivity we avoid constraining the network to a pre-defined topology. At the same time the second layer is connected to the output neurons which result in the agent's actions such as movement and pheromone laying. The first twelve neurons correspond to the input layer. The top three neurons in are receptor neurons which react to the direction of the pheromone. The next three neurons can locate the nest. The middle of the nest is indicated by the queen receptor. The reward and punishment for the ant are determined by the reward receptors and nociceptors. Whenever food or the wall is seen the green and red photoreceptors are activated correspondingly. The heartbeat neuron stimulates the network in every timestep by supplying a small direct current in order to maintain activity in the network. The middle layer comprises twenty neurons, which are connected in an all-to-all manner. The four output neurons are fully connected with the middle layer and control the movement (rotate left, rotate right, go forward) as well the activation to drop the pheromone. The total fitness of the ant colony optimization is calculated by summing up the rewards and punishments received by each of its ants in the colony during the course of a simulation. An ant receives a small positive reward for touching a food patch, and a large reward for returning with food to the nest. This induces the ants to quickly return to their nest, whenever they find food. A small punishment is given at every time step to encourage the colony to complete the task as quickly as possible. At the same time, the ants receive a small punishment for every selected action to avoid excessive reactions, such as depositing too much pheromone or exhibiting too much movement. The simulation takes \(T = 2000\) steps to finish; if the ants are able to collect all the food before the simulation ends they receive a reward as well. This reward is the result of the difference between the total simulation time and the time \(T_{s}\) the ants require to collect the food multiplied with a scalar \(\eta\). The whole process can be formalized as: \[f_{i} = \sum_{t=1}^{T_{s}} \left(\sum_{j=1}^{J} \mathcal{N}^{(t)}_{i,j} + \mathcal{F}^{(t)}_{i,j}-\mathcal{C}^{(t)}_{i,j} \right) + \eta \left(T-T_{s}\right), \label{eq:fitness_antcolony}\] where \(t=1, \ldots, T_{s}\) is the simulation step with \(T_{s}\leq T\) and \(T\) the total simulation time, \(\eta\) is a scalar to weight the speed of the ants collecting the food. Within the sum, \(i\) is a specific individual, \(j=1,\ldots, J\) is an ant and \(J\) is the total number of ants in a colony. \(\mathcal{N}\) is the positive reward value for coming back to the nest with food, \(\mathcal{F}\) is a positive value for touching the food and \(\mathcal{C}\) is the punishment cost. The behavioural reward and cost values are specified in the Supplementary Material (Table S1) and the other experimental parameters can be found in Supplementary Material (Table S2).
# Results {#sec:results}
In the following, we show the result of executing simulations of the foraging task using the L2L framework. See Supplementary Material for further details on the execution. We explore and compare three models:
1. an SNN-driven model, where each ant is controlled by identical copies of a common network,
2. the same SNN-driven model but with the pheromone pathway deactivated (sensing and depositing of pheromone is disabled),
3. a rule-driven model based on the implementation in NetLogo by . In this model each ant follows a set of predefined simple rules.
## Fitness and emergent usage of pheromone {#subsec:fitness_pheromone}
shows the evolution of the mean and best fitness over \(1700\) generations for the SNN-driven experiments mentioned above. In the first SNN-driven model the pheromone pathway is enabled, i.e. the capability of the ants to deposit and sense the pheromone is active. For this model, we observe that the mean fitness is increasing over the generations. After the \(1000\)th generation the mean converges and oscillates around a value of \(8100\) with a standard deviation of \(3970\). Specific individuals reach a fitness higher than \(30000\) with the highest fitness at \(38004\) in generation \(1420\) (see green data series). The second SNN-driven model has no active pheromone pathway, the ants are not able to perceive or deposit the pheromone. Here, we observe an increasing mean, which converges after \(700\) generations around a value of \(5800\) with a standard deviation of \(3950\). In comparison to the first model, specific individuals reach a high fitness in early generations (e.g. generation \(390\) with a fitness of \(21716\)) and the highest fitness at \(24450\) in generation \(1150\). The increase of performance of both models demonstrates that the genetic algorithm in our approach is capable of evolving the SNNs to achieve a much higher performance than that of a random network initialization, which exhibited a negative fitness. Furthermore, the colonies that are able to communicate via pheromones can be adapted via learning to achieve a higher fitness and performance than the colony without the ability to utilize pheromone. Beyond the evolution of the fitness, we examined the emergence of coordination via pheromones. The heatmap in shows the pheromone concentration for an example individual over different generations. Each simulation runs for \(2000\) steps and at step \(1900\) we take a snapshot of the best individual with the position of the ants, the pheromone and the food patches, as well as the pheromone concentration and the food amount. In the early generations (a) the concentration is low, while most of the food patches are uncollected. The ants are exploring the environment and, from visual observation (as seen in a), about half of the population is close to the nest as well as to the food pile on the right side (\(x:35, \, y:30\)). Small pheromone patches can already be observed around the nest and the closest food patch. This is also visible in the marginal plots depicting the pheromone density. The peaks are at the position of the nest (\(x:22, \, y:30\)) and the right food pile. In generation \(200\) (b) it can be observed that there is a correlation between the pheromone concentration and the position of the food piles and the nest. The marginal plots have \(4\) peaks, at the nest and at the left and right food piles (bottom left \(x: 5, \, y: 5\), top left \(x: 5, \, y: 52\), right \(x:35, \, y:28\)). In generation \(600\) a higher pheromone usage is visible (c) in comparison to generation \(200\). The concentration is nearly equally distributed over the environment, which is also visible in the density plots. The final snapshot shows generation \(1500\), which is similar to generation \(200\), but the pheromone is more concentrated and exhibits a clear pheromone path. In comparison to earlier generations, the smaller number of food patches indicate that more have been collected, thus a better task performance. The density plots depict the highest pheromone concentration in the middle, where the nest is positioned and at the left bottom.
## Performance comparison {#subsec:performance_comparison}
To assess the performance of the ant colony, we first selected the best performing individual of the last generation after the optimization procedure. Secondly, we calculated over \(100\) simulations (trials) the average amount of food foraged by the colony (and brought back to the nest) within the simulation time of \(2000\) steps. The trials have different seeds which determine the initial position of the ants in each trial. There are three piles of food positioned in different areas of the simulated 2D-world as shown in . Each pile contains \(50\) units of food making a total of \(150\) units, which is the maximum score that a colony can reach in this experimental setup. We tested the three models introduced in the beginning of . Furthermore, we added a forth experiment in which we impaired the pheromone pathway for the first SNN based model. The results are shown in . The first bar illustrates the foraging performance of a rule based system implemented in NetLogo , i.e., the ant foraging behaviour follows pre-defined action rules. The self-organized behaviour of the ants emerges from the interactions of individual colony members. In this setting, the colony reaches a mean of \(124.5\) foraged units of food with a standard deviation of \(19.6\). We define the rule based system as our baseline for comparison. The second bar depicts the SNN based model 1 which reaches both a higher mean score (\(135.7\)) and a lower standard deviation (\(9.0\)) than the rule-based system. To measure the effect of the pheromone on the colony coordination, we disable the pheromone sensing in the first SNN model after evolving it. We observe that the foraging performance drops significantly and the ants are not able to collect the food (mean: \(0.6\), SD: \(0.9\)) as shown in the third bar. The fourth bar shows the performance of the second SNN based model. It reaches a mean score of \(97.1\) with a standard deviation of \(60.0\) and performs less efficient than the first SNN based model and the rule based one. Based on these results, we can conclude our evolved model with a self-coordination system is able to complete the foraging task in a shorter amount of time than both the rule-based system and the second SNN model which is solely using the visual sensory pathways to navigate and forage for food. The ants are able to learn to communicate via pheromones, which increases their performance when foraging for food, despite the fact that there is no hardwired mapping that defines the usage of pheromones for communication purposes. The communication emerges within the ant colony over the generations as a swarm strategy.
## Input-output correlation analysis {#subsec:network_analysis}
In order to better understand how the ant behaviour changes over the generations, we correlate all input spike trains with all output spike trains for all ants of the best individual for specific generations. The results of this correlation analysis are limited and provide just an initial overview of the relationships between input and output in the network. A more complete assessment of the correlations in a recurrent network with dynamic inputs requires a complex analysis which considers not only snapshots of inputs and outputs but also the internal state of the model and its transitions through time. This type of analysis is out of scope for this work and will be the topic of further publications. First, we create a histogram to bin the input and output spike trains using the Electrophysiology Analysis Toolkit. As a second step, we obtain the mean Pearson product-moment correlation coefficients. depicts the correlation heatmap for the input (x-axis) and the output (y-axis) activity (see also for the network description) for four different generations. While a red color indicates a high correlation, a blue color denotes a negative correlation. The nomenclature follows the naming scheme in , i.e. for the output activity *Left* is rotate left, *Move* means move forward etc. and similarly for the input activity. Across all generations, *Smell Middle* is positively correlated with *Move*, with a value of \(0.265\) in generation \(10\) and a reduced value of \(0.116\) in generation \(1600\). One can also see that e.g. *Smell Right* is negatively correlated with *Move*, it has a value of \(-0.487\) in generation \(10\), but at generation \(1600\) the strength of this anti-correlation is reduced to \(-0.182\). Interestingly, *Smell Left* is negatively correlated with *Left* in generation \(10\) (\(-0.213\)) but shows a positive correlation in later generations (\(0.219\) in generation \(1600\)). A similar observation can be made for *Smell Right* and *Right*. We observe in early generations a negative correlation between *Move* and *On Nest* (\(-0.478\) in generation \(10\)), while in later generations there is a trend towards becoming uncorrelated (\(-0.0017\) in generation \(1600\)). *Nest Middle* and *Left* are anti-correlated in early generations (\(-0.187\) in generation \(10\)) and positively correlated in later generations (\(0.135\) generation \(1600\)). The relationship between *Visual Green* and *Pheromone* indicates a slight positive correlation in early generations (\(0.136\) in generation \(10\)) but increases in later generations (\(0.313\) in generation \(1600\)). The fitness function punishes an extensive pheromone drop which leads to a rather restrained usage. The input-output relationship is changing continuously across the generations due to the evolutionary algorithm, i.e. certain input-output relationships get enhanced or weakened. Some of these mappings align with observations in nature, e.g. *Visual Green* and *Pheromone* have an increasing correlation value across the generations, which reflects the positive reinforcement when the ant perceives food. This is also in agreement with the attractor effect of the pheromone and links to the correlations between smelling pheromones and ant movement, serving as guiding signals to take the ants back to the food patches. Although the pheromone heatmap provides insight into ant behaviour for specific generations, we wanted to see how the behaviour evolves over all generations. In order to compare different input-to-output correlations and validate our results obtained from the heatmap, we calculate the correlation coefficients between input and output spike trains for the first ant of the best individual every tenth generation. We then fit a linear regression on the coefficients, which returns a p-value and the slope s; a low p-value rejects the null-hypothesis that the slope of the curve is zero, i.e. there is not a correlation. Additionally, we calculate the standard deviation \(\sigma\) of all points. The visualization in enables us to determine trends in the ant behavior over generations. For example, we observe that *Visual Red* is initially positively correlated to *Move* but there is a trend towards anti-correlation over the generations. While *Visual Green* and *Move* show an increasing trend, *Visual Red* and *Move* become increasingly negatively correlated over the generations. Likewise, *Smell Right* shows a trend towards an increasing positive correlation with *Right* and a negative correlation with *Left*. *Smell Left* is initially positively correlated to *Move*, but this correlation decreases for later generations. In contrast, the correlation between *Smell Left* and *Nest Left* increases, while it decreases for *Nest Right*. The receptor for *Smell Middle* shows an increasing trend towards positive correlation with *Move* *Nest Middle* has an increasing positive correlation with *Left*, while *Nest Right* has an increasing negative correlation with *Left* as the colony evolves. Pheromone is initially dropped in the vicinity of the nest. However, the pheromone output activity, when the ant is in the middle of the nest, decreases over the generations. In contrast, *Visual Green* and *Pheromone* show an increasing positive trend. Observing the performance of the ant colony at the end of the evolutionary process, pheromone paths seem to emerge and form attractors and trails with higher pheromone concentration resulting from a combination of optimized movement trends while sensing pheromone, the evaporation constant of the pheromone, and the evolutionary pressure of the fitness metric towards the pattern explore-get food-return food to nest. This resembles the actual behaviour of ant colonies observed in nature.
[^1]: Our simulation and analysis scripts are publicly available at <https://github.com/Meta-optimization/L2L/tree/neuroevolution/> | {'timestamp': '2022-12-19T02:12:08', 'yymm': '2212', 'arxiv_id': '2212.08484', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08484'} |
# Introduction and background
Let \(F\) be an arbitrary field and \(N\) a positive integer. By \(M_N(F)\) we denote the set of all \(N \times N\) matrices over \(F\), and by \(\mathcal{N}_N\) we denote the set of all nilpotent elements in \(M_N(F)\). Note that the group \(GL_N(F)\) acts on \(\mathcal{N}_N\) by conjugation, i.e., \(g\cdot X=gXg^{-1}\) for \(g \in GL_N(F)\) and \(X \in \mathcal{N}_N\). Recall that each nilpotent \(N\times N\) matrix over \(F\) is similar to a nilpotent matrix \(J\) in the Jordan canonical form. If we assume that Jordan blocks of \(J\) are ordered by non-increasing size, then such \(J\) is uniquely defined. Hence the set of all the partitions \({\bf{\lambda}}=(\lambda _1,\ldots ,\lambda _t)\), \(\lambda _1\geq\ldots \geq\lambda _t\), \(\sum_{i=1}^t \lambda_i=N\), of the integer \(N\) is in a bijective correspondence with the set of all orbits of \(\mathcal{N}_N\) under the action of the group \(GL_N(F)\). The orbit corresponding to the partition \({\bf{\lambda}}\) is denoted by \(\mathcal{O}_{{\bf{\lambda}}}\). It was first proved by Gerstenhaber that the dominance ordering on the set of partitions of number \(N\) coincides with the inclusion ordering of orbits of \(\mathcal{N}_N\). The algebraic and combinatorial properties of nilpotent orbits over various fields and in other reductive Lie algebras have been exhaustively studied, see e.g.,. In the last decade, there has been an extensive investigation of Jordan structures of multiplication maps on finite-dimensional commutative algebras. If either the algebra is local and \(x\) is an element of the maximal ideal or the algebra is graded and \(x\) has a positive degree (for example, \(x\) is a linear form), then the multiplication map by \(x\) is a nilpotent linear map on the algebra, whose Jordan structure is called the Jordan type of the element \(x\). The "generic\" Jordan type is an important invariant of the algebra, which is tightly connected to the Lefschetz properties, see e.g.,. Recently, the pairs of commuting nilpotent operators have been extensively explored, see e.g.,. Let \(B\in \mathcal{N}_N\) be an arbitrary nilpotent matrix. Let \(\mathcal{C}(B)\) denote the *centralizer of \(B\)*, i.e., \(\mathcal{C}(B)=\{A\in M_N(F);AB=BA\}\), and let \[\mathcal{N}(B)=\mathcal{C}(B)\cap \mathcal{N}_N\] denote the *nilpotent centralizer of \(B\)*, i.e., the set of all nilpotent elements of \(\mathcal{C}(B)\). It is known that \(\mathcal{N}(B)\) is an irreducible variety, . One of the classical questions in the matrix theory is to determine the intersection of the nilpotent centralizer with the nilpotent orbits.
In general, this seems to be a difficult question, which has been answered so far only for some specific cases of matrices \(B\). For example, the nilpotent centralizer of a single Jordan block \(J_{(N)}\in {\mathcal{O}_{(N)}}\) intersects only \(N\) nilpotent orbits. This is the consequence of the fact that the only matrices commuting with \(J_{(N)}\) are polynomials in \(J_{(N)}\), see e.g.,. Thus the partitions \({\bf{\lambda}}\) that correspond to the matrices commuting with \(J_{(N)}\) are precisely the ones whose parts differ by at most 1, i.e., \(\lambda _1-\lambda _t\le 1\). Such partitions are called *almost rectangular*, . Following, we denote the unique almost rectangular partition of \(N\) into \(k\) parts by \[[N]^k=\left((q+1)^r,q^{k-r}\right),\] where \(q=\left\lfloor\frac{N}{k}\right\rfloor\), \(r=N\mod k\), and \(m^t\) abbreviates \(t\) copies of \(m\) in the partition. On the other hand, the matrices over a finite field or over an algebraically closed field of characteristic 0 with largest intersections with nilpotent orbits were characterized in ; if \(N\geq 4\), then \(B^2=0\) if and only if \(\mathcal{N}(B) \cap \mathcal{O}_{{\bf{\lambda}}} \ne \emptyset\) for all partitions \({\bf{\lambda}}\) of \(N\). For a general matrix \(B\), only partial answers to Question [\[commuting_partitions\]](#commuting_partitions){reference-type="ref" reference="commuting_partitions"} were given. The lower bound for the number of parts of \({\bf{\lambda}}\) was determined in and the upper bound for \(\lambda_1\) was given in . Moreover, it was proved in that if \({\bf{\mu}}'\) is the conjugate partition of the Jordan canonical form of \(B\), then \(\mathcal{N}(B) \cap \mathcal{O}_{{\bf{\mu}}'} \ne \emptyset\). Since \(\mathcal{N}(B)\) is an irreducible variety, there exists a partition \({\bf{\lambda}}\) of \(N\) such that \(\mathcal{O}_{{\bf{\lambda}}}\cap \mathcal{N}(B)\) is a dense open subset of \(\mathcal{N}(B)\). If \({\bf{\mu}}\) is the partition of \(N\) that corresponds to the Jordan canonical form of \(B\), then \(\mathcal{Q}({\bf{\mu}})\) is defined to be the unique partition \({\bf{\lambda}}\) such that \(\mathcal{O}_{{\bf{\lambda}}}\cap \mathcal{N}(B)\) is a dense open subset of \(\mathcal{N}(B)\), which is the maximal partition in the dominance ordering, for which its nilpotent orbit intersects \(\mathcal{N}(B)\). Recently, the conjectured recursive algorithm to compute \(\mathcal{Q}({\bf{\mu}})\) (see results and discussions in ) has been proved to work for general \({\bf{\mu}}\), . Furthermore, the preimage \(\mathcal{Q}^{-1}(\lambda)\) was completely determined for partitions \(\lambda\) with two parts which differ by at least \(2\) . For generalization of \(\mathcal{Q}(\mu)\) to other simple Lie algebras, see . Motivated by these partial results and open problems, this paper considers and completely resolves the following special case of Question [\[commuting_partitions\]](#commuting_partitions){reference-type="ref" reference="commuting_partitions"}.
There are some known partitions \({\bf{\lambda}}\), for which the answer to Question [\[nn\]](#nn){reference-type="ref" reference="nn"} is positive. Namely, since the centralizers of square-zero matrices intersect all nilpotent orbits, any partition of the form \({\bf{\lambda}}=(2^t,1^s)\), \(2t+s=2n\), gives rise to a nilpotent orbit that intersects \({\mathcal N}(B)\), . Moreover, we can assume that \(B\) is already in its Jordan canonical form, thus \(B=J_{(n)}\oplus J_{(n)}\). Since every matrix commutes with its polynomial, it follows that \(B\) commutes with \(p(J_{(n)})\oplus q(J_{(n)})\), for any polynomials \(p,q\in F[X]\). By the above results, it follows that \({\bf{\lambda}}=([n]^t,[n]^s)\), where \(1\leq t,s\leq n\), gives a positive answer to Question [\[nn\]](#nn){reference-type="ref" reference="nn"} as well. Also, in it was proved that any nilpotent orbit corresponding to an almost rectangular partition of \((2n)\) intersects \({\mathcal N}(B)\). Furthermore, it was proved in the same paper that all nilpotent orbits that give a positive answer to Question [\[nn\]](#nn){reference-type="ref" reference="nn"} with the corresponding partition \({\bf{\lambda}}\ne (2n)\) have \(\lambda_1\leq n+1\), and \({\mathcal O}_{(n+1,n-1)}\) has a non-empty intersection with \(\mathcal{N}(B)\). It was noted in that these partitions do not constitute the complete list of nilpotent orbits intersecting \({\mathcal N}_{B}\). Some of the partitions \({\bf{\lambda}}\) obtained by block antidiagonal and block upper triangular nilpotent matrices commuting with \(B\) were listed in . The main objective of this paper is to completely resolve Question [\[nn\]](#nn){reference-type="ref" reference="nn"}, and to give a complete list of nilpotent orbits \(\mathcal{O}_{{\bf{\lambda}}}\) having a non-empty intersection with \({\mathcal N}(B)\), where \(B\) has the Jordan canonical form \((n,n)\). We will prove the following theorem.
# Proof of Theorem [\[main\]](#main){reference-type="ref" reference="main"} {#sec:proof}
Let \(B\in \mathcal{N}_{2n}\) be a nilpotent matrix with the Jordan canonical form \((n,n)\). By Lemma [\[A=B\^zC\]](#A=B^zC){reference-type="ref" reference="A=B^zC"}, any \(A\in {\mathcal N}(B)\) is of the form \(A=B^zC\) for some non-negative integer \(z\) and \(C\in {\mathcal N}(B)\). If \(A=0\), then the corresponding partition is \([2n]^{2n}\), which is listed in Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1). In the rest of the proof we therefore assume that \(A\) is non-zero, and hence we assume that \(z\) is maximal possible such that \(C\) is nilpotent. We will use the Weyr form and write \(B\) as in [\[eq:B-Weyr\]](#eq:B-Weyr){reference-type="eqref" reference="eq:B-Weyr"} and \(C,C_i\) as in Lemma [\[A=B\^zC\]](#A=B^zC){reference-type="ref" reference="A=B^zC"}, or in the polynomial notation and write \[B=\left[ \begin{array}{cc}X&0\\0&X \end{array} \right], C=\left[ \begin{array}{cc}a(X)&b(X)\\c(X)&d(X) \end{array} \right], \text{ and } A= \left[ \begin{array}{cc} X^za(X)&X^zb(X)\\ X^zc(X)&X^zd(X) \end{array} \right],\] where \(a(X)=a_0+\cdots +a_{n-z-1}X^{n-z-1}\), \(b(X)=b_0+\cdots +b_{n-z-1}X^{n-z-1}\), \(c(X)=c_0+\cdots +c_{n-z-1}X^{n-z-1},\) and \(d(X)=d_0+\cdots +d_{n-z-1}X^{n-z-1}\) by Lemma [\[A=B\^zC\]](#A=B^zC){reference-type="ref" reference="A=B^zC"}. In the proof we consider two possibilities for \(C_0\), whether \(C_0\) is a non-zero or a zero matrix.
## Case 1: \(C_0\) is non-zero.
Conjugating matrices \(A\), \(B\) and \(C\) in the Weyr form by \(G=\left[ \begin{array}{ccc}g\\&\ddots\\&&g \end{array} \right]\in GL_{2n}(F)\) for any \(g\in GL_2(F)\) changes neither the matrix \(B\) nor the partitions corresponding to nilpotent orbits in which the matrices lie. On the other hand, each diagonal block of \(G^{-1}CG\) is equal to \(g^{-1}C_0g\). Since \(C\) is nilpotent, \(C_0\) is nilpotent as well . Therefore \(C_0\) is similar to the nilpotent \(2\times 2\) Jordan block over any field \(F\), and hence we can assume that \(C_0=\left[ \begin{array}{cc}0&1\\0&0 \end{array} \right]\). In polynomial notation this means that \(a(X),c(X),\) and \(d(X)\) are divisible by \(X\) and the constant term of \(b(X)\) is 1. Choose now \(G=\left[ \begin{array}{cc}b(X)&0\\-a(X)&1 \end{array} \right]\), which is invertible in \(M_2(S)\), since the constant term of \(b(X)\) is non-zero. Then \[G^{-1}AG=\left[ \begin{array}{cc}0&X^z\\X^z\left(b(X)c(X)-a(X)d(X)\right)&X^z\left(a(X)+d(X)\right) \end{array} \right],\] so without any loss of generality we can assume that \(a(X)=0\) and \(b(X)=1\). Note that the assumption that \(c(X)\) and \(d(X)\) are both divisible by \(X\) does not change after the above conjugation, and that after conjugation we still assume that the degrees of the polynomials \(c(X)\) and \(d(X)\) are bounded by \(n-z-1\). Let \(l,m\le n-z\) be the largest positive integers such that \(X^l\) divides \(c(X)\) and \(X^m\) divides \(d(X)\) and let \(p(X),q(X)\in R\) be unique polynomials satisfying \(c(X)=X^lp(X)\) and \(d(X)=X^mq(X)\). Polynomials \(p(X)\) and \(q(X)\) have non-zero constant terms if \(l,m<n-z\), while for \(l=n-z\) (respectively, \(m=n-z\)) the corresponding polynomial \(p(X)\) (respectively, \(q(X)\)) is zero. We first consider the special case when \(p(X)\) is zero. If \(p(X)=q(X)=0\), then \(A^k=0\) for each \(k\ge 2\) and \(\mathrm{rank}\, A=n-z\), so \(A\) corresponds to the partition \((2^{n-z},1^{2z})=[2n]^{n+z}\). This shows that Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1) is possible whenever we decompose \(2n\) into an almost rectangular partition with at least \(n\) parts. As noted in the introduction and proved in, the corresponding nilpotent orbit intersects every nilpotent centralizer. If \(p(X)=0\) and \(q(X)\) is non-zero, then an easy induction shows that for each \(k\ge 1\) we have \[A^k=\left[ \begin{array}{cc}0&X^{kz+(k-1)m}q(X)^{k-1}\\0&X^{kz+km}q(X)^k \end{array} \right],\] which has rank \(n-kz-(k-1)m\) whenever \(k\le \frac{n+m}{z+m}\), and 0 otherwise. Let \(\alpha =\lceil \frac{n+m}{z+m}\rceil\) and \(\beta =\alpha (z+m)-(n+m)\) (so that \(0\le \beta <z+m\)). Then \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}= \begin{cases} n+z&k=0,\\ m+z&1\le k\le \alpha-2,\\ m+z-\beta&k=\alpha-1,\\ 0&k\ge \alpha, \end{cases}\] which implies that \(A\) corresponds to the partition \((\alpha ^{m+z-\beta},(\alpha-1)^{\beta},1^{n-m})=([n+m]^{m+z},1^{n-m})\). These partitions belong to Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P2, \(l=n-z\)), and it follows from the above proof that for each of these partitions \({\bf{\lambda}}\) there exists \(A\in \mathcal{N}(B)\cap {\mathcal O}_{{\bf{\lambda}}}\). In the rest of Case 1 we assume that \(p(X)\) is non-zero, or equivalently that \(l<n-z\). We determine the partition corresponding to \(A\) using the same argument as above. First we compute the powers of \(A\). By Lemma [\[A\^k\]](#A^k){reference-type="ref" reference="A^k"}, we have \[A^k=\left[ \begin{array}{cc}X^{kz}c(X)s_{k-1}(X)&X^{kz}s_k(X)\\X^{kz}c(X)s_k(X)&X^{kz}s_{k+1}(X) \end{array} \right]\] for each positive integer \(k\), where \(s_0(X)=0\), \(s_1(X)=1\), and \[\label{eq:s_k-recursion} s_{k+1}(X)=d(X)s_k(X)+c(X)s_{k-1}(X)\quad \mathrm{for}\, k\ge 1.\] Moreover, \[\begin{aligned}
s_{2k-1}(X)&=&\sum_{i=1}^k{2k-i-1\choose i-1}X^{2km+i(l-2m)-l}p(X)^{i-1}q(X)^{2k-2i}\quad \mathrm{and}\label{eq:s_k-odd}\\ s_{2k}(X)&=&\sum_{i=0}^{k-1}{2k-i-1\choose i}X^{2km+i(l-2m)-m}p(X)^iq(X)^{2k-2i-1}\quad \mathrm{for}\, k\ge 1\label{eq:s_k-even}. \end{aligned}\] Now we consider two subcases.
### **Case 1.1: \(l<2m\)**
In this case, [\[eq:s_k-odd\]](#eq:s_k-odd){reference-type="eqref" reference="eq:s_k-odd"} and [\[eq:s_k-even\]](#eq:s_k-even){reference-type="eqref" reference="eq:s_k-even"} imply that the highest powers of \(X\) dividing \(s_{2k-1}(X)\) and \(s_{2k}(X)\) are \(X^{kl-l}\) and \(X^{kl-l+m}\), respectively, unless \(q(X)=0\) when \(s_{2k}(X)=0\) for each \(k\ge 1\). For each \(k\ge 1\), we can write \(s_{2k-1}(X)=X^{kl-l}r_{2k-1}(X)\) and \(s_{2k}(X)=X^{kl-l+m}r_{2k}(X)\) for some polynomials \(r_{2k-1}(X),r_{2k}(X)\in R\), where \(r_{2k-1}(X)\) is uniquely determined and has non-zero constant term, while \(r_{2k}(X)\) is such if \(q(X)\ne 0\). Lemma [\[A\^k\]](#A^k){reference-type="ref" reference="A^k"} then implies that the equalities \[\begin{aligned}
r_{2k-1}(X)r_{2k+1}(X)-X^{2m-l}r_{2k}(X)^2&=&p(X)^{2k-1},\label{eq:1.1det-odd}\\ X^{2m-l}r_{2k}(X)r_{2k+2}(X)-r_{2k+1}(X)^2&=&-p(X)^{2k}\label{eq:1.1det-even} \end{aligned}\] hold in \(R\) for each \(k\ge 1\). Furthermore, the last equality holds also for \(k=0\) if we define \(r_0(X)=0\). Now we first consider the odd powers of \(A\): \[A^{2k'-1}=\left[ \begin{array}{cc} X^{(2k'-1)z+k'l-l+m}p(X)r_{2k'-2}(X)&X^{(2k'-1)z+k'l-l}r_{2k'-1}(X)\\X^{(2k'-1)z+k'l}p(X)r_{2k'-1}(X)&X^{(2k'-1)z+k'l-l+m}r_{2k'}(X) \end{array} \right].\] By definition, the polynomial \(r_{2k'-1}(X)\) has non-zero constant term, so it is invertible in the quotient ring \(S=R/(X^n)\). Let \(r_{2k'-1}(X)^{-1}\) denote its inverse in \(S\). Applying Gaussian elimination to \(A^{2k'-1}\) and using [\[eq:1.1det-even\]](#eq:1.1det-even){reference-type="eqref" reference="eq:1.1det-even"} we now get \[\begin{aligned}
&&\left[ \begin{array}{cc}r_{2k'-1}(X)^{-1}&0\\-X^mr_{2k'-1}(X)^{-1}r_{2k'}(X)&1 \end{array} \right]\cdot A^{2k'-1}\cdot \left[ \begin{array}{cc}r_{2k'-1}(X)&0\\-X^mp(X)r_{2k'-2}(X)&1 \end{array} \right]=\\ &=&\left[ \begin{array}{cc} 0&X^{(2k'-1)z+k'l-l}\\ X^{(2k'-1)z+k'l}p(X)\Big(r_{2k'-1}(X)^2-X^{2m-l}r_{2k'-2}(X)r_{2k'}(X)\Big)&0 \end{array} \right] = \\ &=&\left[ \begin{array}{cc} 0&X^{(2k'-1)z+k'l-l}\\ X^{(2k'-1)z+k'l}p(X)^{2k'-1}&0 \end{array} \right]. \end{aligned}\] Matrices \(\left[ \begin{array}{cc}r_{2k'-1}(X)^{-1}&0\\-X^mr_{2k'-1}(X)^{-1}r_{2k'}(X)&1 \end{array} \right]\) and \(\left[ \begin{array}{cc}r_{2k'-1}(X)&0\\-X^mp(X)r_{2k'-2}(X)&1 \end{array} \right]\) are invertible in \(M_2(S)\cong \mathcal{C}(B)\), therefore, \[\begin{aligned}
\mathrm{rank}\, A^{2k'-1}&=&\mathrm{rank} \left[ \begin{array}{cc}0&X^{(2k'-1)z+k'l-l}\\X^{(2k'-1)z+k'l}p(X)^{2k'-1}&0 \end{array} \right]\notag\\ &=&\begin{cases} 2n-(2k'-1)(2z+l) &1\le k'\le \frac{n+z}{l+2z},\\ n-(2k'-1)z-k'l+l &\frac{n+z}{l+2z}\le k'\le \frac{n+l+z}{l+2z},\\ 0 &\frac{n+l+z}{l+2z}\le k'. \end{cases}\label{eq:1.1rank-odd} \end{aligned}\] For the further use we denote \[\alpha :=\left\lceil \frac{n+z}{l+2z}\right\rceil \text{ and } \beta :=\alpha (l+2z)-(n+z),\] and in this case \(0\le \beta < l+2z\) and \(\alpha >1\). Moreover, \(\frac{n+z}{l+2z}\le \frac{n+l+z}{l+2z}\le \frac{n+z}{l+2z}+1\), so \[\left\lceil \frac{n+l+z}{l+2z}\right\rceil \in \{\alpha,\alpha+1\}.\] Now we consider the even powers of \(A\): \[A^{2k'}=\left[ \begin{array}{cc} X^{2k'z+k'l}p(X)r_{2k'-1}(X)&X^{2k'z+k'l-l+m}r_{2k'}(X)\\X^{2k'z+k'l+m}p(X)r_{2k'}(X)&X^{2k'z+k'l}r_{2k'+1}(X) \end{array} \right]\] for \(k'\ge 1\). To compute the rank of such a matrix we have to distinguish between two cases. **Case 1.1(a): \(m\le l\).** In this case \(q(X)\) is non-zero as well. With the same argument as above we see that \(r_{2k'}(X)\) is invertible in \(S\) for \(k'\ge 1\), and let \(r_{2k'}(X)^{-1}\) be its inverse. Using [\[eq:1.1det-odd\]](#eq:1.1det-odd){reference-type="eqref" reference="eq:1.1det-odd"}, for each \(k'\ge 1\) we compute \[\begin{aligned}
&&\left[ \begin{array}{cc}r_{2k'}(X)^{-1}&0\\-X^{l-m}r_{2k'}(X)^{-1}r_{2k'+1}(X)&1 \end{array} \right]\cdot A^{2k'}\cdot \left[ \begin{array}{cc}r_{2k'}(X)&0\\-X^{l-m}p(X)r_{2k'-1}(X)&1 \end{array} \right]=\\ &=&\left[ \begin{array}{cc} 0&X^{2k'z+k'l-l+m}\\X^{2k'z+k'l+l-m}p(X)\Big(X^{2m-l}r_{2k'}(X)^2-r_{2k'-1}(X)r_{2k'+1}(X)\Big)&0 \end{array} \right]=\\ &=&\left[ \begin{array}{cc}0&X^{2k'z+k'l-l+m}\\-X^{2k'z+k'l+l-m}p(X)^{2k'}&0 \end{array} \right], \end{aligned}\] therefore \[\begin{aligned}
\mathrm{rank}\, A^{2k'}&=&\mathrm{rank}\, \left[ \begin{array}{cc}0&X^{2k'z+k'l-l+m}\\-X^{2k'z+k'l+l-m}p(X)^{2k'}&0 \end{array} \right]\notag \\ &=&\begin{cases} 2n-2k'(2z+l) &k'\le \frac{n+m-l}{l+2z},\\ n-k'(2z+l)+l-m &\frac{n+m-l}{l+2z}\le k'\le \frac{n+l-m}{l+2z},\\ 0 &\frac{n+l-m}{l+2z}\le k'. \end{cases}\label{eq:1.1rank-even} \end{aligned}\] Moreover, it is obvious that equality [\[eq:1.1rank-even\]](#eq:1.1rank-even){reference-type="eqref" reference="eq:1.1rank-even"} holds for \(k'=0\), too. We now compare fractions in [\[eq:1.1rank-even\]](#eq:1.1rank-even){reference-type="eqref" reference="eq:1.1rank-even"} with \(\alpha\). Since \(\frac{n+z}{l+2z}-1\le \frac{n+m-l}{l+2z}\le \frac{n+z}{l+2z}\), we get \[\left\lceil \frac{n+m-l}{l+2z}\right\rceil\in\{\alpha-1,\alpha \}.\] Similarly, \(\frac{n+m-l}{l+2z}\le \frac{n+l-m}{l+2z}\le \frac{n+l+z}{l+2z}\), so \[\left\lceil \frac{n+l-m}{l+2z}\right\rceil\in \{\alpha-1,\alpha, \alpha +1\}.\] We consider various cases.
1. Assume first that \(\lceil \frac{n+l-m}{l+2z}\rceil=\alpha +1\), or equivalently \(\beta <l-m-z\), and hence in particular \(l>m+z\). Then it is clear that \(\lceil \frac{n+l+z}{l+2z}\rceil =\alpha+1\) as well. Furthermore, since \(l<2m\), we have \(\frac{n+l-m}{l+2z}-1\le \frac{n+m-l}{l+2z}\), so \(\lceil \frac{n+m-l}{l+2z}\rceil =\alpha\). It follows from [\[eq:1.1rank-odd\]](#eq:1.1rank-odd){reference-type="eqref" reference="eq:1.1rank-odd"} and [\[eq:1.1rank-even\]](#eq:1.1rank-even){reference-type="eqref" reference="eq:1.1rank-even"} that \[\mathrm{rank}\, A^{2k'}=\begin{cases} 2n-2k'(l+2z)&k'\le \alpha-1,\\ l-m-z-\beta&k'=\alpha,\\0&k'\ge \alpha+1 \end{cases}\] and \[\mathrm{rank}\, A^{2k'-1}= \begin{cases} 2n-( 2k'-1)(l+2z )& k' \leq \alpha-1,\\ l-\beta & k'=\alpha, \\ 0,&k'\geq \alpha +1. \end{cases}\] In particular, \(A^{2\alpha+1}=0\) and \(\mathrm{rank}\, A^{2\alpha}=l-m-z-\beta\) which is strictly positive by assumption, so \(2\alpha +1\) is the nilpotency index of \(A\). Moreover, \[\mathrm{rank}A^k-\mathrm{rank}\, A^{k+1}= \begin{cases} l+2z&k\le 2\alpha-3,\\ l+2z-\beta&k=2\alpha-2,\\ m+z&k=2\alpha-1,\\ l-m-z-\beta&k=2\alpha,\\ 0&k\ge 2\alpha+1, \end{cases}\] which implies that \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P3) \[\begin{aligned}
&&((2\alpha +1)^{l-m-z-\beta},(2\alpha)^{2m-l+2z+\beta},(2\alpha-1)^{l-m+z-\beta},(2\alpha-2)^{\beta})\\ &=&([n+l-m+\alpha (2m-l)]^{m+z},[n+m-l-\alpha(2m-l)]^{l-m+z}), \end{aligned}\] where \(0\le z\le n-1\), \(0<m\le l<n-z\), \(l<2m\), \(\alpha >1\) and \(0\le \beta <l-m-z\), where \(0<m\le l\) is redundant. We note that \(n+m-l-\alpha(2m-l)\) is indeed positive for \(\alpha \ge 1\), as it is equal to \(\alpha (2l-2m+2z)+m-l-z-\beta \ge l-m+z-\beta>0\).
2. Assume that \(\lceil \frac{n+l-m}{l+2z}\rceil=\alpha\), or equivalently \(l-m-z\le \beta <2l-m+z\). We know that \(\lceil \frac{n+m-l}{l+2z}\rceil \in\{\alpha-1,\alpha\}\) and \(\lceil \frac{n+l+z}{l+2z}\rceil\in \{\alpha ,\alpha +1\}\). We consider all options.
- Assume \(\lceil \frac{n+m-l}{l+2z}\rceil =\alpha\) and \(\lceil \frac{n+l+z}{l+2z}\rceil=\alpha +1\), which is equivalent to \(m+z>\beta\) and \(l>\beta\). In this case \[\mathrm{rank}\, A^{2k'}=\begin{cases} 2n-2k'(l+2z)&k'\le \alpha-1,\\0&k'\ge \alpha \end{cases}\] and \[\mathrm{rank}\, A^{2k'-1}=\begin{cases} 2n-( 2k'-1)(l+2z ) & k' \leq \alpha-1,\\ l-\beta & k'=\alpha, \\ 0&k'\ge \alpha+1. \end{cases}\] We therefore get \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\begin{cases} l+2z&k\le 2\alpha-3,\\l+2z-\beta&k=2\alpha-2,\\l-\beta&k=2\alpha-1,\\0&k\ge 2\alpha, \end{cases}\] so \(A \in {\mathcal O}_{{\bf{\lambda}}}\), where \[{\bf{\lambda}} = ((2\alpha)^{l-\beta},(2\alpha-1)^{2z},(2\alpha-2)^{\beta}),\] corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P6), and \(0\le z\le n-1\), \(0<m\le l<n-z\), \(l<2m\) and \(\max\{l-m-z,0\}\le \beta <\min\{m+z,l\}\). (Again, the condition \(\beta <2l-m+z\) is redundant.)
- Assume \(\lceil \frac{n+m-l}{l+2z}\rceil =\lceil \frac{n+l+z}{l+2z}\rceil=\alpha\), which is equivalent to \(l\le \beta <m+z\). Then \[\mathrm{rank}\, A^{2k'}= \begin{cases} 2n-2k'(l+2z)&k'\le \alpha-1,\\ 0&k'\ge \alpha, \end{cases}\] \[\mathrm{rank}\, A^{2k'-1}=\begin{cases} 2n-( 2k'-1)(l+2z ) & k' \leq \alpha-1,\\ 0&k'\geq \alpha, \end{cases}\] and \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\begin{cases} l+2z&k\le 2\alpha-3,\\ 2l+2z-2\beta&k=2\alpha-2,\\0&k\ge 2\alpha-1. \end{cases}\] Consequently, \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1) \[\left( (2\alpha-1)^{2l+2z-2\beta},(2\alpha-2)^{2\beta-l}\right)=[2n]^{l+2z},\] where \(0\le z\le n-1\), \(0< m\le l<n-z\), \(l<2m\) and \(l\le \beta <m+z\) (while the conditions \(l-m-z\le \beta <2l-m+z\) and \(\beta<l+2z\) are redundant).
- Assume \(\lceil \frac{n+m-l}{l+2z}\rceil =\alpha-1\) and \(\lceil \frac{n+l+z}{l+2z}\rceil =\alpha +1\), which is equivalent to \(m+z\le \beta <l\). It follows that \[\mathrm{rank}\, A^{2k'}=\begin{cases} 2n-2k'(l+2z)&k'\le \alpha-2,\\ 2l-m+z-\beta &k'=\alpha-1,\\0&k'\ge \alpha, \end{cases}\] \[\mathrm{rank}\, A^{2k'-1}=\begin{cases} 2n-(2k'-1)(l+2z)&k'\le \alpha-1,\\ l-\beta&k'=\alpha,\\ 0&k'\ge \alpha+1 \end{cases}\] and \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\begin{cases} l+2z&k\le 2\alpha-4,\\ l+m+3z-\beta&k=2\alpha-3,\\ l-m+z&k=2\alpha-2,\\ l-\beta&k=2\alpha-1,\\ 0&k\ge 2\alpha. \end{cases}\] Consequently, \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P4) \[\begin{aligned}
&&((2\alpha)^{l-\beta},(2\alpha-1)^{z-m+\beta},(2\alpha-2)^{2m+2z-\beta},(2\alpha-3)^{\beta-m-z})\\ &=&([n+m-\alpha (2m-l)]^{l-m+z},[n-m+\alpha(2m-l)]^{m+z}), \end{aligned}\] where \(0\le z\le n-1\), \(0<m\le l<n-z\), \(l<2m\) and \(m+z\le \beta <l\) (while the conditions \(l-m-z\le \beta <2l-m+z\) and also \(m\le l\) are redundant).
- It remains to consider the case when \(\lceil \frac{n+m-l}{l+2z}\rceil =\alpha-1\) and \(\lceil \frac{n+l+z}{l+2z}\rceil =\alpha\), which is equivalent to \(\beta \ge \max\{m+z,l\}\). Now we have \[\mathrm{rank}\, A^{2k'}=\begin{cases} 2n-2k'(l+2z)&k'\le \alpha-2,\\ 2l-m+z-\beta &k'=\alpha-1,\\ 0&k'\ge \alpha, \end{cases}\] \[\mathrm{rank}\, A^{2k'-1}=\begin{cases} 2n-( 2k'-1)(l+2z ) & k' \leq \alpha-1,\\ 0&k'\geq \alpha, \end{cases}\] and \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\begin{cases} l+2z&k\le 2\alpha-4,\\ l+m+3z-\beta&k=2\alpha-3, \\2l-m+z-\beta&k=2\alpha-2, \\0&k\ge 2\alpha-1. \end{cases}\] Hence, \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P7) \[\left( (2\alpha-1)^{2l-m+z-\beta},(2\alpha-2)^{2m-l+2z},(2\alpha-3)^{\beta-m-z}\right),\] where \(0\le z\le n-1\), \(0<m\le l<n-z\), \(l<2m\) and \(\max\{m+z,l\}\le \beta <\min\{2l-m+z,l+2z\}\) (and the condition \(\beta \ge l-m-z\) is redundant).
3. Assume now that \(\lceil \frac{n+l-m}{l+2z}\rceil=\alpha-1\), or equivalently \(\beta \ge 2l-m+z\), so in particular we have \(z>l-m\), since \(\beta <l+2z\). Since \(l\ge m\), we get also \(\lceil \frac{n+m-l}{l+2z}\rceil=\alpha-1\). Moreover, the conditions \(m\le l\) and \(z\ge 0\) imply \(2l-m+z\ge l\), so \(\beta \ge l\), which is equivalent to \(\lceil \frac{n+l+z}{l+2z}\rceil =\alpha\). We thus get \[\mathrm{rank}\, A^{2k'}=\begin{cases} 2n-2k'(l+2z)&k'\le \alpha-2,\\ 0&k'\ge \alpha-1, \end{cases}\] \[\mathrm{rank}\, A^{2k'-1}= \begin{cases} 2n-(2k'-1)(l+2z)&k'\le \alpha-1,\\ 0&k'\ge \alpha, \end{cases}\] and therefore \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}= \begin{cases} l+2z&k\le 2\alpha-4,\\ 2n-(2\alpha-3)(l+2z)&k=2\alpha-3,\\ 0&k\ge2\alpha-2. \end{cases}\] It follows that \(A \in {\mathcal O}_{{\bf{\lambda}}}\), where \[{\bf{\lambda}}=((2\alpha-2)^{2n-(2\alpha-3)(l+2z)},(2\alpha-3)^{(2\alpha-2)(l+2z)-2n})=[2n]^{l+2z}\] as in Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1), for \(0\le z\le n-1\), \(0<m\le l<n-z\), \(l<2m\) and \(2l-m+z\le \beta <l+2z\).
At the end of Case 1.1(a) we observe that it immediately follows from the above proof that for each of the partitions \({\bf{\lambda}}\) obtained in (i)-(iii) there exists \(A\in \mathcal{N}(B)\cap {\mathcal O}_{{\bf{\lambda}}}\). Note that in the special case when \(l=1\), the condition (a) forces \(m=1\). Moreover, the option (i) above and the last two cases in (ii) are not possible, and in all other cases we get the partition \([2n]^{2z+1}\). Hence, in particular we get all odd cases in the first set of partitions listed in the theorem, except \([2n]^{2n-1}\), since \(1<n-z\). However, the existence of this partition was shown at the beginning of the proof in the case when \(p(X)=q(X)=0\). The existence of all even cases will be proved in Case 2 below, but we note here that we get almost all of them also from the partitions above in the case \(l=2\). Indeed, in this case the condition \(m\le l<2m\) implies \(m=2\), again the option (i) and the last two cases in (ii) are not possible, and the only case when we get a partition different from \([2n]^{2z+2}\) is when \(\beta=1\). Note also that for \(n\ge 2\) the partition \([2n]^{2n-2}\) is not covered by the above cases if \(l=2\), but it has been established when we considered the case \(p(X)=q(X)=0\). **Case 1.1(b): \(n-z\ge m>l\)**. We proceed as in Case 1.1(a). As above, the polynomial \(r_{2k'+1}(X)\) has non-zero constant term, therefore it is invertible in \(S=R/(X^n)\) for each \(k'\ge 0\). Let \(r_{2k'+1}(X)^{-1}\) be its inverse in \(S\). For \(k'\ge 1\) we compute, using [\[eq:1.1det-odd\]](#eq:1.1det-odd){reference-type="eqref" reference="eq:1.1det-odd"}, that \[\begin{aligned}
&&\left[ \begin{array}{cc}1&-X^{m-l}r_{2k'+1}(X)^{-1}r_{2k'}(X)\\0&r_{2k'+1}(X)^{-1} \end{array} \right]\cdot A^{2k'}\cdot \left[ \begin{array}{cc}r_{2k'+1}(X)&0\\X^mp(X)r_{2k'}(X)&1 \end{array} \right]=\\ &=&\left[ \begin{array}{cc} X^{2k'z+k'l}p(X)\Big(r_{2k'-1}(X)r_{2k'+1}(X)-X^{2m-l}r_{2k'}(X)^2\Big)&0\\0&X^{2k'z+k'l} \end{array} \right]=\\ &=&\left[ \begin{array}{cc}X^{2k'z+k'l}p(X)^{2k'}&0\\0&X^{2k'z+k'l} \end{array} \right], \end{aligned}\] so \[\begin{aligned}
\mathrm{rank}\, A^{2k'}&=&\mathrm{rank}\,\left[ \begin{array}{cc}X^{2k'z+k'l}p(X)^{2k'}&0\\0&X^{2k'z+k'l} \end{array} \right]=\notag \\ &=&\left\{ \begin{array}{ll} 2n-2k'(l+2z)&k'\le \frac{n}{l+2z},\\ 0&k'\ge \frac{n}{l+2z}, \end{array} \right.\label{eq:1.1(b)rank-even} \end{aligned}\] and clearly \(\mathrm{rank}\, A^0=2n\), so [\[eq:1.1(b)rank-even\]](#eq:1.1(b)rank-even){reference-type="eqref" reference="eq:1.1(b)rank-even"} holds for \(k'=0\). Since \(\frac{n+z}{l+2z}-1\le \frac{n}{l+2z}\le \frac{n+z}{l+2z}\), we get \[\left\lceil \frac{n}{l+2z}\right\rceil\in \{\alpha-1,\alpha\}.\] On the other hand, we know that \(\lceil \frac{n+l+z}{l+2z}\rceil \in \{\alpha ,\alpha +1\}\). Moreover, \(\frac{n+l+z}{l+2z}-\frac{n}{l+2z}\le 1\), therefore \(\lceil \frac{n+l+z}{l+2z}\rceil-\lceil \frac{n}{l+2z}\rceil \in \{0,1\}\). We consider all options.
1. Assume first that \(\lceil \frac{n+l+z}{l+2z}\rceil =\alpha +1\), which is equivalent to \(\beta <l\). Then it is clear that \(\lceil \frac{n}{l+2z}\rceil =\alpha\). Furthermore, we get from [\[eq:1.1rank-odd\]](#eq:1.1rank-odd){reference-type="eqref" reference="eq:1.1rank-odd"} and [\[eq:1.1(b)rank-even\]](#eq:1.1(b)rank-even){reference-type="eqref" reference="eq:1.1(b)rank-even"} that \[\mathrm{rank}\, A^{2k'}=\left\{ \begin{array}{ll}2n-2k'(l+2z)&k'\le \alpha-1,\\0&k'\ge \alpha, \end{array} \right.\] \[\mathrm{rank}\, A^{2k'-1}=\left\{ \begin{array}{ll}2n-(2k'-1)(l+2z)&k'\le \alpha-1,\\ l-\beta&k'=\alpha,\\ 0&k'\ge \alpha+1, \end{array}\right.\] and \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}= \left\{ \begin{array}{ll}l+2z&k\le 2\alpha-3,\\ l+2z-\beta&k=2\alpha-2,\\ l-\beta&k=2\alpha-1,\\ 0&k\ge 2\alpha, \end{array} \right.\] and the corresponding partition is Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P6) \[\left( (2\alpha)^{l-\beta},(2\alpha-1)^{2z},(2\alpha-2)^{\beta}\right),\] for \(0\le z\le n-1\), \(0< l<m\le n-z\) and \(0\le \beta <l\).
2. Assume now that \(\lceil \frac{n+l+z}{l+2z}\rceil =\lceil \frac{n}{l+2z}\rceil =\alpha\), which is equivalent to \(l\le \beta <l+z\). It follows from [\[eq:1.1rank-odd\]](#eq:1.1rank-odd){reference-type="eqref" reference="eq:1.1rank-odd"} and [\[eq:1.1(b)rank-even\]](#eq:1.1(b)rank-even){reference-type="eqref" reference="eq:1.1(b)rank-even"} that \[\mathrm{rank}\, A^{2k'}=\left\{ \begin{array}{ll}2n-2k'(l+2z)&k'\le \alpha-1,\\0&k'\ge \alpha, \end{array} \right.\] \[\mathrm{rank}\, A^{2k'-1}=\left\{ \begin{array}{ll}2n-(2k'-1)(l+2z)&k'\le \alpha-1,\\0&k'\ge \alpha, \end{array}\right.\] and \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}= \left\{ \begin{array}{ll}l+2z&k\le 2\alpha-3,\\2n-2(\alpha-1)(l+2z)&k=2\alpha-2,\\0&k\ge 2\alpha-1, \end{array} \right.\] so the corresponding partition is Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1) \[\left( (2\alpha-1)^{2n-2(\alpha-1)(l+2z)}, (2\alpha-2)^{-2n+(2\alpha-1)(l+2z)}\right)=[2n]^{l+2z},\] where \(0\le z\le n-1\), \(0<l<m\le n-z\) and \(l\le \beta <l+z\).
3. It remains to consider the case when \(\lceil \frac{n}{l+2z}\rceil =\alpha-1\) (and hence \(\lceil \frac{n+l+z}{l+2z}\rceil =\alpha\)), which is equivalent to \(\beta \ge l+z\). In this case we have \[\mathrm{rank}\, A^{2k'}=\left\{ \begin{array}{ll}2n-2k'(l+2z)&k'\le \alpha-2,\\0&k'\ge \alpha-1, \end{array} \right.\] \[\mathrm{rank}\, A^{2k'-1}=\left\{ \begin{array}{ll}2n-(2k'-1)(l+2z)&k'\le \alpha-1,\\0&k'\ge \alpha, \end{array}\right.\] and \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}= \left\{ \begin{array}{ll}l+2z&k\le 2\alpha-4,\\2n-(2\alpha-3)(l+2z)&k=2\alpha-3,\\0&k\ge 2\alpha-2, \end{array} \right.\] and the corresponding partition is Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1) \[\left( (2\alpha-2)^{2n-(2\alpha-3)(l+2z)}, (2\alpha-3)^{-2n+(2\alpha-2)(l+2z)}\right)=[2n]^{l+2z},\] where \(0\le z\le n-1\), \(0<l<m\le n-z\) and \(\beta \ge l+z\).
As before, it follows from the above proof that for each of the partitions obtained in (i)-(iii) there indeed exists \(A\in \mathcal{N}(B)\) that corresponds to the partition.
### **Case 1.2: \(l\ge 2m\)**
Then \(q(X)\) has non-zero constant term. Moreover, it follows from [\[eq:s_k-odd\]](#eq:s_k-odd){reference-type="eqref" reference="eq:s_k-odd"} and [\[eq:s_k-even\]](#eq:s_k-even){reference-type="eqref" reference="eq:s_k-even"} that, for each \(k\ge 1\), the polynomial \(s_k(X)\) is divisible by \(X^{km-m}\). For each \(k\ge 1\), we write \(s_k(X)=X^{km-m}r_k(X)\) for some uniquely determined polynomial \(r_k(X)\in R\), and let \(r_0(X)=0\). Then \[\label{eq:r_k-recursion} r_{k+1}(X)=q(X)r_k(X)+X^{l-2m}p(X)r_{k-1}(X)\] for \(k\ge 1\) by [\[eq:s_k-recursion\]](#eq:s_k-recursion){reference-type="eqref" reference="eq:s_k-recursion"}. Furthermore, Lemma [\[A\^k\]](#A^k){reference-type="ref" reference="A^k"} implies that \[A^k=\left[ \begin{array}{cc} X^{kz+km-2m+l}p(X)r_{k-1}(X)&X^{kz+km-m}r_k(X)\\X^{kz+km-m+l}p(X)r_k(X)&X^{kz+km}r_{k+1}(X) \end{array} \right]\] for \(k\ge 1\), and \[\label{eq:1.2det} r_{k-1}(X)r_{k+1}(X)-r_k(X)^2=(-1)^kX^{(k-1)(l-2m)}p(X)^{k-1}\] holds in \(R\) for each \(k\ge 2\). Clearly, [\[eq:1.2det\]](#eq:1.2det){reference-type="eqref" reference="eq:1.2det"} holds also for \(k=1\). If \(l>2m\), then it is also clear from [\[eq:s_k-odd\]](#eq:s_k-odd){reference-type="eqref" reference="eq:s_k-odd"} and [\[eq:s_k-even\]](#eq:s_k-even){reference-type="eqref" reference="eq:s_k-even"} that \(r_k(X)\) has non-zero constant term for each \(k\ge 1\). As we will see below, it is convenient to define the numbers \[\alpha =\left\lceil\frac{n+m}{z+m}\right\rceil, \beta =\alpha (z+m)-(n+m), \gamma =\left\lceil \frac{n-m}{z+l-m}\right\rceil, \mathrm{ and }\, \delta =\gamma(z+l-m)-(n-m).\] Then \(0\le \beta <z+m\) and \(0\le \delta < z+l-m\). Moreover, since \(l\ge 2m\), we get \(z+l-m\ge z+m\), hence \(\frac{n+m}{z+m}\ge \frac{n-m}{z+l-m}\), and consequently \(\alpha\geq \gamma.\) Furthermore, \(l<n-z\) implies \(\gamma >1\). Suppose first that \(k\ge 1\) is such that \(r_k(X)\) is not divisible by \(X\). (As noted above, this happens always if \(l>2m\).) Then it is invertible in \(S=R/(X^n)\), and let \(r_k(X)^{-1}\) be its inverse in \(S\). Using [\[eq:1.2det\]](#eq:1.2det){reference-type="eqref" reference="eq:1.2det"} we now compute \[\begin{aligned}
&&\left[ \begin{array}{cc} r_k(X)^{-1}&0\\-X^mr_k(X)^{-1}r_{k+1}(X)&1 \end{array} \right]\cdot A^k\cdot \left[ \begin{array}{cc}r_k(X)&0\\-X^{l-m}p(X)r_{k-1}(X)&1 \end{array} \right]=\\ &=&\left[ \begin{array}{cc} 0&X^{kz+km-m}\\X^{kz+km+l-m}p(X)\Big(r_k(X)^2-r_{k-1}(X)r_{k+1}(X)\Big)&0 \end{array} \right]=\\ &=&\left[ \begin{array}{cc} 0&X^{kz+km-m}\\(-1)^{k+1}X^{kz+kl-km+m}p(X)^k&0 \end{array} \right]. \end{aligned}\] We can therefore compute \[\begin{aligned}
\label{eq:14} \mathrm{rank}\, A^{k}&=&\mathrm{rank}\, \left[ \begin{array}{cc} 0 & X^{kz+km-m}\\ (-1)^{k+1}X^{kz+kl-km+m}p(X)^{k} & 0 \end{array} \right]\notag \\ &=& \begin{cases} 2n-k(l+2z)& k\leq \frac{n-m}{z+l-m},\\ n+m-k(m+z)& \frac{n-m}{z+l-m} \leq k\le \frac{n+m}{z+m}, \\ 0& k\ge \frac{n+m}{z+m}. \end{cases} \end{aligned}\] If \(k\ge 1\) and \(r_k(X)\) is not divisible by \(X\), we therefore get \[\label{eq:1.2rankA^k-not_divisible} \mathrm{rank}\, A^k=\left\{ \begin{array}{ll}2n-k(l+2z)&k\le \gamma-1,\\ n+m-k(m+z)&\gamma\le k\le \alpha-1,\\ 0&k\ge \alpha, \end{array} \right.\] and it is clear that the above holds also for \(k=0\). Suppose now that \(X\) divides \(r_k(X)\) for some \(k\ge 1\). We already know that this forces \(l=2m\), hence \(\gamma =\lceil \frac{n-m}{z+m}\rceil\), and in particular \(\alpha \le \gamma +2\). Moreover, if \(X\) divides \(r_{k-1}(X)\), too, then [\[eq:r_k-recursion\]](#eq:r_k-recursion){reference-type="eqref" reference="eq:r_k-recursion"} implies that \(r_j(X)\) is divisible by \(X\) for any \(j\). (Recall that \(p(X)\) is not divisible by \(X\).) However, this contradicts \(r_1(X)=1\). It follows that no two consecutive polynomials \(r_j(X)\) and \(r_{j-1}(X)\) are divisible by \(X\). Let \(t\) be the largest positive integer smaller than or equal to \(m+1\) such that \(X^t\) divides \(r_k(X)\), and let \(\widetilde{r_k}(X)\in R\) be the unique polynomial satisfying \(r_k(X)=X^t\widetilde{r_k}(X)\). If \(t\le m\), then the polynomial \(\widetilde{r_k}(X)\) has non-zero constant term, so it is invertible in \(S=R/(X^n)\). Let \(\widetilde{r_k}(X)^{-1}\) be its inverse in \(S\). For \(t\le m\), we get by [\[eq:1.2det\]](#eq:1.2det){reference-type="eqref" reference="eq:1.2det"} that \[\begin{aligned}
&&\left[ \begin{array}{cc}\widetilde{r_k}(X)^{-1}&0\\-X^{m-t}\widetilde{r_k}(X)^{-1}r_{k+1}(X)&1 \end{array} \right] \cdot A^k\cdot \left[ \begin{array}{cc}\widetilde{r_k}(X)&0\\-X^{m-t}p(X)r_{k-1}(X)&1 \end{array} \right]=\\ &=&\left[ \begin{array}{cc}0&X^{kz+km-m+t}\\X^{kz+km+m-t}p(X)\Big(r_k(X)^2-r_{k-1}(X)r_{k+1}(X)\Big)&0 \end{array} \right]=\\ &=&\left[ \begin{array}{cc}0&X^{kz+km-m+t}\\(-1)^{k+1}X^{kz+km+m-t}p(X)^k&0 \end{array} \right], \end{aligned}\] so \[\label{eq:1.2rankA^k-divisible} \mathrm{rank}\, A^k= \begin{cases} 2n-2k(m+z)&k\le \frac{n-m+t}{m+z},\\ n+m-t-k(m+z)&\frac{n-m+t}{m+z}\le k\le \frac{n+m-t}{m+z},\\ 0&k\ge \frac{n+m-t}{m+z}. \end{cases}\] On the other hand, for \(t=m+1\) we use the fact that \(r_{k+1}(X)\) has non-zero constant term, which was shown above. Hence \(r_{k+1}(X)\) is invertible in \(S\) and we obtain \[\begin{aligned}
&&\left[ \begin{array}{cc}1&-X\widetilde{r_k}(X)r_{k+1}(X)^{-1}\\0&r_{k+1}(X)^{-1} \end{array} \right]\cdot A^k\cdot \left[ \begin{array}{cc}r_{k+1}(X)&0\\-X^mp(X)r_k(X)&1 \end{array} \right]=\\ &=&\left[ \begin{array}{cc}X^{kz+km}p(X)\Big(r_{k-1}(X)r_{k+1}(X)-r_k(X)^2\Big)&0\\0&X^{kz+km} \end{array} \right]=\\ &=&\left[ \begin{array}{cc}X^{kz+km}p(X)^k&0\\0&X^{kz+km} \end{array}, \right] \end{aligned}\] hence \[\mathrm{rank}\, A^k= \begin{cases} 2n-2k(m+z)&k\le \frac{n}{m+z},\\ 0&k\ge \frac{n}{m+z}, \end{cases}\] which is a special case of [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} when \(t=m\). If \(t=m+1\), we therefore get no additional partitions corresponding to orbits intersecting \(\mathcal{N}(B)\), so we do not consider this case below. Suppose that \(k\le \gamma-1\). If \(r_k(X)\) is not divisible by \(X\), then [\[eq:1.2rankA\^k-not_divisible\]](#eq:1.2rankA^k-not_divisible){reference-type="eqref" reference="eq:1.2rankA^k-not_divisible"} shows that \(\mathrm{rank}\,A^k=2n-k(l+2z)\). On the other hand, if \(X^t\) is the highest power of \(X\) that divides \(r_k(X)\) for some \(1\le t\le m\), then \(l=2m\) and it is clear that \(k\le \frac{k-m+t}{m+z}\), so [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} implies that again we have \(\mathrm{rank}\, A^k=2n-2k(m+z)=2n-k(l+2z)\). Similarly, the rank of \(A^k\) in [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} is zero if \(k\ge \alpha\). It follows that the ranks of \(A^k\) in [\[eq:1.2rankA\^k-not_divisible\]](#eq:1.2rankA^k-not_divisible){reference-type="eqref" reference="eq:1.2rankA^k-not_divisible"} and in [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} are different only if \(\gamma \le k\le \alpha-1\). Now we consider two cases. **Case 1.2(a): \(r_k(X)\) is not divisible by \(X\) for any \(k\in \{\gamma,\ldots,\alpha-1\}\).** As noted above, this case in particular includes the case when \(l>2m\). We consider various cases with respect to \(\alpha\) and \(\gamma\):
1. If \(\alpha \ge \gamma +2\), then [\[eq:1.2rankA\^k-not_divisible\]](#eq:1.2rankA^k-not_divisible){reference-type="eqref" reference="eq:1.2rankA^k-not_divisible"} implies \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll}l+2z&k\le \gamma-2,\\ l+2z-\delta&k=\gamma-1,\\ m+z&\gamma\le k\le\alpha-2,\\ m+z-\beta&k=\alpha-1,\\ 0&k\ge \alpha, \end{array} \right.\] hence the matrix \(A\) belongs to \({\mathcal O}_{\bf{\lambda}}\), where \[{\bf{\lambda}}=(\alpha^{m+z-\beta},(\alpha-1)^{\beta},\gamma^{l-m+z-\delta},(\gamma-1)^{\delta})=([n+m]^{m+z},[n-m]^{l-m+z}),\] as in Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P2), for \(0\le z\le n-1\), \(0<2m\le l<n-z\) and \(\alpha \ge \gamma+2\).
2. If \(\alpha =\gamma +1\), then we again use [\[eq:1.2rankA\^k-not_divisible\]](#eq:1.2rankA^k-not_divisible){reference-type="eqref" reference="eq:1.2rankA^k-not_divisible"} to get \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}= \begin{cases} l+2z&k\le \gamma-2,\\ l+2z-\delta&k=\gamma-1,\\ m+z-\beta&k=\gamma,\\ 0&k\ge \gamma +1. \end{cases}\] It follows that \(A \in {\mathcal O}_{\bf{\lambda}}\), where \[{\bf{\lambda}}=((\gamma +1)^{m+z-\beta},\gamma ^{l-m+z+\beta-\delta},(\gamma-1)^{\delta})=([n+m]^{m+z},[n-m]^{l-m+z}),\] as in Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P2), for \(0\le z\le n-1\), \(0<2m\le l<n-z\) and \(\alpha = \gamma+1\).
3. If \(\alpha =\gamma\), then \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll}l+2z&k\le \gamma-2,\\ 2n-(\gamma-1)(l+2z)&k=\gamma-1,\\ 0&k\ge \gamma, \end{array} \right.\] so \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1) \[(\gamma ^{2n-(\gamma-1)(l+2z)},(\gamma-1)^{\gamma (l+2z)-2n})=[2n]^{l+2z},\] where \(0\le z\le n-1\), \(0<2m\le l<n-z\) and \(\alpha= \gamma\).
Conversely, it follows from the above proof that for each partition obtained in (i)-(iii) there indeed exists \(A\in \mathcal{N}(B)\) that corresponds to this partition. The only case of existence that might not be obvious is when \(l=2m\) and \(r_k(X)\) is not divisible by \(X\) for any \(k\ge 1\). However, in this case we can define \(p(X)=q(X)=1\), and then [\[eq:r_k-recursion\]](#eq:r_k-recursion){reference-type="eqref" reference="eq:r_k-recursion"} implies that \(r_k(X)\) is a strictly positive constant for each \(k\ge 1\), and hence non-zero. **Case 1.2(b): \(l=2m\) and there exists \(j\in \{\gamma,\ldots,\alpha-1\}\) such that \(r_j(X)\) is divisible by \(X\).** Since \(r_2(X)=q(X)\) has non-zero constant term, we immediately get \(j\ge 3\), and consequently \(\alpha \ge 4\). As shown above, we may assume that \(r_j(X)=X^t\widetilde{r_j}(X)\) for some \(t\le m\) and some polynomial \(\widetilde{r_j}(X) \in R\) with non-zero constant term. We have shown above also that \(r_{j-1}(X)\) and \(r_{j+1}(X)\) are not divisible by \(X\). Moreover, since \(l=2m\), we have \(\frac{n-m}{l-m+z}+2=\frac{n+m+2z}{m+z}\ge \frac{n+m}{m+z}\), so \(\gamma +2\ge \alpha\). It follows that \(j\in \{\alpha-2,\alpha-1\}\). We consider each of the two options.
1. Assume first that \(j=\alpha-2\). Then \(\gamma =\alpha-2\) and \(\delta =\beta-2z\), so \(\beta \ge 2z\) and \(\delta <m-z\). (Recall that \(\beta <m+ z\).) Furthermore, from [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} we get \[\mathrm{rank}\, A^{\alpha-2}=\left\{ \begin{array}{ll} 2n-2(\alpha-2)(m+z)& m\ge t>\delta, \\ n+m-t-(\alpha-2)(m+z)& t\le \delta. \end{array} \right.\] Note that the third option in [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} is not possible, since it would imply \(t\ge2m-\delta > m+z\ge m\), which contradicts our assumption on \(t\). If \(t\le \delta\), then [\[eq:1.2rankA\^k-not_divisible\]](#eq:1.2rankA^k-not_divisible){reference-type="eqref" reference="eq:1.2rankA^k-not_divisible"} and [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} imply \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll} 2(m+z)&k\le \alpha-4,\\ 2m+2z+t-\delta&k=\alpha-3,\\ m+z-t&k=\alpha-2,\\ m+z-\beta&k=\alpha-1,\\0&k \ge \alpha, \end{array} \right.\] so \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P8), \[\begin{aligned}
{\bf{\lambda}}&=&\left(\alpha^{m-z-\delta},(\alpha-1)^{\delta+2z-t},(\alpha-2)^{m+z+2t-\delta}, (\alpha-3)^{ \delta-t }\right)=\\ &=&\left([n+m-(\alpha-1)t]^{m+z-t},[n-m+(\alpha-1)t]^{m+z+t}\right), \end{aligned}\] where \(0\le z\le n-1\), \(0<2m<n-z\), \(\alpha\ge 5\), \(\gamma=\alpha-2\) and \(t\le \delta\). On the other hand, if \(m\ge t>\delta\), then \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll}2(m+z)&k\le \alpha-3,\\ m+z-\delta &k=\alpha-2,\\ m+z-\beta&k=\alpha-1,\\ 0&k \ge \alpha, \end{array} \right.\] so \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P8), \((\alpha^{m-z-\delta},(\alpha-1)^{2z},(\alpha-2)^{m+z+\delta})\), which is a special case (for \(t=\delta\)) of the partition considered in the case when \(t\le \delta\).
2. Assume now that \(j=\alpha-1\), which is (because of the assumption \(\gamma \le j\le \alpha-1\)) equal either to \(\gamma\) or to \(\gamma +1\). Using [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"}, we get \[\mathrm{rank}\, A^{\alpha-1}=\left\{ \begin{array}{ll} 2n-2(\alpha-1)(m+z)&t>m-z+\beta, \\ m+z-t-\beta& t\le \min\{m-z+\beta ,m+z-\beta\},\\ 0 & m\ge t>m+z-\beta. \end{array} \right.\] We consider two cases:
- Assume first that \(j=\gamma+1\). Then, as in (i), we get \(\gamma =\alpha-2\) and \(\delta =\beta-2z\ge 0\). In particular, we get \(t\le m\le m-z+\beta\). If \(t\le m+z-\beta\), then we get from [\[eq:1.2rankA\^k-not_divisible\]](#eq:1.2rankA^k-not_divisible){reference-type="eqref" reference="eq:1.2rankA^k-not_divisible"} and [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} that \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll} 2(m+z)&k\le \alpha-4,\\ 2m+2z-\delta&k=\alpha-3,\\ m+z+t & k=\alpha-2,\\ m+z-t-\beta&k=\alpha-1,\\0&k \ge \alpha, \end{array} \right.\] so \(A\in {\mathcal O}_{\bf{\lambda}}\), where \[\begin{aligned}
{\bf{\lambda}}&=&\left(\alpha^{m+z-t-\beta},(\alpha-1)^{ \beta+2t},(\alpha-2)^{m+z-t-\delta},(\alpha-3)^{ \delta}\right)=\\ &=&\left([n+m+\gamma t]^{m+z+t},[n-m-\gamma t]^{m+z-t}\right) \end{aligned}\] as in Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P9) for \(0\le z\le n-1\), \(0<2m<n-z\), \(\gamma=\alpha-2\ge 2\), and \(t\le m+z-\beta\). On the other hand, if \(m\ge t>m+z-\beta\), then \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll} 2(m+z)&k\le \alpha-4,\\ 2m+2z-\delta&k=\alpha-3,\\ 2m-\delta & k=\alpha-2,\\ 0&k \ge \alpha-1, \end{array} \right.\] so \(A\) corresponds to the partition \(\left( (\alpha-1)^{ 2m-\delta },(\alpha-2)^{2z},(\alpha-3)^{\delta }\right)\), which is a special case (for \(t=m+z-\beta\)) of the partition considered above.
- Assume now that \(j=\gamma\), so \(\gamma =\alpha-1\), or, equivalently, \(z-m\le \beta <2z\). More precisely, we have \(\beta =z-m+\delta\). If \(t\le \min\{m-z+\beta,m+z-\beta\}\), then we get from [\[eq:1.2rankA\^k-not_divisible\]](#eq:1.2rankA^k-not_divisible){reference-type="eqref" reference="eq:1.2rankA^k-not_divisible"} and [\[eq:1.2rankA\^k-divisible\]](#eq:1.2rankA^k-divisible){reference-type="eqref" reference="eq:1.2rankA^k-divisible"} that \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll} 2(m+z)&k\le \alpha-3,\\ m+3z+t-\beta&k=\alpha-2,\\ m+z-t-\beta&k=\alpha-1,\\0&k \ge \alpha, \end{array} \right.\] so \(A\in{\mathcal O}_{\bf{\lambda}}\), where \[\begin{aligned}
{\bf{\lambda}}&=&\left(\alpha^{m+z-t-\beta},(\alpha-1)^{ 2z+2t },(\alpha-2)^{m-z-t+\beta}\right)=\\ &=&\left(\alpha^{m+z-t-\beta},(\alpha-1)^{2t+\beta },(\alpha-1)^{2z-\beta},(\alpha-2)^{m-z-t+\beta}\right)=\\ &=&\left([n+m+(\alpha-2)t]^{m+z+t},[n-m-(\alpha-2)t]^{m+z-t}\right) \end{aligned}\] corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P9) for \(0\le z\le n-1\), \(0<2m<n-z\), \(\gamma=\alpha-1\ge 3\) and \(t\le \min\{m+z-\beta,m-z+\beta\}\). If \(m\ge t>m+z-\beta\), then \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll} 2(m+z)&k\le \alpha-3,\\ 2n-2(\alpha-2)(m+z)&k=\alpha-2,\\ 0&k \ge \alpha-1, \end{array} \right.\] so \(A\) corresponds to the partition \[\left((\alpha-1)^{2n-2(\alpha-2)(m+z)},(\alpha-2)^{2(\alpha-1)(m+z)-2n}\right)=[2n]^{2(m+z)},\] which is of the form Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1) and whose existence will be shown in Case 2. Similarly, if \(m\ge t>m-z+\beta\), then \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll} 2(m+z)&k\le \alpha-2,\\ 2n-2(\alpha-1)(m+z)&k=\alpha-1,\\ 0&k \ge \alpha, \end{array} \right.\] so \(A\) corresponds to the partition \((\alpha^{2n-2(\alpha-1)(m+z)},(\alpha-1)^{2\alpha (m+z)-2n})\), which is again of the form \([2n]^{2(m+z)}\).
To finish the proof of Case 1 it remains to show that when \(F=\mathbb{C}\) the partitions obtained in the case 1.2(b) are indeed all possible. First we set the notation. Given polynomials \(p(X),q(X)\in R\), we define the sequence of polynomials \(r_k(X)\in R\) by \[r_0(X)=0,\quad r_1(X)=1,\quad \mathrm{and}\quad r_{k+1}(X)=q(X)r_k(X)+p(X)r_{k-1}(X)\quad \mathrm{for}\quad k\ge 1.\] Note that Lemma [\[A\^k\]](#A^k){reference-type="ref" reference="A^k"} then holds for the polynomials \(p(X)\), \(q(X)\), and \(r_0(X), r_1(X),\ldots\). Now let \(z\le n-1\) be a non-negative integer, let \(m<\frac{n-z}{2}\) be a positive integer, and let \(\alpha =\left\lceil \frac{n+m}{m+z}\right\rceil\) and \(\beta =\alpha(m+z)-(n+m)\). To show that all partitions obtained in the case 1.2(b) are indeed possible it suffices to show that for \(j=\alpha-2\ge 3\) and \(1\le t \le\min\ 2z+\beta,m\ \), and for \(j=\alpha-1\ge 3\) and \(1\le t\le \min\{ m+z-\beta,m-z+\beta\}\) there exist polynomials \(p(X),q(X)\in R\) with non-zero constant terms such that \(X\) does not divide \(r_k(X)\) for any \(k<\alpha\) satisfying \(k\ne j\) and that \(X^t\) is the highest power of \(X\) that divides \(r_j(X)\). We will show that this happens if we define \(q(X)=1\) and \(p(X)=X^t+u\) for a suitable non-zero \(u\in \mathbb{C}\). In this case Lemma [\[A\^k\]](#A^k){reference-type="ref" reference="A^k"} implies that for \(k\ge 1\) we have \[r_{2k-1}(X)=\sum _{i=1}^k{2k-i-1\choose i-1}u^{i-1}+\sum _{i=1}^k(i-1){2k-i-1\choose i-1}u^{i-2}X^t+\sum_{i\geq 1}v_iX^{t+i}\] and \[r_{2k}(X)=\sum_{i=0}^{k-1}{2k-i-1\choose i}u^i+\sum_{i=0}^{k-1}i{2k-i-1\choose i}u^{i-1}X^t+\sum_{i\geq 1}v'_iX^{t+i}\] for some \(v_i,v_i'\in \mathbb{C}\), and let \(r_0(X)=0\). We define \(h_0(u):=0\) and for \(k\ge 1\) let \[h_{2k-1}(u):=\sum_{i=1}^k {2k-i-1\choose i-1}u^{i-1} \text{ and } h_{2k}(u):=\sum_{i=0}^{k-1}{2k-i-1\choose i}u^i.\] It follows that \[\label{r<->c} r_k(X)=h_k(u)+h_k'(u)X^t+\, \mathrm{higher}\, \mathrm{terms}\] for all \(k\ge 0\) (where \(h_k'\) is the derivative of \(h_k\)). Since \(r_k(X)=r_{k-1}(X)+(X^t+u)r_{k-2}(X)\) for \(k\ge 2\), we therefore get \[h_0(u)=0 ,\quad h_1(u)=1, \quad \mathrm{and}\quad h_k(u)=h_{k-1}(u)+uh_{k-2}(u)\,\, \mathrm{ for}\, k\ge 2.\] Note that we have assumed that \(u\ne 0\). Assume now that \(u\ne-\frac{1}{4}\) as well. Then we can show, either by solving the above difference equation or by induction, that \[h_k(u)=\frac{1}{\sqrt{1+4u}}\left(\left(\frac{1+\sqrt{1+4u}}{2}\right)^k-\left(\frac{1-\sqrt{1+4u}}{2}\right)^k\right)\] for each \(k\ge 0\). Here we define the square root as the one that has argument within \([0,\pi)\). An easy computation shows that \[\label{zeros_of_c} h_k(u)=0\Leftrightarrow \sqrt{1+4u}=i\, \mathrm{tg}\, \frac{k'\pi}{k}\quad \text{for some }\, k'\in\{1,\ldots,k-1\}\setminus\left\{\frac{k}{2}\right\}.\] (Note that \(k'=0\) is not possible, since we have assumed that \(u\ne-\frac{1}{4}\).) Fix now \(j\ge 3\) and assume that either \(j=\alpha-2\) and \(t\le \min\{2z+\beta,m\}\) or \(j=\alpha-1\) and \(t\le \min\{ m+z-\beta,m-z+\beta\}\) (and hence \(\alpha \ge 4\) in both cases). Let \(u_j=-\frac{1}{4\cos^2\frac{\pi}{j}}\). Then \(\sqrt{1+4u_j}=i\, \mathrm{tg}\, \frac{\pi}{j}\) and \(h_j(u_j)=0\) by ([\[zeros_of_c\]](#zeros_of_c){reference-type="ref" reference="zeros_of_c"}). We now consider \(k\ne j\). We will show that \(h_k(u_j)\ne 0\). Assume first that \(k<j\). If \(k'>\lfloor \frac{k}{2}\rfloor\), then \(\mathrm{tg}\, \frac{k'\pi}{k}<0\) and therefore \(i\, \mathrm{tg}\, \frac{k'\pi}{k}\) cannot be a square root and hence equal to \(\sqrt{1+4u_j}\). On the other hand, if \(k'\le \lfloor \frac{k}{2}\rfloor\), then \(\frac{k'}{k}\ge \frac{1}{k}>\frac{1}{j}\) and hence \(i\, \mathrm{tg}\, \frac{k'\pi}{k}\ne i\, \mathrm{tg}\, \frac{\pi}{j}=\sqrt{1+4u_j}\). On the other hand, if \(k>j\), then the only possibility is \(j=\alpha-2\) and \(k=\alpha-1\), as \(k<\alpha\). If \(k-1\ge k'>\lfloor \frac{\alpha-1}{2}\rfloor\), then \(\mathrm{tg}\, \frac{k'\pi}{k}<0\) and hence \(i\, \mathrm{tg}\, \frac{k'\pi}{k}\ne \sqrt{1+4u_j}\). On the other hand, it is clear that \(i\, \mathrm{tg}\, \frac{\pi}{\alpha-1}\ne \sqrt{1+4u_j}\) and that for \(2\le k'\le \lfloor \frac{\alpha-1}{2}\rfloor\) we have \(\frac{k'}{\alpha-1}>\frac{1}{\alpha-2}\), which again implies \(i\, \mathrm{tg}\, \frac{k'\pi}{k}\ne \sqrt{1+4u_j}\). Using ([\[zeros_of_c\]](#zeros_of_c){reference-type="ref" reference="zeros_of_c"}) we therefore see that \(h_k(u_j)=0\) if and only if \(k=j\), and ([\[r\<-\>c\]](#r<->c){reference-type="ref" reference="r<->c"}) then implies that \(r_k(X)\) is not divisible by \(X\) for \(k\ne j\) and that \(r_j(X)\) is divisible by \(X^t\) if we define \(p(X)=X^t+u_j\). It remains to show that \(r_j(X)\) is not divisible by \(X^{t+1}\), which is by ([\[r\<-\>c\]](#r<->c){reference-type="ref" reference="r<->c"}) equivalent to \(h_j'(u_j)\ne 0\). In the neighborhood of \(1-\frac{1}{\cos^2\frac{\pi}{j}}\) the square root is a holomorphic function, so we can compute \[\begin{aligned}
h_j'(u)=&-\frac{2}{(1+4u)^{\frac{3}{2}}}\left(\left(\frac{1+\sqrt{1+4u}}{2}\right)^j-\left(\frac{1-\sqrt{1+4u}}{2}\right)^j\right)+\\ &+\frac{j}{1+4u}\left(\left(\frac{1+\sqrt{1+4u}}{2}\right)^{j-1}+\left(\frac{1-\sqrt{1+4u}}{2}\right)^{j-1}\right)=\\ =&-\frac{2}{1+4u}h_j(u)+\frac{j}{1+4u}\left(\left(\frac{1+\sqrt{1+4u}}{2}\right)^{j-1}+\left(\frac{1-\sqrt{1+4u}}{2}\right)^{j-1}\right). \end{aligned}\] In particular, \[h_j'(u_j)=\frac{j}{1+4u_j}\left(\left(\frac{1+i\, \mathrm{tg}\, \frac{\pi}{j}}{2}\right)^{j-1}+\left(\frac{1-i\, \mathrm{tg}\, \frac{\pi}{j}}{2}\right)^{j-1}\right)=\frac{j\cos\frac{(j-1)\pi}{j}}{2^{j-2}(1+4u_j)\left(\cos\frac{\pi}{j}\right)^{j-1}}\ne 0\] for \(j\ge 3\), which had to be proved. It follows that all the partitions obtained in case 1.2(b) indeed correspond to nilpotent orbits in \(\mathcal{N}(B)\).
## Case 2: \(C_0\) is a zero matrix
Since \(A\) is non-zero and \(z\) is maximal possible, we get that \(C_1\) is not nilpotent. We consider two subcases.
### **Case 2.1: \(C_1\) is invertible**
In this case it is clear that \(\mathrm{rank}\, A^k=2n-2k(z+1)\) for \(k<\frac{n}{z+1}\) and \(A^k=0\) for \(k\ge \frac{n}{z+1}\). Let \(\alpha=\left\lceil \frac{n}{z+1}\right\rceil\). Since we have assumed at the beginning of the proof that \(A\ne 0\), we have \(\alpha \ge 2\), and hence \[\mathrm{rank}\, A^k-\mathrm{rank}\, A^{k+1}=\left\{ \begin{array}{ll} 2(z+1)&k\le \alpha-2,\\ 2n-2(\alpha-1)(z+1) & k=\alpha-1,\\ 0&k \geq \alpha, \end{array} \right.\] so \(A\) corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P1) \[\left(\alpha^{ 2n-2(\alpha-1)(z+1)},(\alpha-1)^{2\alpha(z+1)-2n}\right)=[2n]^{2(z+1)}.\] Conversely, it is clear from the above proof that for each partition \([2n]^{2(z+1)}\) where \(0\le z\le n-1\) there exists \(A\in \mathcal{N}(B)\) corresponding to this partition. This concludes the proof for the even cases in the first set of partitions stated in the theorem.
### **Case 2.2: \(C_1\) is not invertible**
Since \(C_1\) is not nilpotent and the multiplication with a non-zero constant has no influence on the Jordan canonical form of a nilpotent matrix, we may assume that \(C_1\) is idempotent, so \(C_1^2=C_1\). Moreover, we may conjugate \(C\) by an invertible matrix in \(\mathcal{C}(B)\) to assume that \(C_1=\left[ \begin{array}{cc}1&0\\0&0 \end{array} \right]\) or, equivalently, \(a_1=1\) and \(b_1=c_1=d_1=0\). We write matrix \(C\) as \(C=\left[ \begin{array}{cc}Xa'(X)&X^2b'(X)\\X^2c'(X)&X^2d'(X) \end{array} \right]\) for some polynomials \(a'(X),b'(X),c'(X),d'(X)\). As the constant term of \(a'(X)\) is 1, it is invertible in \(S\). Therefore, by defining first the constant term, then the linear term, and so on, we may find \(f(X),r(X)\in S\) such that the equalities \[\begin{aligned}
a'(X)\left(f(X)-d'(X)\right)&=&-Xc'(X)b'(X)-Xf(X)d'(X)+Xf(X)^2,\\ a'(X)r(X)&=&X\left(c'(X)+d'(X)r(X)-b'(X)r(X)^2\right) \end{aligned}\] are satisfied in \(S\). We furthermore define \[e(X)=a'(X)+Xb'(X)r(X) \text{ and } q(X)=-Xb'(X)\left(a'(X)-Xf(X)\right)^{-1}\] in \(S\), and the matrices \(P=\left[ \begin{array}{cc}1&q(X)\\r(X)&1 \end{array} \right]\) and \(D=\left[ \begin{array}{cc}Xe(X)&0\\0&X^2f(X) \end{array} \right]\). Note that \(e(X)\) is invertible in \(S\), while \(q(X)\) and \(r(X)\) are divisible by \(X\), which implies that matrix \(P\) is invertible in \(M_2(S)\). Furthermore, a short calculation shows that \(CP=PD\), so we may replace \(C\) by a diagonal matrix \(D=P^{-1}CP=\left[ \begin{array}{cc}Xe(X)&0\\0&X^mg(X) \end{array} \right]\) for some \(m\ge 2\) where \(e(X),g(X)\in S\) are invertible or \(g(X)=0\) and \(m=n\). Consequently, \(P^{-1}AP=\left[ \begin{array}{cc} X^{z+1}e(X)&0\\0&X^{z+m}g(X) \end{array} \right]\) and it is clear that such a matrix corresponds to the partition Table [1](#tab:nn){reference-type="ref" reference="tab:nn"}(P5) \[([n]^{z+1},[n]^{z+m})\] and that all such partitions are possible. Note that here we have \(m\ge 2\); the case \(m=1\) was considered in the case 2.1.
# Examples and open questions {#sec:examples}
We conclude our paper with a few examples that illustrate the methods and results obtained by our work.
Moreover, we also provide the complete list of \({\mathcal O}_{{\bf{\lambda}}}\) that have non-empty intersection with the nilpotent centralizer of a nilpotent matrix \(B\in {\mathcal O}_{(7,7)}\) in Table [3](#tab:77){reference-type="ref" reference="tab:77"}. It shows that also partitions of type (P6), (P8), and (P9) appear for \(n=7\).
Note that examples \(n=6\) and \(n=7\) show different structures of partitions \(\lambda\) for which the nilpotent orbit \({\mathcal O}_{\lambda}\) intersects \({\mathcal N}(B)\). Although partitions of type (P7) do not appear in these two cases, note that for example for \(n=22\) (\(z=1\), \(m=3\), \(l=5\), \(\alpha=4\), \(\beta=5\)) the partition \((7^3,6^3,5)\) is of the form (P7) but not (P1)-(P6). This shows that, although types (P1)-(P9) of partitions in Table [1](#tab:nn){reference-type="ref" reference="tab:nn"} are not complementary, they are all significant to describe \({\mathcal N}(B)\).
Although we were able to characterize nilpotent orbits \({\mathcal O}_{{\bf{\lambda}}}\) that have non-empty intersection with the nilpotent centralizer of a nilpotent matrix \(B\in {\mathcal O}_{(n,n)}\), the partitions \({\bf{\lambda}}\) in Table [1](#tab:nn){reference-type="ref" reference="tab:nn"} are presented in a rather scattered way.
Moreover, there are plenty of possible generalizations to Theorem [\[main\]](#main){reference-type="ref" reference="main"} that would be interesting to resolve. | {'timestamp': '2022-12-19T02:10:47', 'yymm': '2212', 'arxiv_id': '2212.08425', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08425'} |
# Introduction
Whenever the statistics of a measurement on a composite quantum state contradict the assumptions of local realism, thus violating a Bell-type inequality, the correlations are referred to as nonlocal. These nonlocal correlations are used to certify private randomness in device-independent quantum key distribution (DIQKD) and device-independent randomness generation (DIRNG). For quantifying randomness, estimating the guessing probability is often an important task. The guessing probability is the probability with which an adversary can guess an outcome of another party's measurement. If the guessing probability is less than 1, the adversary cannot predict the outcome with certainty. This implies the presence of intrinsic randomness in the system. However, bounding the guessing probability is not an easy task. Typically it is not possible to explicitly compute the guessing probability, but one can only provide an upper bound by solving a semi-definite optimization problem. Usually, one bounds the guessing probability from a given Bell inequality, and the corresponding quantum violation . Here, one needs to use the hierarchical structure of the quantum correlations to solve the semi-definite optimization problem. The complexity of this optimization problem is increasing and becoming computationally demanding with the number of settings and outcomes. In this paper, motivated by the outstanding recent progress in utilizing machine learning in the field of quantum information, we develop deep learning (DL) models that predict the guessing probability along with the optimal Bell inequalities (used to upper bound the guessing probability) from an observed probability distribution using supervised machine learning. A crucial element of supervised machine learning is to generate sample data input and output to train the model. Here, we sample random quantum probability distributions and use them as the input of the training data. With this data, using the two-step method of Ref., we estimate the upper bound of the guessing probability and the optimal Bell inequality, and use it as the output of the training data. After sufficient training, our DL approach can recognize the pattern and predict the guessing probability and the optimal Bell inequality with high accuracy and low average statistical error. We organize this work as follows. We start in Sec. [2](#Bell setup){reference-type="ref" reference="Bell setup"} by explaining the generalized Bell set-up, types of correlations and Bell inequalities. We introduce the guessing probability and show how to estimate it by solving a semi-definite programme in Sec. [3](#Guessing probability){reference-type="ref" reference="Guessing probability"}. We introduce our deep learning approach in Sec. [4](#sec ML approach){reference-type="ref" reference="sec ML approach"}. We discuss how to sample quantum probability distributions from the quantum correlation space, which are then used as input for supervised learning. We build several deep learning models for predicting the guessing probability and the Bell inequality for various Bell scenarios and measure their efficiency to show the model's utility.
# Generalized Bell Set-up {#Bell setup}
In this section, we introduce a generalized Bell set-up. In each measurement round, two parties, Alice and Bob, share a quantum state \(\rho_{AB}\) acting on \(\mathcal{H}_{A} \otimes \mathcal{H}_{B}\). In the presence of an eavesdropper Eve, her side information \(E\) is described via the purification of the joint system \(\rho_{ABE}\) acting on \(\mathcal{H}_{A} \otimes \mathcal{H}_{B} \otimes \mathcal{H}_{E}\) where \(\mathrm{Tr}_{E}\left(\rho_{ABE}\right)=\rho_{AB}\). Each party selects locally an input (a measurement setting) which produces an output (a measurement outcome). We refer to this scenario as a Bell scenario. Alice performs measurements specified by her input \(x \in X = \{1,\cdots,m\}\), where each input has \(k\) possible outcomes \(a \in A= \{ 1, \cdots, k \}\). Similarly, Bob performs measurements specified by his input \(y \in Y = \{1,\cdots,m\}\) and produces the outputs \(b \in B= \{ 1, \cdots, k \}\). We denote this scenario as \([m,k]\) Bell scenario, i.e. \(m\) measurement settings with \(k\) outcomes each; see Fig. [\[Bell Scenario figure\]](#Bell Scenario figure){reference-type="ref" reference="Bell Scenario figure"} for visualization. After many repetitions, the conditional probability \(P(ab|xy)\) can be estimated. The Bell scenario is completely characterized by the set \(\textbf{P}:=\{ P(ab|xy) \} \subset \mathds{R}^{m^2k^{2}}\) of all joint conditional probabilities which we refer to as a behavior. Thus, the following constraints are imposed: positivity \(P(ab|xy) \geq 0\) \(\forall\) \(a,b,x,y\) and the normalization \(\sum_{a,b=1}^{k} P(ab \arrowvert xy)=1\) for all \(x\) and \(y\). We say the behavior is no-signaling if the input-output correlation obeys \[\label{No-signalling eq} \begin{aligned} \sum_{b=1}^{k}P(ab|xy) &= P(a|x) \qquad \forall a,x,y \quad \textup{and} \\ \sum_{a=1}^{k}P(ab|xy) &= P(b|y) \qquad \forall b,x,y \,. \end{aligned}\] The set of all correlations satisfying the no-signaling constraints forms a convex polytope \(\mathcal{NS}\). A behavior is said to be local if it can be written as a convex mixture of deterministic strategies. The set of all local correlations forms a convex polytope \(\mathcal{P}\). There exist inequalities of the form \[\label{Bell Inequality} \sum_{a,b,x,y} C_{abxy}P(ab|xy) \leq \mathcal{I}_{L} \,,\] which separate the set of all local correlations (in other words, the convex polytope \(\mathcal{P}\)) from the nonlocal behaviors. These inequalities are called Bell inequalities. A Bell inequality is specified by the coefficients \(C_{abxy} \in \mathds{R}\). We denote a Bell inequality as \(B\), and \(\sum_{a,b,x,y}C_{abxy}P(ab|xy)\) as the Bell value \(B[\textbf{P}]\) in this paper. Here, \(\mathcal{I}_{L}\) is the classical bound, which is the maximal value over all local behaviors. Thus, a behavior with a classical origin, i.e. \(\{P(ab|xy)\} \in \mathcal{P}\), cannot violate this inequality. The Born rule of quantum theory postulates that a behavior is quantum if there exists a quantum state \(\rho_{AB}\) acting on a joint Hilbert space \(\mathcal{H}_A \otimes \mathcal{H}_B\) of arbitrary dimension and measurement operators (POVM elements) \(\{M_{a|x}\}\) with \(M_{a|x} \ge 0\) and \(\sum_a M_{a|x} = \mathds{1}\) \(\forall x\), and \(\{M_{b|y}\}\) with analogous properties such that \[P(ab|xy) = \mathrm{Tr}(\rho_{AB} M_{a|x}\otimes M_{b|y})\,.\] The set of all quantum correlations forms a convex set \(\mathcal{Q}\). If a behavior \(\{P(ab|xy)\} \in \mathcal{Q} \setminus \mathcal{P}\), it violates at least one Bell inequality of the form in Eq. ([\[Bell Inequality\]](#Bell Inequality){reference-type="ref" reference="Bell Inequality"}). The sets \(\mathcal{P}\), \(\mathcal{Q}\) and \(\mathcal{NS}\) obey the following relation: \(\mathcal{P} \subsetneq \mathcal{Q} \subsetneq \mathcal{NS}\); see Fig. [\[CQNS\]](#CQNS){reference-type="ref" reference="CQNS"} for a pictorial representation.
# Guessing probability {#Guessing probability}
In an adversarial black box scenario framework, the adversary Eve tries to guess some outcomes obtained by Alice and Bob. The probability that Eve can correctly guess the outcome is called the guessing probability. Here, we denote the guessing probability as \(P_g(a|x,E)\), which is the guessing probability of Eve about Alice's outcome \(a\) corresponding to her measurement setting \(x\). In Ref., it is shown that \(P_g(a|x,E)\) can be upper bounded by a function \(G_x\) of the observed Bell value \(B[\textbf{P}]\) of a particular Bell inequality \(B\) by semi-definite programming, i.e. \(P_g(a|x,E) \leq G_x(B[\textbf{P}])\). One crucial element to bound the guessing probability \(P_g(a|x,E)\) is to choose a suitable Bell inequality. We follow the two-step procedure of where the Bell inequality is constructed from the input-output probability distribution **P** that leads to the maximum Bell violation for that particular measurement statistics. This is achieved by solving the linear program: \[\label{linear_optimization} \begin{aligned} & \underset{\textbf{h},c}{\text{max}} & & \textbf{h}^T \textbf{P}-c \, ,\\ & \text{subject to} & & \textbf{\textbf{h}}^T \textbf{v}_{p} \leq c \quad \forall \quad p \in \{1,\cdots,k^{2m} \} \,, \\ &&& \textbf{h}^T \textbf{P} > c \, ,\\ &&&-1 \leq h_i \leq 1 \quad \forall \quad i \in \{1,\cdots,m^2k^2 \}\,. \end{aligned}\] Here \(\textbf{h}\) is the hyperplane specifying the Bell inequality \(B\), **P** denotes the measurement data, \(\textbf{v}_{p}\) corresponds to the \(p\) vertices of the classical polytope \(\mathcal{P}\) and \(c\) is the classical bound. Thus the Bell inequality \(B\) found by the optimization of Eq. ([\[linear_optimization\]](#linear_optimization){reference-type="ref" reference="linear_optimization"}) and specified by the hyperplane vector **h** is given as: \[\label{Hyperplane BI} \sum_{a, b, x, y} h_{abxy}P(ab|xy) \leq c \,,\] where \(a \in A\), \(b \in B\), \(x \in X\), \(y \in Y\). We will use the Bell inequality \(B\) and corresponding Bell value \(B[\textbf{P}] =\sum_{a, b, x, y} h_{abxy}P(ab|xy)\) to upper bound the guessing probability \(P_g(a|x,E)\) by solving the following semidefinite program: \[\label{SDP} \begin{aligned}[t] & \underset{ \rho_{AB},\{A(a|x)\}, \{B(b|y)\}}{\text{max}}\quad P_{g}(a|x,E)\\ & \text{subject to:} \quad \mathrm{Tr}(\rho_{AB} \mathcal{G})=B[\textbf{P}] \,. \end{aligned}\] In the optimization problem of Eq. ([\[SDP\]](#SDP){reference-type="ref" reference="SDP"}), the guessing probability is bounded using the NPA-hierarchy up to level 2. The optimization is performed using standard tools YALMIP, CVX and QETLAB. Note that, \(A(a|x)\) and \(B(b|y)\) are the measurement operators of Alice and Bob, respectively, and \(\rho_{AB}\) is the state shared between them. \(\mathcal{G}\) is the Bell operator defined as: \[\mathcal{G}=\sum_{a, b, x, y} h_{abxy} A(a|x) B(b|y) \,.\] Let us denote \(P^{*}_{g}(a|x,E)\) as the upper bound of the guessing probability, which is the solution to the optimization problem of Eq. ([\[SDP\]](#SDP){reference-type="ref" reference="SDP"}).
# Machine learning approach {#sec ML approach}
Providing an upper bound for the guessing probability by solving a semi-definite program is a computationally demanding task. It becomes arduous when the Bell scenario raises its complexity, i.e. for an increased number of measurement inputs and/or outputs in the Bell scenario. Thus, in this paper, we approach solving the problem via machine learning (ML) (see Ref. for detailed discussions on the concepts of machine learning) so that the trained model can estimate the guessing probability \(P^{*}_g(a|x,E)\) from the input-output probability distribution \(\{P(ab|xy)\}\). We are going to use the supervised learning technique. In a supervised ML approach, the first step is generating the training points. We use random bipartite quantum probability distributions as the supervised ML model's input (features), after generating them from facet Bell inequalities using the *weighted vertex sampling* method. Since the guessing probability for local behaviors is always 1 (i.e. Eve can guess the right outcome with probability 1), we do not need to train the machine to perform well on those. Thus we only take samples from the nonlocal part of the no-signaling set, i.e. \(\mathcal{NS} \setminus \mathcal{P}\). To single out the input-output correlation with a quantum realization, we reduce the samples using the NPA hierarchy to approximate the quantum realizable probability distribution. Explicitly, we generate samples from the quantum set \(\mathcal{Q}\) as follows. For the \([m,k]\) Bell scenario (i.e. \(m\) measurements, \(k\) outcomes each), the classical polytope \(\mathcal{P}\) is specified by \(k^{2m}\) local vertices. The classical polytope can also be described by its facets, which represent the hyperplanes (or Bell inequalities) that separate any non-classical (quantum and no-signaling) behavior from the classical ones. These facets are called facet Bell inequalities or tight Bell inequalities; see Fig. [\[CQNS\]](#CQNS){reference-type="ref" reference="CQNS"} for a pictorial representation. For the \([2,2]\) scenario, eight facet Bell inequalities exist, all equivalent to the CHSH inequality. For the \([3,2]\) Bell scenario, there are 648 facet Bell inequalities. These facet Bell inequalities are found using the formulation of Ref. [^1]. Note that all the 648 facet Bell inequalities correspond to two classes of independent facet Bell inequalities, i.e. the CHSH inequalities and the \(I3322\) inequalities. We consider all facet Bell inequalities for the \[2,2\] and \[3,2\] Bell scenario while generating training points for the supervised machine learning problem. For the \([4,2]\) Bell scenario, there are 174 independent facet Bell inequalities. Since there are many (\>10000) equivalent facets , we will only consider the independent ones. These facet Bell inequalities are spanned by some of the local vertices of the classical polytope [^2]. These vertices provide the maximum classical bound of the corresponding facet Bell inequality. Consider the case that \(n\) local vertices span a facet Bell inequality, where we denote the set of \(n\) vertices as \(\left\{P_{i}^{\mathcal{L}}(ab|x y)\right\}_{i=1}^{n}\). We denote the PR-box of the corresponding facet Bell inequality as \(P^{\mathrm{PR}}(ab|xy)\), see Fig. [\[CQNS\]](#CQNS){reference-type="ref" reference="CQNS"} for visualization. The PR-Box \(P^{\text{PR}}(ab|xy)\) can be defined as the probability distribution that provides the maximal no-signaling bound of the corresponding facet Bell inequality. We take uniformly random weighted mixtures of the \(n+1\) vertices (\(n\) vertices that span the facet Bell inequality and the corresponding PR-box) with an \(n\)-fold weight on the PR-box. Formally, the sample behavior from the set \(\mathcal{NS}\setminus \mathcal{P}\) can be generated as: \[\textbf{P}:=P(a b|x y)=\frac{n w_{0} P^{\mathrm{PR}}(ab|xy)+\sum_{i=1}^{n} w_{i} P_{i}^{\mathcal{L}}(ab|xy)}{n w_{0}+\sum_{i=1}^{n} w_{i}}\] where the \(w_{i} \in [0,1]\) with \(i=0,1,\cdots,n\), are uniformly drawn random numbers. This process is done for all facet inequalities. From this set of samples, we only select the ones with a \(\mathcal{Q}_2\) realization (the second level of NPA hierarchy ). Here we work under the assumption that \(\mathcal{Q}_2\) provides a good approximation for the original quantum set \(\mathcal{Q}\). We store the probability distribution {\(P(ab|xy)\)} and use it as the input (features) of the supervised machine learning problem, i.e. \[\label{input of model} \textbf{X}:=\{ P(ab \arrowvert xy) \}^{a,b=1,\cdots,k}_{x,y=1,\cdots,m} \,.\] We calculate the guessing probability of each input **P** using the two-step method (see Sec. [3](#Guessing probability){reference-type="ref" reference="Guessing probability"} for details), and use it as the output (target), i.e. \[\label{output of model 1} y= P^{*}_g(a|x,E)\,.\] Without loss of generality, we have always calculated the guessing probability of Alice's first measurement setting. We use a deep neural network to assess the dataset and make predictions. We fed the input-output pair \(\{\textbf{X},y\}\) (see Eq. ([\[input of model\]](#input of model){reference-type="ref" reference="input of model"}) and Eq. ([\[output of model 1\]](#output of model 1){reference-type="ref" reference="output of model 1"})) into an artificial neural network (ANN) to learn the best possible fit. For an elaborate explanation of an artificial neural network, see Ref.. Following the standard approach, we divide the dataset into two parts. The first part of the dataset is for training and validation (80%), and the second is for testing (20%). We choose a 'linear' [^3] ANN with several layers as our model; see Fig. [\[FFNN_oneNN\]](#FFNN_oneNN){reference-type="ref" reference="FFNN_oneNN"} for visualization. The input layer has \(m^2k^2\) neurons corresponding to the elements in {\(P(ab\arrowvert xy)\)}. The output (last) layer has only one neuron since we only have to predict one element: the guessing probability \(P^{*}_g(a|x,E)\). We perform 100 rounds of training using the optimizer ADAM, of which the first 50 rounds have a fixed learning rate of 0.001. For the next 50 rounds, we reduce the learning rate by 90% in every tenth round. We choose the activation function ReLu (Rectified linear unit) [^4] in the input and the hidden layers while using the sigmoid activation function[^5] in the output layer. The ReLu activation function introduces non-linearity and the sigmoid activation function keeps the output within the range of \[0,1\]. The mean squared error (MSE) [^6] is used as our loss function, which is minimized during the training process.
The trained model generates the predicted value of the guessing probability \(P^{\text{pred}}_g(a|x,E)\). To check the efficiency of our approach, we have used the mean absolute error (MAE) \[\label{Eq. MAE} \begin{aligned} &\text{MAE}\left[ P^{*}_g(a|x,E),P^{\text{pred}}_g(a|x,E) \right] \\ &= \frac{1}{N_{\mathrm{test}}}\sum_{i=1}^{N_{\mathrm{test}}} \left | P^{*}_g(a|x,E)_i-P^{\text{pred}}_g(a|x,E)_i \right| \,, \end{aligned}\] and the mean squared error (MSE) \[\label{Eq. MSE} \begin{aligned} &\text{MSE}\left[P^{*}_g(a|x,E),P^{\text{pred}}_g(a|x,E)\right] \\ &= \frac{1}{N_{\mathrm{test}}}\sum_{i=1}^{N_{\mathrm{test}}} \left( P^{*}_g(a|x,E)_i-P^{\text{pred}}_g(a|x,E)_i \right)^2 \,, \end{aligned}\] as a performance measure. \(N_{\mathrm{test}}\) is the number of data points in the test set. We analyze the results for different bipartite Bell scenarios and list the errors in Table [1](#Table_prob_pGuess_error){reference-type="ref" reference="Table_prob_pGuess_error"}. The average error is in the order of \(10^{-4}\) to \(10^{-2}\). Such high accuracy and small error without knowing the Bell inequality are truly remarkable.
We also compare the runtime performance of the neural network model with the frequently used SDP solver Mosek (that can be used to upper bound the guessing probability by solving the optimization problem of Eq. ([\[SDP\]](#SDP){reference-type="ref" reference="SDP"})) in Table [2](#runtime comparison 1){reference-type="ref" reference="runtime comparison 1"}. The Mosek task is generated and solved using the Ncpol2sdpa. The results are evaluated over 10000 unknown samples and performed on a personal computer [^7] under comparable conditions. Once the neural network is trained, we get a speed-up of \(10^{3}-10^{5}\) for obtaining a prediction about a new instance, compared to the runtime of the usual method for solving the optimization problem; see Table [2](#runtime comparison 1){reference-type="ref" reference="runtime comparison 1"}. This follows from the fact that the number of variables in the optimization process of Eq. ([\[SDP\]](#SDP){reference-type="ref" reference="SDP"}) increases exponentially with the number of measurement settings (or outcomes per measurement) in the Bell scenario. Thus, it takes more computational time to perform the SDP using a classical solver like Mosek. A trained neural network only calculates the functional output using the optimized weights and biases. Only the neural network size affects the computational time needed to complete the prediction task.\
However, the upper bound on the guessing probability calculated from a trained machine learning model only provides an estimate of its real value. Thus, we cannot use this estimate to bound the secret key rate. The predicted Bell inequality on the other side that generates a non-zero Bell violation (for a particular measurement statistics) can be used to bound the guessing probability (see Sec. [3](#Guessing probability){reference-type="ref" reference="Guessing probability"} for details) and the secret key rate. That's why in the next step, we use deep learning to predict the associated optimal Bell inequality \(B\), which is then used to upper bound the guessing probability (see Sec. [3](#Guessing probability){reference-type="ref" reference="Guessing probability"} for details). For this purpose, we again use the neural network architecture where supervised learning is incorporated. We start by preparing the dataset where our input features are \[\label{input of model 2} \textbf{X}:=\{ P(ab \arrowvert xy)\}^{a,b=1,\cdots,k}_{x,y=1,\cdots,m} \,.\] Note that, the input is identical to Eq. ([\[input of model\]](#input of model){reference-type="ref" reference="input of model"}). The outputs are now the coefficients of the optimal Bell inequality \(B\) (specified by \(\{h_{abxy}\}^{a,b=1,\cdots,k}_{x,y=1,\cdots,m}\), see Eq. ([\[linear_optimization\]](#linear_optimization){reference-type="ref" reference="linear_optimization"})) and the guessing probability \(P^{*}_g(a|x,E)\), i.e. \[\label{output of model 2} \textbf{y}:=\left[\{h_{abxy}\}^{a,b=1,\cdots,k}_{x,y=1,\cdots,m},P^{*}_g(a|x,E)\right]\,.\] Here we use two types of neural network architectures. The first neural network is a usual 'linear' feed forward neural network (see for details, schematically represented in Fig. [\[FFNN_oneNN\]](#FFNN_oneNN){reference-type="ref" reference="FFNN_oneNN"}). For \([m,k]\) Bell scenario, the input layer has \(m^2k^2\) neurons (corresponds to the elements of \(\{P(ab\arrowvert xy)\}\)). The input layer is followed by several hidden layers. Unlike in the previous scenario, the output layer has \(m^2k^2+1\) neurons in this case, where \(m^2k^2\) neurons correspond to the coefficients of the Bell inequality \(h_{abxy}\), and one neuron corresponds to the guessing probability \(P^{*}_g(a|x,E)\). In this paper, we denote this construction of the 'linear' deep neural network as \(\mathrm{NN}_1\). Following the standard approach, we divide the dataset \(\{\textbf{X},\textbf{y}\}\) (see Eq. ([\[input of model 2\]](#input of model 2){reference-type="ref" reference="input of model 2"}) and Eq. ([\[output of model 2\]](#output of model 2){reference-type="ref" reference="output of model 2"})) into two sets; the first part of the dataset is for training and validation (80%), and the second part is for testing (20%). Similar to the training of the previous network, we perform 100 rounds (first 50 rounds with a 0.001 learning rate and then reduce the learning rate by 90% in every tenth round) of training using the gradient solver ADAM. Similar to the previous scenario, we use the activation function ReLu in the input and the hidden layers. In the output layer, the linear activation function [^8] is used for \(m^2k^2\) neurons that correspond to the optimal Bell inequality and the sigmoid activation function is incorporated for the neuron that corresponds to the guessing probability. As the cost function, we use the Mean Squared Error (MSE) which is minimized during the training process. In addition, we use another neural network architecture with two parallel sub-models (by using branching) to interpret parts of the output that share the same input. In this construction, the input layer has \(m^2k^2\) neurons corresponding to the elements of the probability distribution \(\{P(ab|xy)\}\) of the \([m,k]\) Bell scenario. The input layer is followed by hidden layers consisting of multiple neurons. Then we bifurcate one hidden layer to create two branches. Several hidden layers then follow both branches; see Fig. [\[FFNN_twoNN\]](#FFNN_twoNN){reference-type="ref" reference="FFNN_twoNN"} for visualization. The first branch of the network is for predicting the coefficients of the optimal Bell inequality \(\{h_{abxy}\}^{a,b=1,\cdots,k}_{x,y=1,\cdots,m}\) and thus has \(m^2k^2\) neurons. The second branch of the network is for predicting the guessing probability. Thus, the output layer will have only one neuron corresponding to \(P^{*}_g(a|x,E)\). In this paper, we refer to this neural network as \(\mathrm{NN}_2\) which is built using the Keras functional API. In \(\mathrm{NN}_2\), we use the ReLu activation function in the input and all the hidden layers. The linear activation function is used in the output layer of the first branch (which predicts the coefficients of the Bell inequality) while the sigmoid activation function is used in the second branch (which predicts the guessing probability). The other details of the training steps are the same as for the \(\mathrm{NN}_1\) neural network stated previously. Both \(\mathrm{NN}_1\) and \(\mathrm{NN}_2\) predict the Bell inequality \(B^{\text{pred}}\) (specified by the predicted coefficients \(\{h_{abxy}^{\text{pred}}\}^{a,b=1,\cdots,k}_{x,y=1,\cdots,m}\)) and the guessing probability \(P^{\text{pred}}_g(a|x,E)\). Since the neural networks \(\mathrm{NN}_1\) and \(\mathrm{NN}_2\) predict two separate entities (the optimal Bell inequality and the guessing probability), we evaluate their performance separately. We use the mean absolute error (see Eq. ([\[Eq. MAE\]](#Eq. MAE){reference-type="ref" reference="Eq. MAE"})) and mean squared error (see Eq. ([\[Eq. MSE\]](#Eq. MSE){reference-type="ref" reference="Eq. MSE"})) as our performance measure of predicting the guessing probability. The errors for different bipartite Bell scenarios are listed in Table [3](#Table_probBIpGuess_pGuess_error){reference-type="ref" reference="Table_probBIpGuess_pGuess_error"}.
Note that, for estimating the guessing probability, \(\mathrm{NN}_2\) yields lower statistical errors than \(\mathrm{NN}_1\). The reason lies in the structure of the neural network architectures. Since we create a branch in the neural network only to estimate the guessing probability, the \(\mathrm{NN}_2\) neural network assigns more nodes to only estimate the guessing probability than \(\mathrm{NN}_1\). In the case of predicting the optimal Bell inequality \(B\) (characterized by its coefficients \(\{h_{abxy}\}^{a,b=1,\cdots,k}_{x,y=1,\cdots,m}\)), we use the performance measure MSE, which reads: \[\begin{aligned} &\text{MSE}\left[B,B^{\text{pred}}\right] \\ &= \frac{1}{m^2k^2} \frac{1}{N_{\mathrm{test}}}\sum_{i=1}^{N_{\mathrm{test}}}\sum_{a,b=1}^{k}\sum_{x,y=1}^{m} \left( \left(h_{abxy}\right)_i-(h_{abxy}^{\text{pred}})_i \right)^2 \,, \end{aligned}\] and MAE, which reads: \[\begin{aligned} &\text{MAE}\left[B,B^{\text{pred}}\right] \\ &= \frac{1}{m^2k^2} \frac{1}{N_{\mathrm{test}}}\sum_{i=1}^{N_{\mathrm{test}}}\sum_{a,b=1}^{k}\sum_{x,y=1}^{m} \left| (h_{abxy})_i-(h_{abxy}^{\text{pred}})_i \right | \,. \end{aligned}\] The errors are listed in Table [4](#Table_probBIpGuess_hyperplane_error){reference-type="ref" reference="Table_probBIpGuess_hyperplane_error"}.
Another way to evaluate the quality of the predicted Bell inequality is to use it for upper bounding the guessing probability problem (see Eq. ([\[SDP\]](#SDP){reference-type="ref" reference="SDP"})). First, we estimate the probability of \(P^{*}_g(a|x,E) < 1\), where \(P^{*}_g(a|x,E)\) is calculated from the predicted Bell inequality \(B^{\text{pred}}\) and the input-output probability distribution \(\{P(ab|xy)\}\) of the test set. We present the results in Table [5](#Table_hyperplane_pGuess_prob){reference-type="ref" reference="Table_hyperplane_pGuess_prob"}.
We also look into the statistical errors between the original guessing probability \(P^{*}_g(a|x,E)\) from the test set and the guessing probability calculated from the predicted Bell inequality \(B^{\text{pred}}\). We use MAE and MSE as the performance measures listed in Table [6](#Table_BIpredictedpGuess_pGuess_error){reference-type="ref" reference="Table_BIpredictedpGuess_pGuess_error"}.
The high probability of generating \(P^{*}_g(a|x,E) < 1\) with the predicted Bell inequalities (see Table [5](#Table_hyperplane_pGuess_prob){reference-type="ref" reference="Table_hyperplane_pGuess_prob"}) and the small statistical errors (see Table [6](#Table_BIpredictedpGuess_pGuess_error){reference-type="ref" reference="Table_BIpredictedpGuess_pGuess_error"}) demonstrate the quality and accuracy of the predicted Bell inequality. We again compare the computational runtime of predicting the optimal Bell inequality using the standard linear optimization of Eq. ([\[linear_optimization\]](#linear_optimization){reference-type="ref" reference="linear_optimization"}) with the neural network \(\mathrm{NN}_1\) and \(\mathrm{NN}_2\). The runtime for different methods is shown in Table [7](#runtime comparison 2){reference-type="ref" reference="runtime comparison 2"}.
Similar to the previous scenario, the runtimes are evaluated over 10000 unknown samples and performed on the same personal computer in the same condition. The linear programming of Eq. ([\[linear_optimization\]](#linear_optimization){reference-type="ref" reference="linear_optimization"}) is performed with the Mosek solver using PICOS python interface. We notice a significant speed-up when using the trained neural network models compared to the Mosek solver. This again follows from the fact that the number of variables in the optimization process of Eq. ([\[linear_optimization\]](#linear_optimization){reference-type="ref" reference="linear_optimization"}) increases with the number of measurement settings (or outcomes per measurement) in the Bell scenario while the computational time for the neural networks only depends on its size.
# Discussion & Conclusion
Estimating the guessing probability is a cornerstone for device-independent quantum key distribution and device-independent randomness generation. This paper introduces a novel method to estimate the guessing probability using trained deep learning models to bypass the computationally complex and cumbersome semi-definite optimization process. Computation with the trained deep learning models is significantly faster than using a conventional solver. With current technology, Bell test event rates are around 100 kHz, which results in new data every 10\(\mu\)s. This frequency is too high for conventional SDP solvers on a single CPU. For those cases, our deep learning approach improves the computation significantly. In principle, optimizing the size of a deep neural network that can process each event as the experiment is being conducted is possible. The deep learning model only provides an estimation of the upper bound of the guessing probability. But it will not provide a certification. Thus, additionally, our DL model provides an estimation of the optimal Bell inequality for which the Bell violation using the measurement statistics certifies the nonlocality of input-output correlations and guarantees that the guessing probability will be less than one. Our trained deep learning models, which significantly speed up the prediction of the Bell inequality compared to a conventional linear program solver, predict a Bell inequality that can generate \(P_g(a|x,E) < 1\) with a very high probability. The mean average error between the guessing probability calculated from the predicted Bell inequality and the optimal Bell inequality (calculated using Eq. ([\[linear_optimization\]](#linear_optimization){reference-type="ref" reference="linear_optimization"})) is in the order of \(10^{-5}-10^{-2}\) (mean squared error is in the order of \(10^{-8}-10^{-3}\)) which shows the quality of this approach such that it can efficiently be used in a DIQKD or DIRNG protocol. We also demonstrate a method for sampling random quantum correlations (correlations which have a realization of NPA hierarchy level of 2) using the facet Bell inequalities, which is then used as input in the supervised machine learning process. Note that, while generating probability distributions, we consider all facet Bell inequalities for the \([2,2]\) and \([3,2]\) Bell scenario. However, since there are more than 10000 facet Bell inequalities for the \([4,2]\) Bell scenario, we only restrict ourselves to generating probability distributions using the independent facet Bell inequalities. To illustrate the benefits of our method, we have applied it to several relevant Bell scenarios. Note that we design and train our neural networks to minimize statistical errors. However, we do not claim that our choice of the trained neural network is optimal for estimating the guessing probability and the associated optimal Bell inequality from the measurement statistics. Other constructions of neural networks will lead to different results. We observed that the statistical errors in the estimation of the guessing probability and the optimal Bell inequality increase with the complexity of the Bell scenario (i.e., the increase in the number of measurements per party). Since there are more inputs and outputs, our neural network architecture might not be able to generalize the extensive system with a limited number of hidden layers and nodes in each layer. To decrease the errors, one can take two steps. First, one can generate a larger dataset to train the model. Second, one can build a more extensive neural network architecture (i.e., more hidden layers or nodes in every layer). However, using a larger dataset for training or/and training a more extensive neural network will result in significantly more computational time. There is also the possibility of overfitting in an extensive network. A larger neural network architecture will also take more time to predict new instances. Therefore, one has to change the network architecture to optimize the speed and precision of a specific scenario. Note that while comparing the runtime for the Mosek optimization solver with the trained neural network for the estimation of the guessing probability (see Table [2](#runtime comparison 1){reference-type="ref" reference="runtime comparison 1"}), we implement the NPA hierarchy of level 2. The difference in computational runtime between the methods will be much more pronounced with increasing hierarchy. Our research demonstrates the applicability of deep learning techniques for Bell nonlocality and upper bounding the guessing probability. We believe that this strategy will create several research lines. The logical next step is to apply our approach to Bell scenarios with a higher number of measurement settings and outcomes. It is also possible to expand our framework to a multipartite scenario. Another direction worth exploring for future work is investigating other neural network constructions. Beyond the advantage in speed, one could use neural network architectures to search for new Bell inequalities. Also, recall that our methodology does not account for uncertainty or offers certification of the output. It remains for future work to use techniques like probabilistic modeling that can certify the correctness of the model's output. | {'timestamp': '2022-12-19T02:12:48', 'yymm': '2212', 'arxiv_id': '2212.08500', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08500'} |
null | null |
# Introduction
In classical relativistic physics, causality is associated with the light-cone structure of Minkowski space-time: no event can be affected by an event lying outside its past light-cone. In relativistic quantum mechanics, the situation is more involved. It is indeed well-known that a relativistic evolution driven by a positive energy Hamiltonian turns instantaneously an initial state having compact spatial support into a distribution having mathematically a non-zero amplitude everywhere in space. Relativistic propagators restricted to the positive energy sector spill outside the light cone: it is only by including the contribution of the anti-particle sector that a causal propagator is obtained. While this observation points to the necessity of having antiparticles in a relativistic quantum theory in order to preserve causality, there are instances in which no negative energies appear. For instance in the Salpeter equation (also known as the relativistic Schrödinger equation), by construction the propagator is restricted to positive energies. This also appears when the solutions of the Klein-Gordon or Dirac equations are unitarily transformed in the Foldy-Wouthuysen (FW) representation. Given the importance of the FW solutions--they are necessary to obtain the classical limit, and it is sometimes claimed that densities constructed from the FW wavefunctions are the only ones having a physical meaning --it is instructive to investigate to which extent there is an effective propagation outside the light cone. Indeed, instantaneous spreading of an initially localized wavefunction is a mathematical fact, but it is often regarded as being physically irrelevant on the ground that beyond the light-cone propagation is extremely small and not detectable in practice for any realistic physical state. In this work, we will investigate numerically the fraction of the wavefunction that effectively propagates outside the light-cone for different initial states characterized by different parameters (width, mean momentum, shape). The common feature to the initial states we will employ is the requirement that they have a compact spatial support. Up to now most works that have investigated this type of propagation, essentially in the context of the Salpeter equation, have used states that have tails at \(t=0\), such as initial Bessel functions (one of the few cases for which analytical solutions can be obtained) or Gaussian wavepackets. If the states have initial compact support, we can meanigfully follow numerically the fraction that remains inside the light cone as time evolves. We will see that although the fraction spilling outside the light cone is small and does so during very short times, it could have observational consequences for elementary particles. To this end, we will first (Sec. [\[sec2\]](#sec2){reference-type="ref" reference="sec2"}) set the context by recalling in which situations one is led to deal only with the propagator of the positive energy sector. We will briefly recall in Sec. [\[sec3\]](#sec3){reference-type="ref" reference="sec3"} the arguments proving the propagator is non-causal. We will give our results in Sec. [\[sec4\]](#sec4){reference-type="ref" reference="sec4"}, describing the method employed and our choice of initial states. We close the paper with a short discussion and conclusive comments (Sec. [\[sec-c\]](#sec-c){reference-type="ref" reference="sec-c"}).
# Positive energy propagation[\[sec2\]]{#sec2 label="sec2"}
## Standard relativistic wave equations
The standard relativistic wave equations for spin-0 and spin-1/2 particles respectively are the Klein-Gordon (KG) and Dirac equations, | {'timestamp': '2022-12-19T02:10:00', 'yymm': '2212', 'arxiv_id': '2212.08400', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08400'} |
# Introduction
20 years ago it had been realized within the lattice QCD community that the large data sets needed for their research projects are not only costly to generate but would also be highly valuable input for research projects of other collaborations. The idea of making the gauge field configurations generated in lattice QCD simulations of various groups available to the wider research community has been brought up at the "\(20^{\rm th}\) International Symposium on Lattice Field Theory" (Lattice 2002) in Boston. It has been suggested to establish within the lattice community an "International Lattice Data Grid" (ILDG), and to use a common metadata schema for markup together with a metadata catalogue for searching. This idea found much attention and first reports on the organization and implementation of ILDG have been given at Lattice 2003 and 2004. Major achievements in the following years were the definition of a community-wide agreed metadata schema (QCDml), and the setup of interoperable storage systems and services. ILDG started as a "grid of grids", i.e. a federation of interoperable but autonomously operated infrastructures and services. The currently five "regional grids" of ILDG are shown in Fig. [\[fig:map\]](#fig:map){reference-type="ref" reference="fig:map"}. Each of them is responsible and free to implement and operate the services in its own (and ILDG-compatible) way. In particular, the regional grids need to acquire the necessary storage resources for their data. ILDG only operates a community-wide user registration and has two working groups which, for instance, elaborate the ILDG-wide metadata schemata and the API for the services of the regional grids. Due to these uniform specifications sharing of data and searching of metadata is possible in a community-wide and seamless way, e.g. through simple web interfaces of JLDG or LDG. Gauge configurations in the US (with somewhat more specific metadata and access ) can be found through the DOE data explorer. The FAIR principles (Findable-Accessible-Interoperable-Reusable) establish essential guiding principles for scientific data management and are becoming a mandatory requirement by funding organisations. FAIR principles realized at all levels of the data management, including intermediate and primary data, is also an important aspect in concepts, like Open Data and Open Science (see also presentations at this conference ). When bringing to life ILDG, the lattice community already made an important step towards FAIR data even before these principles were explicitly formulated and commonly accepted. During the last years, however, usability and usage of ILDG has severely degraded. This has partially been a consequence of the lacking broader uptake of Grid technologies, which made it increasingly difficult keeping the infrastructure and the access to it alive. Instead, Cloud technologies have become the primary choice for realising distributed data infrastructures due to the success of large commercial public Cloud providers. On the other hand, the necessity and interest to share lattice data, at least (but not only) at the level of our costly raw data, is high. This has become evident, for instance, in the parallel session on "Lattice Data" at this conference. More than ten collaborations indicated interest in making order of 500 ensembles openly accessible through ILDG. These sum up to more than 15 million configurations and about 5 PB. Sharing such volumes is a challenging task in view of limited storage resources and person power. Therefore, resuming and joining efforts to restore and improve ILDG seems desirable and timely. Fortunately, ILDG can eventually also leverage on some recent funding, e.g. within the PUNCH4NFDI project, for urgently needed software developments to "go FAIR". The main objectives for "ILDG 2.0" are to improve usability of ILDG and to re-align it with the technological evolution, e.g. towards Cloud storage. Ongoing activities include the transition to token-based authentication (instead of grid certificates), containerization (for easier deployment in different regional grids) and re-factoring of metadata catalogue and user tools, as well as revision of the metadata schemata and support for DOI registration and data publishing. Since some of these topics are not lattice-specific, we also hope to exploit or create synergies with related research-data efforts in other fields.
# Basic Elements of ILDG
## Global Structure and Services
ILDG is a federation of autonomous "Regional Grids" (RG) with a single Virtual Organization (VO). Users can register as members of the VO through a unique registration service which then provides support for further authentication and access to other services. The user registration of ILDG is currently still based on authentication via grid certificates and implemented through a VO Membership Service (VOMS). Moreover, ILDG has agreed on community-wide standards and conventions for (i) a binary storage format of gauge configurations, (ii) a metadata schema and format used for markup of the data, and (iii) a minimal set of services (including their API) to be operated by the regional grids. The VO membership registration service and a web page, where the specifications (i), (ii), and (iii) can be found, are in fact the *only* services operated at a global level by ILDG.
## Services of the Regional Grids
The basic services to be provided and operated by each regional grid are
- a Metadata Catalogue (MDC) with an ILDG-compatible application programming interface (API) to (i) register the unique identifiers for the data (i.e. ensembles and configurations) provided by the RG, (ii) store the corresponding metadata, and (iii) to make it searchable through a powerful query language (Xpath).
- a File Catalogue (FC) which for each configuration identifier returns the list of storage locations (SURLs), where the (possibly replicated) configurations can be accessed through standard protocols.
- storage elements (SE) where the configurations are actually stored.
- a web page with any further RG-specific information.
The URLs of these services from each individual regional grid are kept in a machine actionable file on the ILDG web page, e.g. for use by automated client tools or scripts. The regional grids are responsible for setting up and autonomously operating these basic services following the ILDG specifications. In this way, each RG has large freedom in the implementation choices for services and infrastructure, and vendor locks are avoided. For instance, the regional grids in Japan (JLDG) and Europe (LDG) use rather different solutions for data storage: a global file system (GFARM) in case of JLDG, and several (distributed) storage elements with SRM interfaces being part of the Worldwide LHC Computing Grid in case of LDG. Due to the common API and standard protocols used for the basic services, every regional grid or even individual persons can develop their own client tools or additional services--either according to their own needs and preferences, or as an offer for community-wide use. Examples include convenient web interfaces for simple listing or advanced (meta-)data searches. Of course, neither the Virtual Organization ILDG nor the regional grids are usually direct service providers. Operating the services and, in particular, providing the hardware for storage elements, relies on the support by the home institutions of the scientists or on other sources.
## Organization
ILDG has two working groups: one for metadata-related topics, and one for middleware-related topics. After a phase of inactivity during recent years, the working groups have resumed regular online meetings since beginning of 2022 to discuss status, plans, and ongoing or future activities within the regional grids and ILDG wide. Major topics of the Metadata Working Group (MDWG) currently include adjustments of the metadata schema and support for data publishing (see also Sect. [4.4](#ss:publishing){reference-type="ref" reference="ss:publishing"} below). The Middleware Working Group (MWWG) works on technical aspects including the development, maintenance, and improvement of services, infrastructure, and user tools of ILDG and its RGs (see also Sect. [4.1](#ss:ltools){reference-type="ref" reference="ss:ltools"} below). The ILDG Board is composed of one to two representatives of each RG and the convenors of the working groups. The board meets several times per year to discuss and decide any major strategic and organizational matters of ILDG, including the directions for the activities of the working groups, and the organization of meetings and other outreach activities targeted to the entire the lattice community (e.g. virtual workshops, or the the parallel session on "Lattice Data" and the ILDG lunch meeting at this conference).
## Use Cases and Requirements
Data management and sharing of gauge configurations usually involves two different user perspectives:"data providers", who carry out the simulations and generate the gauge configurations, and "data consumers", who use the configurations for further processing and analysis. Moreover, configuration sharing can take place at different levels: at a collaboration-internal level within one (or, in case of joint projects, few) collaborations, or at a community-wide level. Often, ensembles of gauge configurations which initially are only shared within a collaboration are later on made publicly available for the entire lattice community after some embargo time. The aim of ILDG is to provide a framework that is convenient and beneficial for all these use cases, not only for data consumers or at a community-wide level. At a first glance, this is not obvious, because markup of metadata and packing of configurations according to the schema and format required by ILDG may seem an extra burden for data providers. However, the information required for the ILDG metadata is essentially just the kind of metadata which anyhow is required for any rigorous and high-quality data management and curation, and which can easily be collected in any well organized data production workflow. If this information, basically a key-value list, is properly collected right from the beginning (possibly also in some simple custom format), it is then trivial to convert it into the ILDG markup. Thus, ILDG can help the data providers to guide and organize the workflow for data management and sharing already at the collaboration-internal level in a clear and transparent way, and to guarantee a high level of data quality (e.g. through the checksum and provenance information included in the ILDG metadata). In fact, support for convenient data management and sharing at a collaboration-internal level has been required and intensively used by several collaborations within JLDG or LDG (e.g. ETM or QCDSF) since the beginning of ILDG. Of course, a well-defined and fine-grained access control is a critical technical requirement for this purpose. Then, if configuration sharing through ILDG is already put into place at the collaboration-internal level, making an ensemble eventually community-wide available simply amounts to toggling a flag in the access control settings at the end of the embargo period. Both at collaboration-internal or community-wide level, a suitable way to cite ensembles that were used as input data for further analysis is a basic element of good scientific practice. Moreover, it allows data providers to receive adequate credits and citations, and to properly acknowledge computing grants. Since typically many ensembles, e.g. with different lattice spacings and quark masses, enter into a physics result, citation in form of a lengthy list of ensemble identifiers of ILDG is often not convenient or practical. Therefore, an additional support or workflow to publish (sets of) ensembles in a more standard way, including the assignment of DOIs which then can be conveniently cited, is an important extension of ILDG.
# ILDG goes FAIR
Keeping in mind that ILDG started more than a decade before the FAIR principles were actually formulated in, it is instructive to ask how well the concepts and implementation of ILDG already are compliant with these FAIR principles and where adjustments and extensions are necessary for ILDG 2.0. It is also important to recall that describes guiding principles for scientific data management, not an implementation. In the following we only consider the ten main criteria of which are shown in the text boxes below. A more in-depth analysis, e.g. according to the FAIR Data Maturity Model, would be beyond the scope of this contribution.
## Findable
In ILDG, an identifier of the form \[{\tt lfn:}//\langle rg\rangle/\langle collab\rangle/\langle proj\rangle/\ldots \label{e:lfn}\] or \[{\tt mc:}//\langle rg\rangle/\langle collab\rangle/\langle proj\rangle/\ldots \label{e:mc}\] is assigned to each gauge configuration and to each ensemble, respectively. Including the names of the regional grid (\(rg\)), collaboration (\(collab\)), and project (\(proj\)), together with a suitable convention at the level of each project, guarantees global uniqueness (F1). Persistence of the identifiers is realized in practice for at least the lifetime of the (meta-)data, which in case of the earliest ensembles uploaded in LDG is already 16 years. Gauge configurations and ensembles are described in ILDG by metadata with a rich, hierarchical, and flexible structure (F2) which is concisely defined in terms of two corresponding metadata schemata. As required by (F3), the metadata of ensembles and configurations also include the unique identifiers ([\[e:lfn\]](#e:lfn){reference-type="ref" reference="e:lfn"}) and ([\[e:mc\]](#e:mc){reference-type="ref" reference="e:mc"}). The corresponding metadata elements are called `dataLFN` (data logical file name) and `markovChainURI`, respectively. Since the data files of the configurations are large, they are usually stored in a distributed manner on different systems. In contrast, the metadata is much smaller and kept separately from the data in order to allow efficient searching. The metadata in ILDG is registered and stored in the central Metadata Catalogue (MDC) of each regional grid. The MDC supports complex queries on the content of the metadata and returns the list of matching identifiers ([\[e:lfn\]](#e:lfn){reference-type="ref" reference="e:lfn"}) or ([\[e:mc\]](#e:mc){reference-type="ref" reference="e:mc"}). Thus, the MDC of each regional grid is an essential building block of ILDG to satisfy the requirement (F4).
## Accessible
The Metadata Catalogue in ILDG is a web service (not just a web page) and can be accessed via the standard HTTP(S) protocol and a well-defined API contract. As required by (F4) and (A1), the MDC supports searching through the configuration and ensemble metadata (by Xpath queries), as well as retrieving the *metadata* itself by their `dataLFN` or `markovChainURI`, respectively. Since all metadata in ILDG are public, authentication or authorization is only used for write access (A1.2). The actual *data*, i.e. the gauge configurations, are retrievable by their `dataLFN` (as required by A1) with the help of the File Catalogue (FC), which logically is a separate service in ILDG, but can be also be considered as part of the MDC: first, a query to the FC with a the desired `dataLFN` returns the storage URL (SURL) of the configuration (or a list of SURLs if data is stored redundantly). The SURL includes not only the storage location (address and path), but also a standard protocol (e.g. GSIFTP, HTTP, or SRM) to access the corresponding storage system (A1.1). The data can then be downloaded by standard clients, like `globus-url-copy` or the `gfal2` library. By providing suitable client tools (e.g. scripts), users can conveniently carry out all steps through a single command. The FAIR requirement (A2) is remarkable because it suggests that metadata is to be considered at least as important as the data itself. In ILDG this can and needs to be taken into account, for instance, by frequent (daily) automatic backups of the MDC data. The API of the MDC and FC service for ILDG are defined by the Middleware Working Group. Using the same API in each regional grid ensures compliance with (A1.1) and interoperability[^1] of their services. While the current API is based on the Simple Object Access Protocol (SOAP), the transition to a Representational State Transfer (REST) style is in preparation. In practice, because of IT security concerns, e.g. in data centers, most of the storage systems used by our community also require authentication (A1.2)---even for read access (at least for large amounts of data and fast protocols). Therefore, a global user registration in a VO is essential for providing the authentication and authorization services necessary for seamless and community-wide configuration sharing. The simplified structure of a regional grid in ILDG with findable and accessible data is illustrated in Fig. [\[fig:mdc-schema\]](#fig:mdc-schema){reference-type="ref" reference="fig:mdc-schema"} (without explicitly showing the FC queries and authorization flow). From an abstract point of view, ILDG just implements a database system on a distributed infrastructure by using technologies and concepts of the World Wide Web. The basic entities are configurations and ensembles labeled by unique identifiers. The metadata describing e.g. the physical or administrative properties of the configurations and ensembles---or in case of configurations also the actual data itself---are logically just attributes associated with these identifiers.
## Interoperable
Data and metadata in ILDG satisfy (I1)--(I3) by consistently using well-defined storage and markup formats which are an integral part of the ILDG specifications. Metadata is stored in XML format and must validate against the XSD definition of the corresponding schema. In addition to the XSD specification, further documentation of all elements of the schema can be found on the web page of the ILDG Metadata Working Group. Gauge configurations are stored in binary records with specified order and format (big-endian, 32-or 64-bit IEEE floating-point). Moreover, ILDG specifies a very simple packaging of the data record together with additional metadata (lattice geometry and `dataLFN`, plus optional user-defined metadata) into a single file.
## Reusable
At the highest level, the two metadata schemata of ILDG are grouped into sub-trees, like `management`, `physics`, `algorithm`, `implementation`, or `markovStep`. These are further structured to specify, e.g. lattice geometry, type and parameters of the action and the simulation algorithm, simulation code and hardware, further provenance information (R1.2), checksum, and plaquette value, etc., The two metadata schemata of ILDG seem to meet very well the requirement (R1), in particular (R1.2) and (R1.3). However, a clear data usage license, as required by (R1.1), is missing and must become mandatory in the future, also in order to fully support a data publishing process. Reusability is also related to the important concepts of reproducibility in Open Science (see presentation by E. Bennett at this conference ).
# Towards ILDG 2.0
Main objective of the ongoing reactivation and restart of ILDG, which we coin "ILDG 2.0", is to set up a fully operational, maintainable, FAIR-compliant and user-friendly framework--for data providers as well as data consumers. Improvements needed for this purpose include
- development and maintenance of user tools (e.g. CLI clients), together with up-to-date user documentation
- support for data publishing and DOI registration, in order to make the data citable
- revision of the metadata schemata and data formats (e.g. HDF5)
- support for finding the data through convenient web interfaces or standard search engines
- technology updates, including a VO-wide token-based authentication and authorization
## User Tools {#ss:ltools}
Due to the common API and standard protocols used for the basic services, regional grids or even individual persons are able to develop client tools or additional services. However, coordinating and sharing such efforts in a more coherent way is needed to improve the usability and quality of ILDG at a community-wide level. For the use cases discussed in Sect. 2.4, only a few high-level operations are needed and can be performed by command-line tools, like[^2]
- `lfind` to query the MDC of a regional grid (or of all of them) for ensembles or configurations which have specific properties. Since the MDC supports powerful Xpath queries, one can search for any combination of properties contained in the metadata. For instance, one might want to search for ensembles which have a certain action, number of flavors, lattice size, value of physics or algorithm parameters, etc. In the same way, one might search e.g. for all configurations that were generated by a specific collaboration or code, on a specific machine or in some period of time.
- `lget` with a list of identifiers (`dataLFN` or `markovChainURI`) as argument to either download the corresponding metadata and/or (depending on command-line options) the configuration data itself. In the later case, `lget` should implicitly perform look-up in the FC, authentication (if needed), and verification of the checksum.
- `linit` is a command needed by data providers or project managers to register a new ensemble ID with its metadata, and to set additional attributes, like collaboration and project name, or access rights.
- `lput` can then be used to upload configuration metadata and data. In the later case, the storage element might be selected by default or explicitly specified, and `lput` implicitly carries out registration of the storage URL in the FC, as well as the required authentication and authorization flow
Also the effort to prepare metadata and data for uploading, can be strongly reduced by clearly separating the task into generic and project-specific steps. Project-or collaboration-specific steps usually need to be set up only once. This setup phase amounts to preparing the following prerequisites: a workflow to generate gauge configuration data in ILDG format (directly or by a conversion tool if the simulation code uses a custom format); and creation of a template for the metadata. Such a template can contain all the information, which typically is invariant for an entire set of ensembles (like the description of the action or the algorithm), together with place holders for a limited set of variable values, which can usually be determined in a simple and automatic way, e.g. by trivial scripts or commands to extract them from log files, etc. Then, the repeatedly required steps for actually uploading, namely packing the configurations and generating the metadata, can be handled automatically by two simple and generic tools. A common and shared implementation of these two additional tools can thus greatly simplify the upload in ILDG. Many such convenient tools have been created, either within regional grids or by individual users, in the early days of ILDG. However, due the lack of maintenance, documentation, and porting to newer software or operating system versions, most of them have become unavailable. This can be cured by a careful and well documented re-design of these user tools. They should have minimal dependencies on any specialized software (e.g. by just using scripts, and standard commands, like `curl` and `globus-url-copy`), and/or be packaged into containers (like singularity/apptainer) for convenient deployment on different user platforms (including systems where users do not have root privileges).
## Search Interfaces
A simple command-line tool for searching (like `lfind`) needs as argument a single string which represents the search criteria as an Xpath query. However, properly formulating the query strings for complex searches is not completely straightforward and requires some knowledge of Xpath syntax and of the metadata schema. In addition---and as an alternative---to command-line tools it would, therefore, be important to also provide convenient (web) interfaces, where queries can be formulated at a somewhat higher level, e.g. through a user-friendly selection menu or form similar to INSPIRE-HEP. Since the services (like MDC and FC) of all regional grids have the same API, any implementation of such web interfaces can easily be used at an ILDG-wide level. Extensions of existing attempts, like or the simple listings, can provide a natural and promising starting point.
## Revision of Metadata Schemata and Data Format {#ss:md}
Creating a community-wide agreed metadata schemata at the beginning of ILDG was a challenging task for Metadata Working Group and represents an important achievement of ILDG. The two schemata have now been in use for more than a decade with only minimal changes. Clearly further developments and adjustments are necessary for ILDG 2.0. For instance, some elements of the markup schema, in particular the `action` part in the ensemble metadata, seem to be rather complicated and rigid. This shall be fixed in future revisions by allowing e.g. references to external literature or documents, as well as optional glossaries or annotations. Some metadata elements also need to be adjusted to enable or simplify markup and uploading of data from simulations with new features. Such minor adjustments of the schemata will most efficiently be handled in a use-case driven way: several collaborations, which actively contribute to the reactivation efforts of ILDG, are preparing major uploads. Thus, difficulties or concrete solution proposals, which might arise in this phase, can be directly taken up and finalized by the working groups. On the other hand, major extensions of the metadata schema, or an additional schema, might become necessary in order to fully support a proper data publishing process. For the configuration data, also an alternative packaging of the records according to the widely used HDF5 format and with multiple configurations per file is being discussed as a desirable extension.
## Support for Data Publishing {#ss:publishing}
Possibilities to register persistent identifiers, like DOI, have been discussed in ILDG since a long time. Finding an ILDG-wide solution or at least best practice for this purpose is clearly among the important objectives for ILDG 2.0. Being able to register DOIs for ensembles or sets of ensembles would make citation of these data in publications much more practical and attractive. In turn, this would help to motivate and reward data providers who share and upload their configurations. A closer look at the publishing process, as it is well established for literature, reveals that an analogous process is also desirable and necessary for data. It is not sufficient to just make the data publicly availability on some computer in the internet and to register a DOI. For instance, specific information, i.e. additional metadata, needs to be provided, and a landing page must be generated and hosted. This is not only required for the DOI registration itself, but also for making the data findable from non lattice-specific search engines. Open issues are also the persistence requirements of data and metadata, or explicit support for metadata harvesting, e.g. through an OAI-PMH interface and/or arrangements with portals, like INSPIRE-HEP. Interesting solutions for DOI registration and the entire publishing process have already been explored by JLDG and USQCD, see and. Also services, like Zenodo, seem an attractive option. The Metadata Working Group is now further investigating these approaches in order to find out whether an ILDG-wide solution, or at least an ILDG-wide support and best practice for (possibly different) solutions within the regional grids, can be realized. For instance, already a community-wide schema for collecting the metadata, which is required for DOI registration and for the generation of meaningful landing pages for lattice data, would be an important step forward.
## Technology Upgrades
Several services and building blocks of ILDG have been in use for more than a decade and a transition or upgrade to up-to-date technology is overdue. A key element is the ILDG-wide user management and authentication, which is currently based on X.509 certificates (in parts with VOMS extensions). Switching to a token-based solution (based on OIDC or SAML) is necessary to ensure alignment with technologies commonly used in the Cloud. This will also free the users from the sometimes difficult procedure to obtain a grid certificate. Among the various alternatives that have been investigated by the Middleware Working Group, the INDIGO IAM appears as the most suitable and promising solution. The deployment of a dedicated instance of this service at CNAF in Italy will be further explored and might become a central building block for ILDG 2.0. Further essential elements of ILDG are the Metadata and File Catalogue of each regional grid. The MDC implementation, which has been in continuous operation in LDG since 2008, has recently been re-built and containerized. Thus, it can now be easily deployed also in other regional grids and save redundant development efforts. Integration of the FC functionality into the MDC is currently in progress and can remove the need for a separate service. This implementation of the MDC also provides an additional attribute service for fine-grained access control, which is essential for collaboration-internal configuration sharing during initial embargo periods. Since token-based authentication and authorization, as well as a REST API for MDC and FC are expected to become an ILDG 2.0 standard, a complete re-factoring of the MDC and FC is planned for 2023.
# Outlook
Taking up and evolving the ILDG initiative is a natural, efficient, and timely step to face the data management challenge in Lattice QCD in a way which is fully compliant with FAIR principles and modern data repository standards. The basic concepts of ILDG as a federated system of storage infrastructures and data services are sill and more than ever valid. While some of the components of ILDG and its regional grids are still running after more than a decade of stable operations, e.g. storage elements and the metadata repositories of JLDG and LDG, modernizations and technological updates are necessary and timely. For instance, the transition to token-based authentication and authorization mechanisms is perfectly in line with the ongoing developments of large HEP experiments towards Cloud technologies. Some of the necessary developments in ILDG also need dedicated software expertise, which eventually has become available only thanks to recent funding, e.g. in the context of PUNCH4NFDI. Two crucial elements for the reactivation and modernization of ILDG are the services for the VO-wide user registration and authentication, and the Metadata and File Catalogues. With the completed rebuild and containerization of the MDC and FC of LDG, these can also be deployed in other regional grids. Thus, we expect to achieve fully operational and ILDG-compliant services, at least for JLDG and LDG and possibly other regional grids within the next months. Larger uploads e.g. of configurations from ETMC, HotQCD, and CLS, are then planned, and will serve to improve and test the setup. The volume of the configurations to be eventually uploaded is of the order of several PB. Since neither ILDG nor the regional grids have usually own storage resources, additional storage space may become a concern. It should therefore be emphasized that lattice collaborations also need to apply for additional storage resources at their home institutions or through other sources. Joining efforts in the framework ILDG is likely to be an advantage for this purpose. The further evolution of ILDG as outlined in this contribution, in particular, also development and maintenance of user tools or updated documentation, relies to a large extent on contributions from the user community. Therefore, joint efforts and active participation, e.g. in the working groups is essential and highly welcome. | {'timestamp': '2022-12-19T02:09:46', 'yymm': '2212', 'arxiv_id': '2212.08392', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08392'} |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction
Multifunctional materials are defined as materials that have several interesting and applicable properties. For example, a material can both be a superconductor as well as an insulator, it is just a question of which temperature range the material is operated in. In the same way, the LiFePO\(_4\) (LFPO) compound is both a cathode material for rechargeable batteries in a room temperature regime, while displaying interesting magnetic properties at low temperatures. Originally, the great interest for LFPO is founded in its electrochemical properties, where it is of high interest as a safe and highly stable cathode material in low current conditions (c.f. other Li-ion battery cathode materials). Further, LFPO is a 'Co-free' cathode material where the replacement of Co with Fe is very favorable due to its much higher natural abundance (Fe:Co = 2500:1). Consequently, both lower price as well as less friction on the geopolitical level makes LFPO attractive for industrial scale energy applications. LFPO adopts an olivine type crystal structure with the orthorhombic \(Pnma\) space group (see Fig. [\[LFPO\]](#LFPO){reference-type="ref" reference="LFPO"}). The Li ions are situated in one-dimensional (1D) channels running along the b-axis, which give rise to interesting effects on the lithium intercalation mechanism and dynamics. Such anisotropic intercalation/de-intercalation process during charge/discharge cycles is generally believed to occur topotactically, and thereby create a phase separated cathode material with LFPO and FPO phases, respectively. However, such two-phase separation mechanism has also been called into question by both theoretical as well as experimental studies. At low temperature, LFPO enters into an antiferromagnetic (AFM) phase below \(T_{\rm N}\approx50\) K. It has been shown that a rather complex canted spin order is present (see also Fig. [\[LFPO\]](#LFPO){reference-type="ref" reference="LFPO"}) resulting from competing in-plane and out of plane magnetic couplings (\(J\)). A lot of effort has already been done in order to tune the ion-dynamics (i.e. battery performance) in LFPO. Here, nanostructuring as well as surface coatings (with carbon as well as with other elements and compounds) are among the most common approaches. Such systems have also been studied systematically using both neutron scattering and \(\mu^+\)SR. When it comes to the tuning of low-temperature magnetic properties, much less detailed studies are to be found in the published literature. Indeed, chemical substitution in LFPO is possible, yielding e.g. NaFePO\(_4\), which display a slightly larger unit cell and also an expected reduction in the magnetic ordering temperature (\(T_{\rm N}\)) by a few Kelvin (Fe-Fe distance is larger). Of course, nanostructuring and surface coatings could also induce interesting effects on the magnetic properties, however, other more direct and 'cleaner' approaches could be employed first. In fact, a magnetic field dependence of the AFM magnons in LFPO has recently been presented. Another very common method is to apply hydrostatic pressure, thereby compressing the crystallographic lattice, and investigate how this affects the spin correlations in a material. Combining the penetrating properties of muons, the ultra-high sensitivity of \(\mu^+\)SR to detect changes in spin order/dynamics, along with hydrostatic pressure is a very powerful approach. In this study we have used a similar protocol to conduct the first investigation of the local magnetic properties of LFPO under hydrostatic pressure.
# Experimental Setup
The \(^7\)LiFePO\(_4\) powder sample was prepared using a solid-state reaction technique with \(^7\)Li enriched and reagent grade \(^7\)Li\(_2\)CO\(_3\), Fe(II)C\(_2\)O\(_4\cdot\)`<!-- -->`{=html}2H\(_2\)O, and (NH\(_4\))2HPO\(_4\) as the starting materials. More details on the sample synthesis can be found in Ref. . The neutron powder diffraction (NPD) experiments were performed using the HRPT instrument at the Swiss Spallation Neutron Source (SINQ) of Paul Scherrer Institute (PSI), Switzerland. Approximately 2 grams of sample inside a standard Ø = 6 mm vanadium sample container that was subsequently inserted into a cryofurnace (\(1.5 K\leq{}T\leq550\) K). High resolution data was acquired using the neutron wavelength \(\lambda=1.49\) Å. The NPD patterns were refined using the Fullprof software suite. The \(\mu^+\)SR experiments were performed at \(\mu\)E1 beamline within the S\(\mu\)S muon source of PSI using the GPD instrument. Hydrostatic pressures up to 20 kbar were applied utilizing a double-wall piston cylinder cell (MP35/CuBe alloy). Daphne oil was used as the pressure medium in order to achieve the hydrostatic pressure. Measurements were performed inside a \(^{4}\)He flow cryostat, reaching temperatures down to 2 K. More details about the \(\mu^+\)SR technique and experimental setup can be found in Refs. . Data was analyzed using the open analysis software suite `musrfit`.
# Results & Discussion
Initial neutron powder diffraction (NPD) investigation show that the sample is of high quality. As shown in Fig. [\[NPD\]](#NPD){reference-type="ref" reference="NPD"}, the diffraction pattern acquired at \(T=100\) K is very well refined by the expected \(Pnma\) space group and there are no signs of any impurity peaks. The right panel of Fig. [\[NPD\]](#NPD){reference-type="ref" reference="NPD"} further display the basic crystallographic data obtained from the Rietveld refinement. Weak transverse-field (wTF = 50 G) \(\mu^+\)SR time spectra were collected as a function of temperature for selected pressures (\(p=0, 10, 20\) kbar). Typical wTF time spectra at selected temperatures are shown in Fig. [\[wTFspec\]](#wTFspec){reference-type="ref" reference="wTFspec"} for ambient and \(p=20\) kbar, respectively. A distinct oscillation is observed as well as an offset at lower temperatures. Therefore, wTF data was fitted using a combination of an exponentially relaxing cosine function and an exponentially relaxing signal: \[\begin{aligned}
A_0 \, P_{\rm TF}(t) = A_{\rm TF}\cos(2\pi f_{\rm TF}t + \phi)e^{-\lambda_{\rm TF} t}+A_{\rm S}e^{-\lambda_{\rm S} t}, \label{wtf} \end{aligned}\] where \(A_0\) is the initial asymmetry, which depends on the instrumental configuration, and \(P_{\rm TF}(t)\) is the muon spin depolarisation function under TF. \(A_{\rm TF}\), \(f_{\rm TF}\), \(\phi\) and \(\lambda_{\rm TF}\) are the asymmetry, frequency, phase and relaxation rate originating from the applied TF while \(A_{\rm S}\) and \(\lambda_{\rm S}\) are the asymmetry and relaxation rate originating from the internal magnetic field. Fig. [\[wTFAsym\]](#wTFAsym){reference-type="ref" reference="wTFAsym"} shows the background subtracted and normalized \(A_{\rm TF}\) as a function of temperature for the three different pressures. For all pressures, \(A_{\rm TF}\) exhibits a temperature independent behaviour at low temperatures. A step-like increase is observed for each pressures as the temperature increases. Since \(A_{\rm TF}\) corresponds roughly to the paramagentic fraction, the step like increase is attributed to the magnetic transition. This critical temperature is obtained from a sigmoid fit and is defined as the point in which the paramagnetic fraction \(\simeq\) magnetic fraction function for each pressure \[\(T_{\rm N}(p)\)\]. A clear pressure dependence is observed in \(T_{\rm N}\), which increases with applied hydrostatic pressure. Previous inelastic neutron scattering measurements determined most significant bonds with values: \(J_1=-1.08\) meV, \(J_2=-0.4\) meV and \(J_\perp=0.021\) meV where \(J_1\) and \(J_2\) are in plane interactions along \[001\] and \[010\] while \(J_\perp\) is the perpendicular interaction along \[100\]. Moreover, high pressure XRD at room temperature reported a higher compressibility along the \[100\] direction (a-axis is the longest). Therefore, the increase in \(T_{\rm N}(p)\) with pressure may be attributed to stronger \(J_\perp\) interactions due to shorter bond distance, at least as a first approximation. It is noted that \(J_\perp\) is a ferromagnetic interaction, unlike \(J_1\) and \(J_2\). Therefore, a high pressure AF-FM transition may be present in LFPO at low temperatures. Zero field (ZF) time spectra were collected at base temperature \(T=4\) K for ambient and highest (20 kbar) pressures. The spectra exhibit two distinct oscillations \[see Fig. [\[ZFspec\]](#ZFspec){reference-type="ref" reference="ZFspec"}(a)\], consistent with our previous study. Therefore, they were fitted using a combination of exponentially relaxing cosine functions, an exponential and a exponentially relaxing static gaussian Kubo-Toyabe (\(G^{\rm SGKT}\)): \[\begin{aligned}
A_0 \, P_{\rm ZF}(t) &=& \sum_i^2 A^{\rm AF}_{i}\cos(f^{\rm AF}_{i}t + \phi^{\rm AF}_{i})e^{-\lambda^{\rm AF}_{i} t}+ A_{\rm tail}e^{-\lambda_{\rm tail} t}+ A_{\rm PC}G^{\rm SGKT}e^{-\lambda_{\rm PC} t} \label{zf} \end{aligned}\] where \(A_0\) is the initial asymmetry, which depends on the instrumental configuration, and \(P_{\rm ZF}(t)\) is the muon spin depolarisation function under ZF. \(A^{i}_{\rm i}\), \(f^{\rm AF}_{\rm i}\), \(\phi^{\rm AF}_i\) and \(\lambda^{\rm AF}_{\rm i}\) are the asymmetry, frequency and phase resulting from internal field contributions that are non-parallel to the initial muon spin polarisation. \(A_{\rm tail}\) and \(\lambda_{\rm tail}\) are the asymmetry and relaxation rate of the tail component, that originates from internal field components that are parallel to the initial muon spin polarisation. \(A_{\rm PC}\) and \(\lambda_{\rm PC}\) are the asymmetry and relaxation rates, respectively, associated with the \(G^{\rm SGKT}\) signal, attributed to contributions from the pressure cell. The precession frequencies at ambient pressure, \(f^{\rm AF}_1=45.37\) MHz and \(f^{\rm AF}_2=92.99\) MHz are fully consistent with our previous studies. When a hydrostatic pressure \(p=20\) kbar is applied (i.e. compression of the lattice), the value of both \(f^{\rm AF}_1\) and \(f^{\rm AF}_2\) increases (see Fig. [\[ZFspec\]](#ZFspec){reference-type="ref" reference="ZFspec"}), which is coherent with the TF results showing an increase in \(T_{\rm N}\). Such results are summarized in Fig. [\[pdep\]](#pdep){reference-type="ref" reference="pdep"}(a). An increase in the values of precession frequency, implies an increase in the ordered moment and supports a scenario in which \(J_\perp\) increases in strength with pressure application. This is also very reasonable when considering the NaFePO\(_4\) sister compound mentioned above, which displays a larger Fe-Fe distance, and as a result, slightly lower \(T_{\rm N}\). Several high pressure studies of the atomic structure in LFPO have been published and there is some controversy in the results. An early high pressure XRD investigation reported a structural transition to a space group \(Cmcm\) at a critical pressure below 65 kbar. However, other studies indicate a much higher critical pressure. Further, theoretical calculations based on the \(Cmcm\) structure predict a decrease in the ordered Fe magnetic moment with applied pressure, as shown in Fig. [\[pdep\]](#pdep){reference-type="ref" reference="pdep"}(b). Our experimental result on the contrary show stronger magnetic correlations with applied pressure. Therefore, we may at least conclude that the supposed structural transition is not present up to 20 kbar. Our TF and ZF result also indicate that the spin structure at ambient pressure is most likely preserved up to at least \(p=20\) kbar. We can of course not fully exclude a combination of atomic lattice compression together with a subtle change in the spin canting occurs. However, to discern such in-depth information, it would be necessary to conduct additional and detailed high-pressure neutron diffraction studies. Finally, even if we only have acquired data at three different pressures, we could indeed fit both \(T_{\rm N}(p)\) and \(f^{\rm AF}(p)\) to a straight line. For \(T_{\rm N}(p)\) we can then extrapolate and acquire a critical pressure \(p_c\) for which \(T_{\rm N}=0\) K, i.e. magnetic order in LFPO is fully suppressed. Such procedure yield \(p_c\approx-325\) kbar, i.e. a very large tensile strain (larger lattice). This is clearly a very unreasonable phase space for practically conducting actual experiments, however, it gives an indication and guidance to how LFPO could be chemically tuned in order to suppress magnetic spin order.
# Conclusion
We have performed the first \(\mu^+\)SR investigations of the magnetic properties in LiFePO\(_4\) as a function of hydrostatic pressure and temperature. Measurements in transverse field (TF) and zero field (ZF) confirms the formation of long range antiferromagnetic order within the entire pressure range (\(p=0-20\) kbar). A clear increase in both magnetic ordering temperature (\(T_{\rm N}\)) and ordered Fe magnetic moment is revealed, which we interpret as related to an increase of the out plane interaction strength. Moreover, our results also confirm the absence of any \(Pnma\) to \(Cmcm\) structural transition in LiFePO\(_4\), at least up to \(p=20\) kbar. Finally, we can estimate a critical pressure \(p_{\rm c}\approx-325\) kbar (i.e. tensile strain) for which magnetic spin order in LiFePO\(_4\) is fully suppressed. | {'timestamp': '2022-12-19T02:11:00', 'yymm': '2212', 'arxiv_id': '2212.08439', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08439'} |
null | null |
null | null |
# Introduction
The forced Navier--Stokes equations on the \(3\)-dimensional torus \(\T^3 \simeq \mathbb R^3 / \Z^3\) are given by \[\begin{aligned}
\tag{NS} \label{e:NSE} \partial_t v_{\nu} + v_\nu \cdot \nabla v_\nu + \nabla p_\nu = \nu \Delta v_\nu + F_\nu \\ \diver v_\nu =0, \notag \end{aligned}\] where \(v_\nu: [0, T] \times \T^3 \to \mathbb R^3\) is the velocity field, \(p_\nu: [0, T] \times \T^3 \to \mathbb R\) is the pressure, \(\nu> 0\) is the viscosity parameter and \(F_\nu: [0, T] \times \T^3 \to \mathbb R^3\) is a (divergence-free) force that may depend on \(\nu\). When \(\nu =0\) the Navier--Stokes equations [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"} reduce to the forced Euler equations \[\begin{aligned}
\tag{E} \label{e:E} \partial_t v_0 + v_0 \cdot \nabla v_0 + \nabla p_0 = F_0 \\ \diver v_0 =0. \notag \end{aligned}\] We consider both the Navier--Stokes equations [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"} and the Euler equations [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} with a prescribed initial datum \(v_{\initial}\) which is independent of the viscosity parameter \(\nu\), namely \[\label{e:Cauchy} v_\nu (0, \cdot) = v_{\initial} \,.\] Following we study *smooth* solutions of [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"} (namely \(u_\nu\) and \(F_\nu\) are both \(C^\infty\)), which enjoy uniform in \(\nu\) bounds for \(v_\nu\) and \(F_\nu\) in appropriate function spaces \(X\) and \(Y\). The purpose is to understand which spaces \(X\) and \(Y\) allow for \(u_\nu\) to display *anomalous dissipation*, more precisely whether \[\label{diss_main_Onsager} \limsup_{\nu \downarrow 0} \, \nu \int_0^T \int_{\T^3} | \nabla v_\nu|^2 \, dx\,dt > 0 \,.\] We require that the space \(Y\) rules out anomalous dissipation for solutions of the forced linear Stokes equations under the assumption \(\sup_\nu \|F_\nu\|_Y < \infty\), namely [\[diss_main_Onsager\]](#diss_main_Onsager){reference-type="eqref" reference="diss_main_Onsager"} would not hold if we eliminate the nonlinear advective term \(v_\nu \cdot \nabla v_\nu\) from [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"} and we have uniform bounds for the body forces in the space \(Y\). As it is noticed in \*Section 2 the assumption \[\label{bound_force} \sup_\nu \| F_\nu \|_{L^{1+ \sigma} ([0,1]; C^{\sigma} (\T^3))} <\infty\] for any positive \(\sigma>0\) is in fact sufficient. In the first and fourth authors give examples of smooth solutions \(v_\nu\) to [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"} for which:
- [\[bound_force\]](#bound_force){reference-type="eqref" reference="bound_force"} holds (in fact with the stronger bound \(\sup_\nu \|F_\nu\|_{L^\infty_t (C^{1-\varepsilon})}<\infty\) for any given positive \(\varepsilon\)),
- \(\sup_\nu \|v_\nu\|_{L^\infty} < \infty\),
- and [\[diss_main_Onsager\]](#diss_main_Onsager){reference-type="eqref" reference="diss_main_Onsager"} is satisfied.
In \*Section 2 the authors ask whether this type of behavior is still possible if the uniform \(L^\infty\) bound (ii) is replaced by a uniform bound in some space \(X\) which is close to be "Onsager critical". The Onsager criticality refers to the famous remark by Onsager that if \(\|v\|_{L^\infty (C^{\sfrac{1}{3}+\varepsilon})} < \infty\) and \(u\) solves [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} with \(F=0\), then such solution \(u\) is energy conservative. After a first partial result by Eyink in, the latter was rigorously proved by Constantin, E, and Titi in. It is straightforward to check that, using the arguments in, [\[bound_force\]](#bound_force){reference-type="eqref" reference="bound_force"} and a uniform bound in \(\|v_\nu\|_{L^3 (C^{\sfrac{1}{3}+\varepsilon})}\) is in fact enough to rule out [\[diss_main_Onsager\]](#diss_main_Onsager){reference-type="eqref" reference="diss_main_Onsager"}. Onsager in stated also that the regularity class \(L^\infty_t (C^{\sfrac{1}{3}}_x)\) should in fact be critical, in particular he conjectured the existence of solutions of [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} with \(F=0\) belonging to slightly lower regularity classes of \(L^\infty_t (C^{\sfrac{1}{3}}_x)\) which do not conserve the kinetic energy. After a decade of work in the area which started with [@DLS1; @DLS2], the Onsager conjecture was proved by Isett in (cf. also ) using "convex intgeration methods". While Onsager's conjecture was motivated by the zero-th law of Kolmogorov's fully developed turbulence, which roughly speaking states that [\[diss_main_Onsager\]](#diss_main_Onsager){reference-type="eqref" reference="diss_main_Onsager"} should be a "typical" phenomenon, it seems at the moment very hard to show that at least some of the dissipative solutions of the unforced Euler equations found so far in the literature can actually be approximated by a sequence of regular solutions to the unforced Navier-Stokes. For this reason in the authors suggested to consider the forced versions of both equations. The main result of this paper is to show that indeed [\[diss_main_Onsager\]](#diss_main_Onsager){reference-type="eqref" reference="diss_main_Onsager"} can be achieved for family of solutions \(\{v_\nu\}_{\nu}\) which enjoy a uniform bound in a space which is just below the Onsager-critical \(L^3_t (C^{\sfrac{1}{3}}_x)\), while the corresponding forces \(F_\nu\) also enjoy a bound like [\[bound_force\]](#bound_force){reference-type="eqref" reference="bound_force"} which rules out [\[diss_main_Onsager\]](#diss_main_Onsager){reference-type="eqref" reference="diss_main_Onsager"} for solutions of the linear Stokes equations.
The following open question was also raised in and at present the methods of this work do not seem strong enough to address it.
In view of Remark [\[r:remark-weakly\]](#r:remark-weakly){reference-type="ref" reference="r:remark-weakly"} even producing one such example with force in \(L^1 ((0,2); L^2 (\T^3))\) seems interesting and highly nontrivial.
## Lack of selection principle and non-uniqueness
As in, a byproduct of our techniques is the lack of a selection principle under vanishing viscosity for bounded solutions of the three dimensional forced Euler equations, if the force converges in the vanishing viscosity limit. We say that a weak solution \(v \in L^\infty ((0,T); L^2(\T^3))\) of the forced Euler equations [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} is admissible if \[\begin{aligned}
\label{admissible} \int_{\T^3} |v(x,t)|^2 dx \leq \int_{\T^3} |v_{\initial}(x)|^2 dx + 2 \int_{\T^3} F (x,t) \cdot v(x,t) dx \end{aligned}\] for \(a.e.\) \(t\in (0,T)\). We will show that the problem of uniqueness and vanishing viscosity selection in the class of admissible solutions for [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} is related to having a solution in the space \(L^1_t (W^{1,\infty}_x)\) (this is essentially the threshold for classical "weak-strong" uniqueness results, see e.g. [@W18; @DRIS22]). In particular uniqueness and selection both fail for solutions in \(L^1 ((0,T); C^\alpha (\T^3))\) for any \(\alpha <1\).
If we give up the regularity of the conservative solution \(v_0^{\mbox{cs}}\) it is possible to show nonuniqueness and lack of selection for much smoother forces.
Obviously the following are simple corollaries of the previous theorems.
We remark that, with a totally different method, Vishik in [@VishikI; @VishikII] has produced nonuniqueness examples for the incompressible Euler equations in \(\mathbb R^2\) in vorticity formulation when the solutions have vorticity in \(C ([0,T], L^\infty \cap L^p)\) for any fixed \(p<\infty\), while the curl of the body force belongs to \(L^{1+\sigma} ([0,T], L^p)\) (cf. the lecture notes ). In particular, using classical Calderon-Zygmund estimates, one can easily see that the velocities of these solutions belong to \(C ([0,T], W^{1,p}_{\rm loc})\), while the body forces belong to \(L^{1+\sigma} ([0,T], W^{1,p}_{\rm loc})\). In fact Vishik's techniques have been successfully transposed to even show nonuniqueness of Leray solutions of the forced Navier-Stokes equations *at a fixed positive viscosity \(\nu>0\)*, see. While the nature of the nonuniqueness results in [@VishikI; @VishikII; @ABCDGJK; @ABC] is quite different from the constructions of this paper, they also strongly suggest that all the results of this section are likely to hold for body forces \(\{F_\nu\}\) enjoying uniform bounds in \(L^1 ([0,T], W^{1,p})\) and solutions of [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"} enjoying uniform bounds in \(L^\infty ([0,T], W^{1,p})\). They also suggest that the following question has likely a positive answer.
# Strategy of the proof {#section:heuristics}
We use the same strategy as in [@BDL22; @JY20; @JYo20] and consider a \(2 + \sfrac{1}{2}\)-dimensional Navier-Stokes solution, for which the evolution decouples into a forced \(2d\)-Navier-Stokes system and a scalar advection-diffusion equation. The solution \(v_\nu\) of the forced \(2d\)-Navier-Stokes system is a suitable regularization of the two-dimensional velocity field \(u: [0,2] \times \T^2 \to \mathbb R^2\) constructed in \*Section 4, which is an alternating shear flow, that is, for every \(t \in (0,1)\) we have either \(u(t, x_1, x_2) = (W(t, x_2), 0)\) or \(u(t, x_1, x_2) = (0, W(t, x_1))\). The third component of the \(3d\)-Navier-Stokes solution solves an advection-diffusion equation and will exhibit anomalous dissipation. More precisely, we define the solution of the forced \(3d\)-Navier-Stokes system and the initial condition as \[v_\nu = \begin{pmatrix} u_\nu \\ \tilde \vartheta_\nu \end{pmatrix}\,, \qquad v_{\initial} = \begin{pmatrix} 0 \\ \vartheta_{\initial} \end{pmatrix} \,,\] where \(u_\nu\) is a suitable regularization of \(u\) (to be defined in Section [4](#sec:solution 3d NS){reference-type="ref" reference="sec:solution 3d NS"}) and \(\tilde \vartheta_{\nu}\) solves the advection-diffusion equation with velocity field \(u_\nu\) and initial datum \(\vartheta_{\rm in}\), i.e. \[\begin{aligned}
\begin{cases} \partial_t \tilde \vartheta_\nu + u_\nu \cdot \nabla \tilde \vartheta_\nu = \nu \Delta \tilde \vartheta_\nu, \\ \tilde \vartheta_{\nu} (0, \cdot ) = \vartheta_{\initial} (\cdot )\,. \end{cases} \end{aligned}\] Since \(u_\nu\) is also an alternating shear flow (see Section [4](#sec:solution 3d NS){reference-type="ref" reference="sec:solution 3d NS"}, the nonlinear term \(u_\nu \cdot \nabla u_\nu\) vanishes identically and therefore the velocity field \(v_\nu\) solves the forced \(3d\)-Navier-Stokes system with force \[F_\nu = \begin{pmatrix} \partial_t u_\nu-\nu \Delta u_\nu \\ 0 \end{pmatrix}\,.\] By suitably setting the parameters in the construction of \(u\), we will verify that \[v_\nu \in L^3((0,1); C^\alpha(\T^3))\,, \quad F_\nu \in L^{1+\sigma}((0,1); C^\sigma(\T^3)) \qquad \text{uniformly in \(\nu\),}\] for some \(\sigma>0\), where \(\alpha < 1/3\) is arbitrary. In order to show that \(v_\nu\) exhibits anomalous dissipation, hence concluding the proof of Theorem [\[t_Onsager\]](#t_Onsager){reference-type="ref" reference="t_Onsager"}, we employ \*Theorem A to get \[\label{anomalous} \limsup_{\nu \downarrow 0} 2 \, \nu \int_0^1 \int_{\T^3} |\nabla v_\nu(s,x)|^2\, dx\, ds \ge \limsup_{\nu \downarrow 0} 2 \, \nu \int_0^1 \int_{\T^3} |\nabla \tilde \vartheta_\nu(s,x)|^2\, dx\, ds > 1/2\,.\] To prove that the vanishing viscosity limit does not select a unique solution in the setting of Theorem [\[t_selection2\]](#t_selection2){reference-type="ref" reference="t_selection2"} and Theorem [\[t_selection\]](#t_selection){reference-type="ref" reference="t_selection"} we use the corresponding statement in \*Theorem B which proves lack of selection for solutions of the advection-diffusion equations with velocity field \(u\). More precisely, we prove that the first two components of \(v_\nu\) (namely \(u_\nu\)) strongly converge in \(L^2((0,2) \times \T^3)\) to a unique limit whereas the last component of \(v_\nu\) (namely \(\tilde \vartheta_\nu\)) for a suitable choice of a sequence of viscosity parameters \(\{ \tilde \nu_q \}_{q \in \N}\) exhibits anomalous dissipation [\[anomalous\]](#anomalous){reference-type="eqref" reference="anomalous"} and for another suitable choice of a sequence of viscosity parameters \(\{ \nu_q \}_{q \in \N}\) converges strongly in \(L^2((0,2) \times \T^3)\) to a conservative solution (i.e. the limit satisfies the energy balance [\[eq:conservative\]](#eq:conservative){reference-type="eqref" reference="eq:conservative"} with the first two components of the velocity field).
# Construction and main properties of the \(2d\) velocity field {#section:parameters}
In this section we recall the main properties of the velocity field \(u: [0,1] \times \T^2 \to \mathbb R^2\) constructed in and of the corresponding solution \(\vartheta_\nu: [0,1] \times \T^2 \to \mathbb R\) of the advection-diffusion equation with velocity field \(u\). This velocity field will be used as a building block for the construction of solutions to the forced \(3d\)-Navier-Stokes equations in Theorems [\[t_Onsager\]](#t_Onsager){reference-type="ref" reference="t_Onsager"}, [\[t_selection2\]](#t_selection2){reference-type="ref" reference="t_selection2"} and [\[t_selection\]](#t_selection){reference-type="ref" reference="t_selection"}.
## Choice of the parameters {#ss:para}
Let \(\alpha \in (0,1)\) and \(\beta \in [0,\sfrac{1}{3})\) such that \(\alpha + 2 \beta <1\). We consider parameters \(\epsilon, \delta \in (0,\sfrac14)\) sufficiently small such that
Furthermore we introduce the parameter \(\gamma>0\) as \[\begin{aligned}
\label{d:gamma} \gamma = \frac{3 \beta (1 + 3\epsilon (1 + \delta)) (1 + \delta ) }{1-\delta} + \frac{ \delta}{8} <1 \. \end{aligned}\] Given \(a_0 \in (0,1)\) such that \[\begin{aligned}
\label{c:d_0} a_0^{\epsilon \delta^2} + a_0^{\sfrac{\epsilon \delta}{8} } \leq \frac{1}{20} \,, \end{aligned}\] we define \[\begin{aligned}
\label{d:parameters} a_{q+1} = a_q^{1 + \delta}, \qquad \lambda_q = \frac{1}{2 a_q}. \end{aligned}\]
## Construction of the velocity field
Let us begin by introducing some notation. For any \(f: [0,2] \times \T^2 \to \mathbb R^2\) we denote by \(\supp_{T} (f)\) the temporal support of the function \(f\), namely the projection on the time interval \([0,2]\) of the support of \(f\). The precise definition is \[\supp_{T} (f) := \overline{ \{ t \in [0,2]: \text{ there exists } x \in \T^2 \text{ such that } f(t,x) \neq 0 \}} \,.\] Given \(\{T_q \}_{q \in \N\cup\{-1\}}\), a decreasing sequence of non-negative numbers such that \(T_{-1} = 1\) and \(T_q \downarrow 0\) as \(q\to \infty\), we define the time intervals \[\mathcal{I}_q = [1-T_q, 1-T_{q+1}] \,, \qquad \mathcal{J}_q = [1+T_{q+1}, 1 + T_{q}] \,, \quad \text{for any \(q\in \N \cup \{-1\}\)}\,.\] The results below are taken from.
# Solution of the forced \(3d\)-Navier-Stokes and Euler equations {#sec:solution 3d NS}
Let \(\alpha\), \(\beta\), \(\gamma\), \(\epsilon\), \(\delta\), and \(\{ a_q \}_{q \in \N}\) be as in Section [3.1](#ss:para){reference-type="ref" reference="ss:para"}. We employ the velocity field \(u\) and the initial condition \(\vartheta_{\rm in}\) built in Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"} to produce \((v_\nu, p_\nu, F_\nu)\) a smooth solution to the forced \(3d\)-Navier-Stokes equations [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"}-[\[e:Cauchy\]](#e:Cauchy){reference-type="eqref" reference="e:Cauchy"}. For any \(q\in \N\), we introduce the closed set \(K_q = [0,1-T_q]\cup [1+T_q,2]\) and define \[\label{e:defuq} u_q (t,x ) = u(t,x) \mathbbm{1}_{K_q} (t)\,.\] We observe that \(u_q\) is smooth for any \(q \in \N\). We consider the family of viscosity parameters \({\tilde \nu}_q\) defined in [\[eq: tilde nu\]](#eq: tilde nu){reference-type="eqref" reference="eq: tilde nu"}. For any \(\nu \in (0,a_0^2)\) there exists \(q\in \N\) such that \(\nu \in ( \tilde \nu_{q+1}, \tilde \nu_q]\). Let \(\tilde \vartheta_{\nu}: [0,2] \times \T^3 \to \mathbb R\) be the unique smooth solution to the advection-diffusion equation [\[e:advection-diffusion\]](#e:advection-diffusion){reference-type="eqref" reference="e:advection-diffusion"} with diffusion parameter \(\nu\), initial datum \(\vartheta_{\initial}\), and velocity field \(u_q (t,x )\), i.e. \[\begin{aligned}
\partial_t \tilde{\vartheta}_{\nu} + {u}_{q} \cdot \nabla \tilde{\vartheta}_{\nu} = \nu \Delta \tilde{\vartheta}_{ \nu} \,. \end{aligned}\] We define smooth functions \(F_{\nu}, v_\nu: [0,2] \times \T^3 \to \mathbb R^3\) and \(p_\nu: [0,2] \times \T^3 \to \mathbb R\) as \[\begin{aligned}
F_{\nu}(t,x) & = \begin{pmatrix} \partial_t u_q (t,x)-\nu \Delta u_q (t,x) \\ 0 \end{pmatrix} \\ v_{\nu} (t,x ) & = \begin{pmatrix} u_q(t,x) \\ \tilde{\vartheta}_{\nu} (t,x) \end{pmatrix} \\partial_{\nu} &= 0\,. \end{aligned}\] Finally, we set \[\begin{aligned}
\label{initialdatum_NS} v_{\initial } = \begin{pmatrix} 0 \\ \vartheta_{\initial} \end{pmatrix}\,. \end{aligned}\] Given Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}, the following lemma is immediately checked.
At least formally, we expect \((v_\nu,p_\nu, F_\nu)\) to converge to a solution of the forced \(3d\)-Euler equations [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"}-[\[e:Cauchy\]](#e:Cauchy){reference-type="eqref" reference="e:Cauchy"} when \(\nu\downarrow 0\). We will prove in the next sections that this is the case under suitable assumptions and that \[\begin{aligned}
\label{F0} F_{0}(t,x) & = \begin{pmatrix} \partial_t u (t,x) \\ 0 \end{pmatrix} \\ p_{0} &= 0\,. \end{aligned}\] The following lemma immediately follows from the regularity of \(u\) in \((0,1)\times \T^2\) in Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}.
# Proof of Theorem [\[t_Onsager\]](#t_Onsager){reference-type="ref" reference="t_Onsager"} and Theorem [\[t_selection2\]](#t_selection2){reference-type="ref" reference="t_selection2"} {#section:proofOnsager}
Let \(\alpha \in [0, 1/3)\) be fixed as in Theorem [\[t_Onsager\]](#t_Onsager){reference-type="ref" reference="t_Onsager"} and \(\alpha ' \in [0,1)\) be fixed as in Theorem [\[t_selection2\]](#t_selection2){reference-type="ref" reference="t_selection2"}. Without loss of generality and up to increasing \(\alpha\) or \(\alpha'\), we can assume \(\alpha ' = 3 \alpha\). We fix \(\beta = \alpha\) and choose the parameters \(\epsilon\), \(\delta\), and \(\{ a_q \}_{q \in \N}\) as in Section [3.1](#ss:para){reference-type="ref" reference="ss:para"}. The parameter \(\gamma\) is then determined by [\[d:gamma\]](#d:gamma){reference-type="eqref" reference="d:gamma"}. The viscosity parameter \({\tilde \nu}_q\) has been chosen in [\[eq: tilde nu\]](#eq: tilde nu){reference-type="eqref" reference="eq: tilde nu"}. Let \((v_\nu, p_\nu, F_\nu)\) be the solution to [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"}, with initial datum as in [\[initialdatum_NS\]](#initialdatum_NS){reference-type="eqref" reference="initialdatum_NS"}, built in Section [4](#sec:solution 3d NS){reference-type="ref" reference="sec:solution 3d NS"}. In order to prove Theorem [\[t_Onsager\]](#t_Onsager){reference-type="ref" reference="t_Onsager"} and Theorem [\[t_selection2\]](#t_selection2){reference-type="ref" reference="t_selection2"} we need to show the following facts:
- There exists \(\sigma>0\) such that \[\begin{aligned}
\sup_{\nu \in (0,a_0^2)} \| v_\nu \|_{L^{3} ([0,1]; C^\alpha(\T^3))} + \| v_\nu \|_{L^{\infty } ([0,2] \times \T^3)} + \| F_\nu \|_{L^{1+ \sigma} ([0,2]; C^{\sigma} (\T^3))} < \infty \. \end{aligned}\] Moreover, \(F_\nu \to F_0\) in \(L^{1+\sigma} ([0,2]; C^{\sigma} (\T^3))\).
- Let \(v_0\) be as in Lemma [\[lemma:Euler(0,1)\]](#lemma:Euler(0,1)){reference-type="ref" reference="lemma:Euler(0,1)"}. We have that \(v_\nu \to v_0\) in \(L^2((0,1)\times \T^3)\) as \(\nu \to 0\).
- There exist \(v_0^{\rm ds}\in L^\infty([0,2]\times \T^3)\) solution to [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} with initial datum [\[initialdatum_NS\]](#initialdatum_NS){reference-type="eqref" reference="initialdatum_NS"} and a sequence \(q_k\to \infty\), such that \(v_{\tilde \nu_{q_k}} \to v_0^{\rm ds}\) weakly in \(L^2([0,2]\times \T^3)\). Moreover, \[2 \, \tilde \nu_{q_k} \int_0^1 \int_{\T^3} | \nabla v_{\tilde \nu_{q_k}}|^2 \, dx\,dt > 1/2 \qquad \text{for any \(k\in\N\).}\] In particular \(v_0^{\rm ds}\) is an admissible dissipative solution of [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"}.
- Set \(\nu_q =a_q^{2-\gamma + \delta + 8 \epsilon}\). There exists \(v_0^{\rm cs}\in L^1((0,2); C^{\alpha'}(\T^3))\), a conservative (admissible) solution to [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} with initial datum [\[initialdatum_NS\]](#initialdatum_NS){reference-type="eqref" reference="initialdatum_NS"}, such that \(v_{\nu_q} \to v_0^{\rm cs}\) strongly in \(L^2((0,2)\times \T^3)\) as \(q \to \infty\).
From [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"} and the maximum principle for the advection-diffusion equation (using that the initial datum is bounded), we deduce that \[\sup_{\nu \in (0,a_0^2)} \| v_\nu \|_{L^{\infty } ([0,2] \times \T^3)} < \infty \,.\] Let us now check that \[\begin{aligned}
\sup_{\nu \in (0,a_0^2)} \| v_\nu \|_{L^{3} ([0,1]; C^\alpha(\T^3))} < \infty \,. \end{aligned}\] This is a consequence of \[\label{z5} u \in L^3 ((0,1); C^{\alpha}(\T^3)) \qquad\text{ and }\qquad \sup_{\nu \in (0,a_0^2)} \| \vartheta_\nu \|_{L^{3} ((0,1); C^\alpha(\T^2))} < \infty\] that we now prove. Indeed, \(\tilde \vartheta_\nu \equiv \vartheta_\nu\) in \([0, 1-T_q] \times \T^3\) since \(u_q \equiv u\) in \([0, 1-T_q] \times \T^3\), while \(\tilde \vartheta_\nu(t,\cdot)\) solves the heat equation for \(t \in [1-T_q, 1]\), and the Hölder norm is nonincreasing for solutions of the heat equation. Let us begin by proving the first property in [\[z5\]](#z5){reference-type="eqref" reference="z5"}. By [\[p:support\]](#p:support){reference-type="eqref" reference="p:support"} and [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"} and by interpolation we have \[\begin{aligned}
\| u \|_{L^3((0,1);C^\alpha)}^3 & = \sum_{q=0}^\infty \int_{{\mathcal{I}}_q} \|u (s,\cdot)\|_{C^\alpha(\T^2)}^{3} ds \leq \sum_{q=0}^\infty \int_{{\mathcal{I}}_q} \|u (s,\cdot)\|_{L^\infty(\T^2)}^{3 (1-\alpha)} \|u (s,\cdot)\|_{W^{1,\infty}(\T^2)}^{3 \alpha} ds \\ & \le C\sum_{q=0}^\infty a_q^{ \gamma} a_q^{-3(1-\alpha)(\gamma-1) } a_q^{-3 \alpha (\gamma-1) } a_{q+1}^{-3 \alpha (1 + \epsilon \delta)} \end{aligned}\] and the sum is finite if and only if \[\frac{\gamma}{3} + 1-\gamma-\alpha (1 + \epsilon \delta) (1 + \delta) >0\,,\] which holds thanks to the choice [\[d:gamma\]](#d:gamma){reference-type="eqref" reference="d:gamma"} and the condition [\[c:alpha_beta_eps_kappa\]](#c:alpha_beta_eps_kappa){reference-type="eqref" reference="c:alpha_beta_eps_kappa"}. Let us show the second property in [\[z5\]](#z5){reference-type="eqref" reference="z5"}. Fix \(\nu \in (0,a_0^2)\) and correspondingly let \(q\in \N\) such that \(\nu \in ( \tilde \nu_{q+1}, \tilde \nu_q]\). Thanks to property [\[prop:regularity-solution\]](#prop:regularity-solution){reference-type="eqref" reference="prop:regularity-solution"} of Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"} and using \(\alpha = \beta\), we get \[\begin{aligned}
\| \vartheta_\nu \|_{L^{3 }((0,1); C^\alpha)}^{3} &= \sum_{q=0}^\infty \int_{{\mathcal{I}}_q} \| \vartheta_\nu (s, \cdot ) \|_{ C^\beta}^{3} ds \le C\sum_{q=0}^{\infty} a_q^{ \gamma-\gamma \delta} a_{q+1}^{-3( \beta + 3\beta \epsilon (1+ \delta))} \\ & = C\sum_{q=0}^{\infty} a_q^{ \gamma-\gamma \delta-3( \beta + 3\beta \epsilon (1+ \delta))(1 + \delta)} \end{aligned}\] and the sum is finite and independent of \(\nu\) since as a consequence of [\[d:gamma\]](#d:gamma){reference-type="eqref" reference="d:gamma"} we have \[-\gamma (1-\delta) + 3( \beta + 3\beta \epsilon (1+ \delta))(1 + \delta) <0 \.\] We finally prove that \[F_\nu \in L^{1+\sigma}((0,2); C^\sigma(\T^3))\,, \quad \text{uniformly in \(\nu\in (0,a_0^2)\)}\,,\] for some \(\sigma>0\), and \(F_\nu \to F_0\) in \(L^{1+\sigma}((0,2); C^\sigma(\T^3))\) as \(\nu \to 0\). To this aim, it is enough to show that there exists \(C>0\) such that for any \(\nu \in (\tilde \nu_{q+1}, \tilde \nu_q]\) we have \[\label{e:regforce1} \| \partial_t u \|_{L^{1+ \sigma}((0,2); C^\sigma ( \T^3))} \leq C \qquad \text{ and } \qquad \| \nu \Delta u \|_{L^{1+ \sigma}((K_q; C^\sigma( \T^3))} \leq C a_q^\epsilon \, \,.\] and that \[\label{e:zeroforce1} \| \partial_t u \|_{L^{1+ \sigma}(K_q^c; C^\sigma ( \T^3))} \to 0 \qquad \text{ as \(q \to \infty\).}\] For the first property in [\[e:regforce1\]](#e:regforce1){reference-type="eqref" reference="e:regforce1"}, thanks to [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"} we have \[\begin{aligned}
\| \partial_t u \|_{L^{1+ \sigma}((0,2); C^\sigma( \T^3))}^{ 1+ \sigma} & \leq \sum_{j=0}^{\infty} \int_{\mathcal{I}_j\cup \mathcal{J}_j} \left ( \| \partial_t u (s, \cdot) \|_{L^\infty (\T^3)}^{(1-\sigma)} \| \partial_t u (s, \cdot) \|_{W^{1,\infty} (\T^3)}^{\sigma} \right )^{1+ \sigma} ds \\ & \leq C \sum_{j=0}^{\infty} \int_{\mathcal{I}_j\cup \mathcal{J}_j} \left ( a_j^{1- 2 \gamma } a_{j+1}^{-\sigma (1+ \epsilon \delta)} \right )^{1+ \sigma} ds \\& \leq 4 C \sum_{j=0}^{\infty} a_j^{\gamma} a_j^{(1+ \sigma)(1-2 \gamma-\sigma (1+ \delta) (1+ \epsilon \delta )) } < \infty \end{aligned}\] where we used that \(1-\gamma >0\), and we choose \(\sigma>0\) sufficiently small such that \(\gamma + (1+ \sigma)(1- 2 \gamma-\sigma (1+ \delta) (1 + \epsilon \delta ) )>0\). Property [\[e:zeroforce1\]](#e:zeroforce1){reference-type="eqref" reference="e:zeroforce1"} follows by noticing that \[\| \partial_t u \|_{L^{1+ \sigma}(K_q^c; C^\sigma( \T^3))}^{ 1+ \sigma} \leq \sum_{j=q}^{\infty} \int_{\mathcal{I}_j\cup \mathcal{J}_j} \left ( \| \partial_t u (s, \cdot) \|_{L^\infty (\T^3)}^{(1-\sigma)} \| \partial_t u (s, \cdot) \|_{C^1 (\T^3)}^{\sigma} \right )^{1+ \sigma} ds \quad \to \quad 0 \qquad \text{ as \(q \to \infty\).}\] For the second property in [\[e:regforce1\]](#e:regforce1){reference-type="eqref" reference="e:regforce1"}, thanks to [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"}, we have \[\begin{aligned}
\| \nu \Delta u \|_{L^{1+ \sigma}(K_q; C^\sigma( \T^3))}^{1+ \sigma} & \leq \tilde \nu_q \sum_{j=0}^{q-1} \int_{\mathcal{I}_j\cup \mathcal{J}_j} \left ( \| \Delta u (s, \cdot) \|_{L^\infty (\T^3)}^{1-\sigma } \| \Delta u (s, \cdot) \|_{W^{1,\infty} (\T^3)}^{ \sigma } \right )^{1+ \sigma} ds \\ & \leq C a_q^{2-\frac{\gamma}{1 + \delta} + 4 \epsilon} \sum_{j=0}^{q-1} \int_{\mathcal{I}_j\cup \mathcal{J}_j} \left ( a_j^{1-\gamma} a_{j+1}^{-2 (1 + \epsilon \delta) (1-\sigma) } a_{j+1}^{-3 (1 + \epsilon \delta) \sigma } \right)^{1+ \sigma} ds \\ & \leq C a_q^{2-\frac{\gamma}{1 + \delta} + 2 \epsilon} \sum_{j=0}^{q-1} a_j^\gamma \left ( a_j^{-1-2 \delta-\gamma} a_j^{-\sigma (1+ \epsilon \delta )(1+ \delta) } \right )^{1 + \sigma} \\ & \leq C q a_q^{2 \epsilon} a_{q-1}^{2 + 2 \delta-\gamma} a_{q-1}^{\gamma} a_{q-1}^{-(1+\sigma)(1 + 2 \delta + \gamma + \sigma (1+ \epsilon \delta )(1+ \delta) )} \leq C a_q^{\epsilon}, \end{aligned}\] where we used that \(q a_q^{\epsilon} \leq 1\), \(a_{j+1} = a_j^{1+ \delta}\), \(1-\gamma >0\) and we choose \(\sigma >0\) sufficiently small to guarantee that \(2 + 2 \delta-(1+ \sigma)( 1 + 2 \delta + \gamma + \sigma (1+ \epsilon \delta )(1+ \delta) ) >0\). Recalling [\[e:defuq\]](#e:defuq){reference-type="eqref" reference="e:defuq"}, it suffices to prove that \(\tilde{\vartheta}_{\nu} \to \vartheta_0\) in \(L^2((0,1) \times \T^3)\), as \(\nu \to 0\). Fix \(\nu \in (0,a_0^2)\), and let \(q\in \N\) such that \(\nu \in (\tilde \nu_{q+1}, \tilde \nu_q]\). We employ a standard vanishing viscosity estimate. For any \(0 \le t\le 1-T_q\), we have that \(u_q = u\), hence \[\partial_t (\tilde{\vartheta}_{\nu}-\vartheta_0) + u \cdot \nabla (\tilde{\vartheta}_{\nu}-\vartheta_0) = \nu \Delta \tilde{\vartheta}_\nu \, \qquad \text{ for any } 0 \leq t \leq 1-T_q.\] We multiply the above equation by \(\tilde{\vartheta}_{\nu}-\vartheta_0\) and integrate in space-time to get \[\begin{aligned}
\| \tilde{\vartheta}_{\nu} (t, \cdot)-\vartheta_0 (t, \cdot) \|_{L^2 (\T^3)}^2 & \leq \nu \left | \int_0^t \int_{\T^3} \nabla \tilde{\vartheta}_\nu (s, x) \cdot \nabla \vartheta_0 (s,x) dx ds \right | \\ & \leq \left( \nu \int_0^t \int_{\T^3} \left | \nabla \tilde{\vartheta}_\nu (s, x) \right |^2 dx ds \right )^{1/2} \left( \nu \int_0^t \int_{\T^3} \left | \nabla \vartheta_0 (s, x) \right |^2 dx ds \right )^{1/2}\,. \end{aligned}\] We observe that by the energy equality \[\label{z2} \nu \int_0^t \int_{\T^3} \left | \nabla \tilde{\vartheta}_\nu (s, x) \right |^2 dx ds \leq 1 \qquad \text{for any \(t\in [0,1]\)}\,.\] Let us define \(t(\nu):= 1-T_{k(q)} \le 1-T_q\), where \(k(q)\) is the largest natural number satisfying \[a_q^{2-\frac{\gamma}{1+ \delta}} \exp \left( a_{k(q)}^{2-2\gamma} a_{k(q) + 1}^{-2-2\epsilon\delta } \right) \leq 1\,.\] We claim that \[\label{z3} \nu \int_0^{t(\nu)} \int_{\T^3} \left | \nabla \vartheta_0 (s, x) \right |^2 dx ds \to 0 \qquad \text{ and } \qquad t(\nu) \to 1 \qquad \mbox{ as } \nu \to 0\,.\] This follows by Grönwall inequality and [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"}, since \[\begin{aligned}
\nu \int_0^{t(\nu)} \| \nabla \vartheta_0 (s, \cdot) \|_{L^\infty (\T^3)}^2 ds & \leq \nu \| \nabla \vartheta_{\rm in} (\cdot)\|_{L^\infty(\T^3)}^2 \exp \left (\int_0^{t(\nu)} \| \nabla u (s, \cdot ) \|_{L^\infty}^2 \right ) ds \\ & \leq C a_q^{2-\frac{\gamma}{1+ \delta} + 4 \epsilon} a_0^{-2-2\epsilon \delta} \exp \left( a_{k(q)}^{2-2\gamma} a_{k(q) + 1}^{-2-2 \epsilon\delta }\right) \\ & \leq C a_q^{4 \epsilon} a_0^{-2-2\epsilon \delta} \to 0 \end{aligned}\] as \(q \to \infty\). Finally, \(t(\nu)\to 1\) as \(\nu \to 0\) follows by the fact that \(k(q)\to \infty\) as \(q\to \infty\). Building upon [\[z2\]](#z2){reference-type="eqref" reference="z2"}, [\[z3\]](#z3){reference-type="eqref" reference="z3"}, and the fact that \(\tilde{\vartheta}_\nu\) and \(\vartheta_0\) are uniformly bounded by \(1\), we deduce \[\|\tilde{\vartheta}_{\nu}-\vartheta_0\|_{L^2 ((0,1) \times \T^3)}^2 \le \nu \int_0^{t(\nu)} \| \nabla \vartheta_0 (s, \cdot) \|_{L^\infty (\T^3)}^2 ds + C(1-t(\nu)) \to 0\] as \(q\to \infty\). [\[subsec:ThAC(iii)\]]{#subsec:ThAC(iii) label="subsec:ThAC(iii)"} We observe that the sequence of solutions \(\vartheta_{\tilde \nu_q}\) of the advection-diffusion equation with diffusion parameter \(\tilde \nu_q\), velocity field \(u\), and initial datum \(\vartheta_{\initial}\) satisfies \[\label{z6} 2 \, \tilde \nu_q \int_0^{1-T_q + \bar t_q} \| \nabla \vartheta_{\tilde \nu_q}(s,\cdot)\|_{L^2}^2 \, ds > \frac{1}{2} \qquad \text{for any \(q \in m\N\),}\] as a direct consequence of [\[prop:anomalous\]](#prop:anomalous){reference-type="eqref" reference="prop:anomalous"} in Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}. Therefore \(\tilde{\vartheta}_{\tilde \nu_q}\), the third component of \(v_{\tilde \nu_q}\), satisfies [\[z6\]](#z6){reference-type="eqref" reference="z6"} as well since \(\vartheta_{\tilde \nu_q} = \tilde \vartheta_{\tilde \nu_q}\) in \([0,1-T_q + \bar t_q]\). The first two components of \(v_\nu\) strongly converge to \(u\) in \(L^{\infty} ((0,2 ) \times \T^3)\) since \(\| {u} \|_{L^\infty (K_q^c \times \T^3)} \leq 2 a_q^{1-\gamma} \to 0\) as \(q \to \infty\). It is simple to see that \(\{\tilde \vartheta_{\tilde \nu_q}\}_{q\in \N}\) admits limit points in the weak topology of \(L^2((0,2)\times \T^2)\) and that any such limit point solves the transport equation with velocity field \(u\) and initial datum \(\vartheta_{\rm in}\). Let us fix a limit point and denote it by \(\vartheta^{\rm ds}\). It follows by [\[z6\]](#z6){reference-type="eqref" reference="z6"} that \(\vartheta^{\rm ds}\) is a dissipative solution of the transport equation. We define \[v_{0}^{\rm ds} (t,x ) := \begin{pmatrix} u (t,x) \\ \vartheta^{\rm ds} (t,x) \end{pmatrix} \,, \quad t\in (0,2), \, x\in \T^3\,,\] and check that \((v_0^{\rm ds}, p_0, F_0)\) with \(p_0=0\) solves [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"}. Indeed, since the first two components of \(v_{\tilde \nu_q}\) strongly converge to \(u\) in \(L^{\infty} ((0,2 ) \times \T^3)\) and the last component converges weakly to \(\vartheta^{\rm ds}\), the quadratic term \(v_{\tilde \nu_q} \cdot \nabla v_{\tilde \nu_q}\) converges in the sense of distributions to \(v_0^{\rm ds} \cdot \nabla v_0^{\rm ds}\). It is straightforward to check that all the other terms in the distributional formulation of [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} pass to the limit as \(\tilde \nu_q \to 0\). Finally the admissibility condition [\[admissible\]](#admissible){reference-type="eqref" reference="admissible"} of \(v_0^{\rm ds}\) follows from the fact that it is a weak\* limit in \(L^\infty\) of admissible solutions \(v_\nu\) with force \(F_\nu\) and the forces \(F_\nu\) are strongly converging to \(F_0\) in \(L^1\). Let \(\nu_q =a_q^{2-\gamma + \delta + 8 \epsilon} \in (\tilde \nu_{q+1}, \tilde \nu_{q}]\). As before we have \[v_{\nu_q} (t,x) = \begin{pmatrix} u (t,x) \mathbbm{1}_{K_q} (t) \\ \tilde{\vartheta}_{\nu_q} (t,x) \end{pmatrix} \.\] Recalling the proof of (iii), we only need to prove that the last component of \(v_{\nu_q}\) strongly converges in \(L^2((0,2) \times \T^3)\) to a velocity field \(v_0^{\rm cs}\in L^1((0,2); C^{\alpha'}(\T^3))\) that conserves in time the spatial \(L^2\) norm and the admissibility condition [\[admissible\]](#admissible){reference-type="eqref" reference="admissible"} will directly follow from the conservative property. We show that \(\tilde \vartheta_{\nu_q} \to \vartheta_{0}\) in \(L^2((0,2) \times \T^3)\), where \(\vartheta_0\) is the symmetric solution to the transport equation in Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}(4). To this aim, it is enough to show that \(\| \tilde \vartheta_{\nu_q} (t, \cdot)-\vartheta_{q} (t, \cdot )\|_{L^2(\T^2)} \to 0\) as \(q \to \infty\) for any \(t \in (0,2)\), where \(\vartheta_q\) is the unique solution of the transport equation with velocity field \(u_q\) and initial datum \(\vartheta_{\rm in}\). Indeed, this will entail \[\| \tilde \vartheta_{\nu_q}-\vartheta_0 \|_{L^2 ([0,2] \times \T^2)} \leq \| \tilde \vartheta_{\nu_q}-\vartheta_q \|_{L^2([0,2] \times \T^2)} + \| \vartheta_{q}-\vartheta_0 \|_{L^2([0,2] \times \T^2)} \to 0 \qquad \mbox{ as } q \to \infty \,,\] where the second term \(\| \vartheta_{q}-\vartheta_0 \|_{L^2 ((0,2) \times \T^3)} \to 0\) as \(q \to \infty\), thanks to \(\vartheta_q (t,\cdot)= \vartheta_0 (t,\cdot)\) for any \(t \in [1-T_q, 1+T_q ]^c\), and the \(L^\infty\) bound \(\| \vartheta_0 \|_{L^\infty ((0,2) \times \T^2)} + \| \vartheta_q \|_{L^\infty ((0,2) \times \T^2)} \leq 2\). For any \(t \in (0,2)\), using a standard energy estimate with the regularity bound [\[prop:regularity-solution\]](#prop:regularity-solution){reference-type="eqref" reference="prop:regularity-solution"} and the symmetry property [\[vectorfield_VV\]](#vectorfield_VV){reference-type="eqref" reference="vectorfield_VV"} from Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}, we estimate \[\begin{aligned}
&\| \tilde \vartheta_{\nu_q} (t, \cdot)-\vartheta_q (t, \cdot ) \|_{L^2(\T^2)}^2 \\ \leq & 2 \nu_q \left | \int_0^t \int_{\T^2} \nabla \tilde \vartheta_{\nu_q} (s,x) \cdot \nabla \vartheta_{q}(s,x) dx ds \right | \\ \leq & 2 \left ( \nu_q\int_0^t \int_{\T^2} | \nabla \vartheta_{q} (s,x)|^2 dx ds \right )^{1/2} \\ = & 2 \left (\nu_q \sum_{j = q}^\infty \int_{\mathcal{I}_j\cup \mathcal{J}_j} \int_{\T^2} | \nabla \vartheta_{q} (s,x)|^2 dx ds + \nu_q \sum_{j = 0}^{q-1}\int_{\mathcal{I}_j\cup \mathcal{J}_j} \int_{\T^2} | \nabla \vartheta_{q} (s,x)|^2 dx ds \right )^{1/2} \\ \le & C \left( a_q^{2-\gamma + \delta + 8\epsilon} a_q^\gamma a_q^{-2(1+3\epsilon(1+\delta))} + a_q^{2-\gamma + \delta + 8\epsilon} \sum_{j=0}^{q-1} a_j^{\gamma} a_{j+1}^{-2(1+3\epsilon(1+\delta))} \right)^{1/2} \\ & \leq C a_q^{1-\frac{\gamma}{2} + \frac{\delta}{2} + 4 \epsilon } a_{q-1}^{\gamma/2} a_q^{-1-3 \epsilon (1+\delta)} \leq C a_q^{\frac{\delta}{2}-\frac{\delta \gamma}{2}} \to 0\,, \end{aligned}\] as \(q \to \infty\), where we used \(\gamma < 1\), \(a_{q} = a_{q-1}^{1+ \delta}\), and \(\delta \in (0,\sfrac{1}{8})\). We finally show that \(v_0^{\rm cs}\in L^1((0,2); C^{\alpha'}(\T^3))\). Using [\[p:support\]](#p:support){reference-type="eqref" reference="p:support"} and [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"} we deduce \[\begin{aligned}
\| u \|_{L^1((0,2); C^{\alpha ' })} & = 2 \sum_{q=1}^\infty \int_{{\mathcal{I}}_q} \|u (s,\cdot)\|_{C^{\alpha ' }(\T^2)} ds \leq 2 \sum_{q=1}^\infty \int_{{\mathcal{I}}_q} \|u (s,\cdot)\|_{L^\infty(\T^2)}^{1-\alpha ' } \|u (s,\cdot)\|_{W^{1,\infty}(\T^2)}^{ \alpha ' } ds \\ & \le C\sum_{q=0}^\infty a_q^{ \gamma} a_q^{ 1-\gamma } a_{q+1}^{-\alpha ' (1 + \epsilon \delta)} = C \sum_{q=0}^\infty a_{q}^{1-\alpha ' (1 + \epsilon \delta) (1+ \delta)} < \infty \end{aligned}\] recalling that \(\alpha ' = 3 \alpha\) and \(\alpha = \beta\), and the last inequality holds thanks to the condition [\[c:alpha_beta_eps_kappa\]](#c:alpha_beta_eps_kappa){reference-type="eqref" reference="c:alpha_beta_eps_kappa"}, which implies \(1-3 \alpha (1 + \epsilon \delta) (1+ \delta) >0\). For the last component of \(v_{0}^{\mbox{cs}}\), namely \(\vartheta_0\), we recall that \[\vartheta_0 (t,x) = \vartheta_0 (2-t, x) \qquad \mbox{for any } x \in \T^3 \mbox{ and } t \in (1,2]\,,\] and that it solves the transport equation (namely [\[e:advection-diffusion\]](#e:advection-diffusion){reference-type="eqref" reference="e:advection-diffusion"} with \(\nu =0\)) with velocity field \(u\). Therefore, it is sufficient to estimate \(\vartheta_0\) in \([0,1] \times \T^2\). Using [\[prop:regularity-solution\]](#prop:regularity-solution){reference-type="eqref" reference="prop:regularity-solution"} in Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"} we have \[\begin{aligned}
\| \vartheta_0 \|_{L^1((0,1); C^{\alpha ' })} &= \sum_{q=0}^\infty \int_{{\mathcal{I}}_q} \| \vartheta_0 (s, \cdot ) \|_{ C^{\alpha '}} ds \leq 4 \| \nabla \vartheta_{\initial} \|_{L^\infty} \sum_{q=0}^{\infty} a_q^{ \gamma-\gamma \delta} a_{q+1}^{-\alpha ' (1+ 3 \epsilon (1+ \delta))} \\ & =4 \| \nabla \vartheta_{\initial} \|_{L^\infty} \sum_{q=0}^{\infty} a_q^{ \gamma-\gamma \delta-\alpha ' (1+ 3\alpha ' \epsilon (1+ \delta))(1 + \delta)} < \infty \end{aligned}\] where the last estimate holds thanks to \(\alpha= \alpha ' /3\), [\[c:alpha_beta_eps_kappa\]](#c:alpha_beta_eps_kappa){reference-type="eqref" reference="c:alpha_beta_eps_kappa"}, [\[d:gamma\]](#d:gamma){reference-type="eqref" reference="d:gamma"}, and \(\epsilon < \frac{\delta}{16 (1+ \delta)^2 }\) (a consequence of [\[c:eps_delta\]](#c:eps_delta){reference-type="eqref" reference="c:eps_delta"}).
# Proof of Theorem [\[t_selection\]](#t_selection){reference-type="ref" reference="t_selection"} {#section:proof_selection}
Let \(\alpha ' \in [0,1)\) be as in Theorem [\[t_selection\]](#t_selection){reference-type="ref" reference="t_selection"}. We fix \(\alpha = \alpha '\) and \(\beta =0\) and we choose the parameters \(\delta\), \(\epsilon\), \(\gamma\), and \(\{ a_q \}_{q \in \N}\) as in Section [3.1](#ss:para){reference-type="ref" reference="ss:para"}. The parameters satisfy [\[c:beta_eps_condition_all\]](#c:beta_eps_condition_all){reference-type="eqref" reference="c:beta_eps_condition_all"}, [\[d:gamma\]](#d:gamma){reference-type="eqref" reference="d:gamma"}, [\[c:d_0\]](#c:d_0){reference-type="eqref" reference="c:d_0"}, and the further condition \[\begin{aligned}
\label{c:add_b} 1-\alpha ' (1 + \epsilon \delta)(1 + \delta)-\frac{\delta}{4 } >0 \, \end{aligned}\] which is compatible with all the other conditions. Let \((v_\nu, p_\nu, F_\nu)\) be the solution to [\[e:NSE\]](#e:NSE){reference-type="eqref" reference="e:NSE"}, with initial datum as in [\[initialdatum_NS\]](#initialdatum_NS){reference-type="eqref" reference="initialdatum_NS"}, built in Section [4](#sec:solution 3d NS){reference-type="ref" reference="sec:solution 3d NS"}. In order to prove Theorem [\[t_selection\]](#t_selection){reference-type="ref" reference="t_selection"} we need to show the following facts:
- There holds \[\sup_{\nu \in (0,a_0^2)} \| v_\nu \|_{L^{\infty } ([0,2] \times \T^3)} + \| F_\nu \|_{C^{\alpha'}((0,2)\times \T^3)} < \infty \,.\] Moreover, \(F_\nu \to F_0\) in \(C^{\alpha'}((0,2)\times \T^3)\).
- There exist \(v_0^{\rm ds}\in L^\infty([0,2]\times \T^3)\) solution to [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} with initial datum [\[initialdatum_NS\]](#initialdatum_NS){reference-type="eqref" reference="initialdatum_NS"} and a sequence \(q_k\to \infty\), such that \(v_{\tilde \nu_{q_k}} \rightharpoonup v_0^{\rm ds}\) weakly in \(L^2([0,2]\times \T^3)\). Moreover, \[2 \, \tilde \nu_{q_k} \int_0^1 \int_{\T^3} | \nabla v_{\tilde \nu_{q_k}}|^2 \, dx\,dt \geq 1/2 \, \qquad \text{for any } q_k.\] In particular \(v_0^{\rm ds}\) is an admissible dissipative solution of [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"}.
- Set \(\nu_q =a_q^{2 + 3 \epsilon}\). There exists \(v_0^{\rm cs}\in L^\infty((0,2) \times \T^3)\), an (admissible) conservative solution to [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} with initiald datum [\[initialdatum_NS\]](#initialdatum_NS){reference-type="eqref" reference="initialdatum_NS"}, such that \(v_{\nu_q} \to v_0^{\rm cs}\) in \(L^2((0,2)\times \T^3)\).
Since \(u\) is bounded, more precisely \(\| u (t, \cdot) \|_{L^\infty ((0,2) \times \T^3)} \leq 2 a_0^{1-\gamma} \leq 1\) and \(\| \tilde \vartheta_\nu \|_{L^\infty ((0,2) \times \T^3)} \leq \| \vartheta_{\initial} \|_{L^\infty (\T^3)} = 1\) by the maximum principle, we have \[\| v_\nu \|_{L^\infty((0,2) \times \T^3) } \leq 1\,.\] Let us now show the uniform-in-viscosity regularity of \(F_\nu\). If suffices to prove that there exists \(C>0\) such that for any \(\nu \in (\tilde \nu_{q+1}, \tilde \nu_q]\) we have \[\label{e:regforce2} \| \partial_t u \|_{C^{\alpha '}((0,2) \times \T^3 )} \leq C \qquad\text{ and } \qquad \| \nu \Delta u \|_{C^\alpha(K_q \times \T^3 )} \leq C a_q^\epsilon \.\] We estimate the first term. Thanks to [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"} and the interpolation inequality, we have \[\begin{aligned}
\| \partial_t u \|_{C^{\alpha '} ((0,2) \times \T^3)} & \leq \sup_{j \in \N} \| \partial_t u \|_{L^\infty (\mathcal{I}_j; C^{\alpha'}( \T^3))} + \sup_{j \in \N} \| \partial_t u \|_{L^\infty (\T^3; W^{1,\infty}( \mathcal{I}_j ))} \\ & \leq C \left ( \sup_{j \in \N} \| \partial_t u \|^{1-\alpha' }_{L^\infty (\mathcal{I}_j; L^\infty( \T^3))} \| \partial_t u \|^{ \alpha' }_{L^\infty (\mathcal{I}_j; W^{1,\infty}( \T^3))} + \sup_{j \in \N} a_j^{1-2 \gamma} \right) \\ & \leq C \sup_{j \in \N} a_j^{1-2 \gamma} a_{j+1}^{-\alpha ' (1+ \epsilon \delta)} +1 < \infty \end{aligned}\] where we used [\[c:add_b\]](#c:add_b){reference-type="eqref" reference="c:add_b"} and \(\gamma = \sfrac{\delta}{8} < \sfrac{1}{2}\). This proves the first property in [\[e:regforce2\]](#e:regforce2){reference-type="eqref" reference="e:regforce2"}. In order to show the second property in [\[e:regforce2\]](#e:regforce2){reference-type="eqref" reference="e:regforce2"}, we exploit [\[s:vectorfield_I\_q5_new\]](#s:vectorfield_I_q5_new){reference-type="eqref" reference="s:vectorfield_I_q5_new"} and tha fact that \(\nu \in ( \tilde \nu_{q+1}, \tilde \nu_q]\) to obtain \[\begin{aligned}
\| \nu \Delta u \|_{C^{\alpha'}((0,1-T_q) \times \T^3))} & \leq \tilde \nu_q \sup_{j \leq q-1} \| \Delta u \|_{C^{\alpha '} ( \mathcal{I}_j \times \T^3 )} \\ & \leq C a_q^{2-\frac{\gamma}{1 + \delta} + 4 \epsilon} \sup_{j \leq q-1} ( a_{j}^{1-\gamma} a_{j+1}^{-2-2 \epsilon \delta} a_{j+1}^{-\alpha ' (1+ \epsilon \delta)} + a_{j}^{1-\gamma} a_{j+1}^{-2-2 \epsilon \delta} a_j^{-\gamma}) \\ & \leq C a_q^{ \epsilon} a_{q-1 }^{ 1-2 \gamma-\alpha ' (1 + \epsilon \delta)(1+ \delta)} \leq C a_q^\epsilon \,, \end{aligned}\] where we also used \(a_q^{2 \epsilon} a_q^{-2 \epsilon \delta } \leq 1\), [\[c:alpha_beta_eps_kappa\]](#c:alpha_beta_eps_kappa){reference-type="eqref" reference="c:alpha_beta_eps_kappa"}, and [\[c:add_b\]](#c:add_b){reference-type="eqref" reference="c:add_b"}. The convergence \(F_\nu \to F_0\) in \(C^{\alpha'}((0,2)\times \T^3)\) can be shown along the same lines, by observing that \(\| \partial_t u \|_{C^{\alpha '} (K_q^c \times \T^2)} \to 0\) as \(q \to \infty\). We argue exactly as in the proof of (iii) in Section [5](#section:proofOnsager){reference-type="ref" reference="section:proofOnsager"}. We first notice that \[2 \, \tilde \nu_q \int_0^{1-T_q + \bar t_q} \| \nabla \tilde \vartheta_{\tilde \nu_q}(s,\cdot)\|_{L^2}^2 \, ds = 2 \, \tilde \nu_q \int_0^{1-T_q + \bar t_q} \| \nabla \vartheta_{\tilde \nu_q}(s,\cdot)\|_{L^2}^2 \, ds > \frac{1}{2} \qquad \text{ for any \(q \in m\N\),}\] as a direct consequence of [\[prop:anomalous\]](#prop:anomalous){reference-type="eqref" reference="prop:anomalous"} in Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}. The first two components of \(v_\nu\) strongly converge to \(u\) in \(L^{\infty} ((0,2 ) \times \T^3)\) while \(\{\tilde \vartheta_{\tilde \nu_q}\}_{q\in \N}\) admits a limit point \(\vartheta^{\rm ds}\) in the weak topology of \(L^2((0,2)\times \T^2)\) which solves the transport equation with velocity field \(u\) and initial datum \(\vartheta_{\rm in}\). Setting \[v_{0}^{\rm ds} (t,x ) := \begin{pmatrix} u (t,x) \\ \vartheta^{\rm ds} (t,x) \end{pmatrix} \,, \quad t\in (0,2), \, x\in \T^3\,,\] we can verify that \((v_0^{\rm ds}, p_0, F_0)\) with \(p_0=0\) solves [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"} and \(v_0^{\rm ds}\) is an admissible solution by arguing exactly as in the proof of (iii) in Section [5](#section:proofOnsager){reference-type="ref" reference="section:proofOnsager"}. The first two components of \(v_{\nu_q}\) strongly converge to \(u\) in \(L^\infty ((0,2) \times \T^3)\). We claim that \(\tilde{\vartheta}_{\nu_q}\), the last component of \(v_{\nu_q}\), strongly converges in \(L^2((0,2) \times \T^2)\) to \(\vartheta_0\) (defined as in [\[prop:regularity-solution\]](#prop:regularity-solution){reference-type="eqref" reference="prop:regularity-solution"} of Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}). Setting \[v_{0}^{\rm cs} (t,x ) := \begin{pmatrix} u (t,x) \\ \vartheta_0 (t,x) \end{pmatrix} \,, \qquad t\in (0,2), \, x\in \T^3\] and observing that \(\| \vartheta_0 (t,\cdot)\|_{L^2} = \| \vartheta_{\rm in}\|_{L^2}\) for any \(t\in (0,2)\setminus\{1\}\), \(F_{\nu_q}\to F_0\) in \(C^{\alpha'}((0,2)\times \T^3)\), and \(u_q \to u\) in \(L^2((0,2)\times \T^3)\), the claimed convergence suffices to conclude that \(v_0^{\rm cs}\) is an (admissible) conservative solution to [\[e:E\]](#e:E){reference-type="eqref" reference="e:E"}. We argue as in the proof of (iv) in Section [5](#section:proofOnsager){reference-type="ref" reference="section:proofOnsager"}. Denoting by \(\vartheta_q:(0,2)\times \T^2 \to \mathbb R\) the unique solution to the transport equation with velocity field \(u_q\) and initial datum \(\vartheta_{\rm in}\), we have \[\label{e:lastsplit} \|\tilde \vartheta_{\nu_q}-\vartheta_0 \|_{L^2 ([0,2] \times \T^2)} \leq \| \tilde \vartheta_{\nu_q}-\vartheta_q \|_{L^2 ([0,2] \times \T^2)} + \| \vartheta_{q}-\vartheta_0 \|_{L^2 ([0,2] \times \T^2)} \,.\] We notice that \(\| \vartheta_{q}-\vartheta_0 \|_{L^2 ((0,2) \times \T^3)} \to 0\) as \(q \to \infty\), thanks to \(\vartheta_q (t,x)= \vartheta_0 (t,x)\) for any \(t \in K_q\) and any \(x \in \T^3\) (because of the symmetry of the velocity field \(u\) as in property [\[vectorfield_VV\]](#vectorfield_VV){reference-type="eqref" reference="vectorfield_VV"} of Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}) and to the bound \(\| \vartheta_0 \|_{L^\infty ((0,2) \times \T^2)} + \| \vartheta_q \|_{L^\infty ((0,2) \times \T^2)} \leq 2\). For any \(t \in (0,2)\), we estimate the first term in [\[e:lastsplit\]](#e:lastsplit){reference-type="eqref" reference="e:lastsplit"} relying on the regularity bound [\[prop:regularity-solution\]](#prop:regularity-solution){reference-type="eqref" reference="prop:regularity-solution"} and the symmetry property [\[vectorfield_VV\]](#vectorfield_VV){reference-type="eqref" reference="vectorfield_VV"} in Proposition [\[proposition\]](#proposition){reference-type="ref" reference="proposition"}. We have \[\begin{aligned}
& \| \tilde \vartheta_{\nu_q} (t, \cdot)-\vartheta_q (t, \cdot ) \|_{L^2(\T^2)}^2 \\ \leq & 2 \nu_q \left | \int_0^t \int_{\T^2} \nabla \tilde \vartheta_{\nu_q} (s,x) \cdot \nabla \vartheta_{q}(s,x) dx ds \right | \\ \leq & 2 \left (\nu_q \int_0^t \int_{\T^2} | \nabla \vartheta_{q} (s,x)|^2 dx ds \right )^{1/2} \\ = & 2 \left (\nu_q \sum_{j = q}^\infty \int_{\mathcal{I}_j\cup \mathcal{J}_j} \int_{\T^2} | \nabla \vartheta_{q} (s,x)|^2 dx ds + \nu_q \sum_{j = 0}^{q-1}\int_{\mathcal{I}_j\cup \mathcal{J}_j} \int_{\T^2} | \nabla \vartheta_{q} (s,x)|^2 dx ds \right )^{1/2} \\ \le & C \left( a_q^{2 + 3\epsilon} a_q^\gamma a_q^{-2(1+3\epsilon(1+\delta))} + a_q^{2 + 3\epsilon} \sum_{j=0}^{q-1} a_j^{\gamma} a_{j+1}^{-2(1+3\epsilon(1+\delta))} \right)^{1/2} \\ \le & C \left( a_q^{2 + 3\epsilon} a_{q-1}^\gamma a_q^{-2(1+3\epsilon(1+\delta))} \right)^{1/2} \\ \leq & Cq a_q^{1+ \frac{3\epsilon}{2}} a_{q-1}^\gamma a_q^{-1-6 \epsilon } \leq Cq a_q^{\frac{\delta}{32}-\frac{9 \epsilon}{2}} \to 0 \end{aligned}\] as \(q \to \infty\), where we used \(\gamma = \sfrac{\delta}{8}\), \(a_{q+1} = a_q^{1+ \delta}\), \(\delta \in (0,\sfrac{1}{8})\), \(q a_q^{\sfrac{\delta}{32}} \leq 1\), \(\sum_{j \geq q} a_j^\gamma \leq 2 a_{q}^\gamma\), and \(\sfrac{\delta}{8} > \sfrac{9 \epsilon}{2}\). Therefore, \(v_0^{\mbox{cs}}\) satisfies [\[eq:conservative\]](#eq:conservative){reference-type="eqref" reference="eq:conservative"}. | {'timestamp': '2022-12-19T02:10:25', 'yymm': '2212', 'arxiv_id': '2212.08413', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08413'} |
# Introduction {#s:intro}
In Sir David R. Cox's highly influential paper 'Some Statistical Methods Connected with Series of Events' he invented doubly stochastic Poisson processes obtained by a generalization of Poisson processes where the intensity function \(\Lambda\) that varies over space or time is a stochastic process. These Cox models play nowadays an important role when analysing point patterns in Euclidean spaces or on spheres. In particular, Cox processes driven by a transformed Gaussian process (GP) \(Y\) or independent copies \(Y_1,\ldots,Y_h\) of \(Y\) play a major role: A Log Gaussian Cox process (LGCP) has \(\Lambda(u)=\exp(Y(u))\); LGCPs constitute the most widely used subclass of Cox processes. Further, an interrupted Cox process (ICP) is obtained by an independent thinning of a Poisson process, where the selection probability of a point \(u\) is given by \(\exp(-\sum_{i=1}^hY_i(u)^2)\). Moreover, a permanental Cox point process (PCPP) is obtained if \(\Lambda(u)=\sum_{i=1}^hY_i(u)^2\). In recent years there has been an increasing interest in analysing point patterns on a linear network \(L\), that is, \(L\) is a connected set in \(\mathbb R^k\) (the real coordinate space of dimension \(k\)) given by a finite union of bounded, closed, line segments which can only overlap at their endpoints, see, , the references therein as well as further references given later in the present paper. Figure [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"} shows two examples of point patterns observed on linear networks, which we will use for illustrative purposes throughout the paper. The first dataset was first analysed in and consists of 116 locations of street crimes reported in the period 25 April to 8 May 2002 in an area close to the University of Chicago. The second dataset is one of the six point pattern datasets analysed in and consists of 566 locations of spines on a dendrite tree protruding from a neuron. The contribution of the present paper is the following. We use isotropic covariance function models with respect to the geodesic metric \(d_{\mathcal G}\) or the resistance metric \(d_{\mathcal R}\) on \(L\) as developed in as well as new models developed in the present paper in order to construct models for isotropic GPs and hence new models for LGCPs, ICPs, and PCPPs on linear networks with isotropic pair correlation functions (more details follow in the next paragraph). Also we construct new simulation algorithms for GPs and consider for the first time statistical procedures and applications for parametric families of LGCPs, ICPs, and PCPPs on linear networks. Moreover, in continuation of the considerations in and we discuss which metric \(d_{\mathcal G}\) or \(d_{\mathcal R}\) we recommend for the kind of Cox processes studied in this paper. Incidentally, we also establish some useful results for the resistance metric. In brief, the paper consists of two parts, Sections [2](#s:2){reference-type="ref" reference="s:2"}-[3](#s:GP-cov-fct){reference-type="ref" reference="s:GP-cov-fct"} on our setting for isotropic covariance functions and related GPs, and Sections [4](#s:3){reference-type="ref" reference="s:3"}-[7](#s:conclusion){reference-type="ref" reference="s:conclusion"} on point processes, in particular Cox processes, including the cases of LGCPs, ICPs, and PCPPs on linear networks and how these models can be used for fitting real data. In more detail, the paper is organized as follows. Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"} considers a general metric space \((S,d)\) in order later to compare covariance functions and point processes for the case \(S=L\) with the cases where \(S=\mathbb R^k\) or \(S=\mathbb S^k\) (the \(k\)-dimensional unit sphere). When \(S=L\) we have in mind that \(d\) is either \(d_{\mathcal G}\) or \(d_{\mathcal R}\), where Section [2.2](#s:res-met){reference-type="ref" reference="s:res-met"} gives a summary of results for \(d_{\mathcal R}\), including a useful expression for the metric. Furthermore, when \(S=L\), Section [3.1](#s:iso-cov-fct){reference-type="ref" reference="s:iso-cov-fct"} studies isotropic covariance functions of the form \(c(u,v)=c_0(d(u,v))\) and provides a less technical summary of results from together with examples of isotropic covariance functions not appearing in that paper. Simulation algorithms for GPs on \(L\) with an isotropic covariance function are developed in Section [3.2](#s:sim-GP){reference-type="ref" reference="s:sim-GP"}, where the case with \(L\) a tree is particularly tractable. Our setting for point processes is given in Section [4.1](#s:3.1){reference-type="ref" reference="s:3.1"}, and Section [4.2](#s:3.2){reference-type="ref" reference="s:3.2"} introduces first and higher order intensity functions which become useful when we later study Cox process models. In particular, we focus on the pair correlation function and the related \(K\)-function defined in Section [4.2](#s:3.2){reference-type="ref" reference="s:3.2"}. As discussed in Section [4.3](#s:est-check){reference-type="ref" reference="s:est-check"}, \(g\), \(K\), and other functional characteristics for point processes become useful for statistical inference. Section [5.1](#s:cox-general){reference-type="ref" reference="s:cox-general"} treats Cox processes in our general setting, since moment and many other properties of Cox processes driven by transformed Gaussian processes do not depend on which space we consider-- the dependence of the metric space \((S,d)\) occurs when we want to specify well-defined parametric models of covariance functions, simulate the underlying GPs, and use statistical inference procedures, which for the case \(S=L\) are the topics of Sections [3](#s:GP-cov-fct){reference-type="ref" reference="s:GP-cov-fct"} and [6](#s:appl){reference-type="ref" reference="s:appl"}. Moreover, Section [5.2](#s:models){reference-type="ref" reference="s:models"} studies the properties of LGCPs, ICPs, and PCPPs models, and in Section [6](#s:appl){reference-type="ref" reference="s:appl"} we demonstrate how these models may be fitted to real and simulated data. Finally, Section [7](#s:conclusion){reference-type="ref" reference="s:conclusion"} summarises our findings and discuss some open problems. At this point we should stress the importance of considering an isotropic pair correlation function (pcf) \(g(u,v)=g_0(d(u,v))\):
\(I\)
: it is easier to plot \(g_0\) than \(g\);
\(II\)
: for \(S=L\), the \(K\)-function is only defined when \(g\) is isotropic;
\(III\)
: to the best of our knowledge, nonparametric estimators of \(g\) and \(K\) have always been derived under the assumption that \(g\) is isotropic;
\(IV\)
: moment based estimation procedures such as minimum contrast, composite likelihood, and Palm likelihood methods become computational convenient if \(g\) is isotropic;
\(V\)
: in particular, for LGCPs, ICPs, and PCPPs, isotropy of \(g\) becomes equivalent to isotropy of the covariance function for the underlying GPs;
\(VI\)
: so far flexible model classes for covariance functions have mainly been developed in the isotropic case .
A substantial part of this work was the development of an R-package `coxln`, in which the methods developed in the present paper are implemented. This package is available on Github under the author `gulddahl`. Moreover, we used the R-package `spatstat` extensively throughout the paper, see.
# Metric spaces {#s:2}
## Setting {#s:new-setting}
The state space for the points of the point processes considered in this paper is a metric space \((S,d)\) equipped with a reference measure \(\nu\). We focus on the case where \(S\) is a linear network and compare with the cases where \(S=\mathbb R^k\) is the \(k\)-dimensional Euclidean space or \(S=\mathbb S^k=\{u\in \mathbb R^{k+1}\,|\, \|u\|=1\}\) is the \(k\)-dimensional unit sphere, with \(d\) and \(\nu\) specified as follows.
\(a\)
: The case \(S=\mathbb R^k\): \(d(u,v)=\|u-v\|\) is given by the usual Euclidean distance and \(\nu(A)=\int_A\mathrm du\) is Lebesgue measure.
\(b\)
: The case \(S=\mathbb S^k\): \(d(u,v)=d_\mathcal{G}(u,v)=\arccos(u\cdot v)\) is the geodesic (orthodromic, great-circle, or shortest path) metric, and \(\nu=\nu_k\) is \(k\)-dimensional Hausdorff/surface measure. Here, \(u\cdot v\) is the usual inner product of \(u\) and \(v\).
\(c\)
: The case \(S=L=\cup_{i=1}^m L_i\) of a linear network: We assume \(m<\infty\), each \(L_i\subset\mathbb R^k\) is a closed line segment of length \(l_i\in (0,\infty)\), \(L_i\cap L_j\) is either empty or an endpoint of both \(L_i\) and \(L_j\) whenever \(i\not=j\), and \(L\) is a path-connected set. Furthermore, \(d\) is a 'natural' metric and \(\nu(A)=\int_A\mathrm d_L(u)\) is arc length measure. We let \(|L|=\sum_{i=1}^m l_i=\int_L\mathrm d_L(u)\) denote the length of the linear network.
The remainder of this section are remarks to the case (c). For disconnected linear networks, definitions and results may be applied separately to each connected component of the network if we consider independent Gaussian processes on the connected components and independent point processes on the connected components. Our definition in (c) may be extended to the more abstract case of a graph with Euclidean edges but we avoid this generalization for ease of presentation and since statistical methods have so far only been developed for the case (c). For each line segment \(L_i\), there are two possible arc length parametrisations. We assume one is chosen and given by \(u_i(t)=(1-t/l_i)a_i+(t/l_i)b_i\), \(t\in[0,l_i]\), where \(a_i\) and \(b_i\) are the endpoints of \(L_i\). The definitions and results in this paper will not depend on this choice, including when calculating arc length measure restricted to \(L_i\): For Borel sets \(A\subseteq L_i\), \(\int_A \mathrm d_L(u)=\int_0^{l_i}1(u_i(t)\in A)\,\mathrm dt\) where \(1(\cdot)\) denotes the indicator function. Let \(V\) denote the set of endpoints of \(L_1,\ldots,L_m\) and consider the graph with vertex set \(V\) and edge set \(E\) given by \(L_1,\ldots,L_m\). Thus, two distinct vertices \(u,v\in V\) form an edge if and only if \(\{u,v\}=\{a_j,b_j\}\) for some \(j\in\{1,\ldots,m\}\), in which case we write \(u\sim v\). We have two cases of natural metrics in mind, namely when \(d\) is the geodesic metric \(d_\mathcal{G}\) or the resistance metric \(d_\mathcal{R}\) defined as follows. Let \(u,v\in L\). Then \(d_\mathcal{G}(u,v)=\min\int_{p_{uv}}\mathrm d_L(w)\) where the minimum is over all paths \(p_{uv}\subseteq L\) connecting \(u\) and \(v\). When defining \(d_\mathcal{R}(u,v)\), without loss of generality assume that \(u,v\in V\), since if e.g. \(u\in L_j\setminus V\), we may split \(L_j\) into the two line segments with endpoints \(\{a_j,v\}\) and \(\{v,b_j\}\), and then consider a new graph with vertex set \(V\cup\{u\}\) and edge set \(E\cup\{\{a_j,v\},\{v,b_j\}\}\) (this follows from (B) in Theorem [\[t:res-m\]](#t:res-m){reference-type="ref" reference="t:res-m"} below). Viewing the graph \((V,E)\) as an electrical network with resistor \(l_i\) at edge \(L_i\), \(i=1,\ldots,m\), then \(d_\mathcal{R}(u,v)\) is the effective resistance between \(u\) and \(v\) as obtained by Kirkhoff's laws. In Section [2.2](#s:res-met){reference-type="ref" reference="s:res-met"} below we provide the detailed definition of \(d_\mathcal{R}\). Indeed there are other interesting metric when \(S=L\), including the least-cost metric, but to the best of our knowledge parametric models for isotropic covariance functions \(c(u,v)=c_0(d(u,v))\) have so far only been developed when \(d=d_\mathcal{G}\), \(d=d_\mathcal{R}\), or \(d\) is given by the usual Euclidean distance. However, Euclidean distance is usually not a natural metric on a linear network.
## The resistance metric {#s:res-met}
This section defines the resistance metric \(d_\mathcal{R}\) for a graph with Euclidean edges in the special case of a linear network \(L=\cup_{i=1}^m L_i\) as given in case (c) in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"}. The section also summarises some properties of \(d_\mathcal{R}\) and compares with \(d_\mathcal{G}\). Consider the graph \(G=(V,E)\) and its relation \(\sim\) as defined above, and denote \(d_V\) the classic (effective) resistance metric \(d_V\) on \(V\). Since \(d_\mathcal{R}\) is an extension of \(d_V\) to \(L\), we start by recalling the definition of \(d_V\) using a notation as follows. Let \(u_0\in V\) be an arbitrarily chosen vertex called the origin. For any \(u,v\in V\), define the so-called conductance function by \[{\mathrm{con}}(u,v)=\begin{cases*} 1/\|u-v\| & \text{if } \(u\sim v\) \\ 0 & \text{otherwise} \end{cases*}\] and define a matrix \(\Delta\) with rows and columns indexed by \(V\) so that its entry \((u,v)\) is given by \[\Delta(u,v)=\begin{cases*} 1+c(u) & \text{if } \(u=v=u_0\) \\ c(u) & \text{if } \(u=v\not=u_0\)\\ -{\mathrm{con}}(u,v) & \text{otherwise} \end{cases*}\] where \(c(u)=\sum_{w\in V:\,w\sim u}{\mathrm{con}}(u,w)\) is the sum of the conductances associated to the edges incident to vertex \(u\). In fact \(\Delta\) is symmetric and strictly positive definite, and it is similar to the 'Laplacian matrix' obtained when viewing \(G\) as an electrical network over the nodes with resistors given by the length of each line segment except that \(\Delta\) has the additional \(1\) added at entry \((u_0,u_0)\) (this makes \(\Delta\) invertible). Now, let \(B_0\) be a zero mean Gaussian vector indexed by \(V\) and having covariance matrix \(\Sigma=\Delta^{-1}\). Then the resistance metric on \(V\) is the variogram \[\label{e:d-V} d_V(u,v)=\mathbb{V}{\mathrm{ar}}(B_0(u)-B_0(v))=\Sigma(u,u)+\Sigma(v,v)-2\Sigma(u,v)\quad\mbox{for } u,v\in V.\] Extend \(B_0\) by linear interpolation to a zero mean Gaussian process (GP) \(Z_0\) on \(L\) so that \[Z_0(u)=\frac{\|u-b_i\|}{l_i}B_0(a_i)+\frac{\|u-a_i\|}{l_i}B_0(b_i)\quad\mbox{for }u\in L_i.\] For \(i=1,\ldots,m\), define a zero mean Brownian bridge \(B_i\) on \(L_i\) so that \[\mathbb{C}{\mathrm{ov}}(B_i(u),B_i(v))=\min\{\|u-a_i\|\|v-b_i\|,\|v-a_i\|\|u-b_i\|\}/l_i\quad\mbox{for \(u,v\in L_i\),}\] and define \[Z_i(u)=\begin{cases*} B_i(u) & \text{for } \(u\in L_i\) \\ 0 & \text{for } \(u\in L\setminus L_i\). \end{cases*}\] Finally, the resistance metric on \(L\) is defined by \[\label{e:def-dR} d_{\mathcal{R}}(u,v) =\sum_{i=0}^m \mathbb{V}{\mathrm{ar}}(Z_i(u)-Z_i(v))\quad\mbox{for } u,v\in L.\] For the following theorem, which follows from, we use a terminology as follows. A closed line segment in \(\mathbb R^k\) with endpoints \(a\) and \(b\) is denoted \([a,b]=\{at+b(1-t)\,|\, 0\le t\le 1\}\). A path is a subset of \(L\) of the form \([u,v_1]\cup [v_1,v_2]\cdots\cup [v_{i-1},v_{i}]\cup[v_{i},v]\) where \(u,v\in L\), \(v_1,\ldots,v_i\in V\) are vertices, \(i\ge0\) is an integer, and we interpret \([v_1,v_2]\cdots\cup [v_{i-1},v_{i}]\) as the empty set if \(i=0\). If all vertices in \(G\) are of order two, we say that \(L\) is a loop (since \(L\) is isomorphic to a circle). If there is no loop, we say that \(L\) is a tree.
Regarding the inequality in (C) above, considering e.g. the Chicago street network in Figure [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"}, \(\max d_{\mathcal R}\approx 675\) feet is much smaller than \(\max d_{\mathcal G}\approx 2031\) feet. It is also smaller than the side length of a square surrounding the network which is a little less than 1000 feet. The following proposition and the remarks below show that \(d_{\mathcal{R}}(u,v)\) is nicely behaving.
A few final remarks are in order. It follows from [\[e:d-R1\]](#e:d-R1){reference-type="eqref" reference="e:d-R1"} and [\[e:d-R2\]](#e:d-R2){reference-type="eqref" reference="e:d-R2"} that once \(\Sigma\) has been calculated, \(d_{\mathcal{R}}(u,v)\) can be quickly calculated. Moreover, if \(V\) is not too large or if \(\Delta\) is sparse, there are quick methods for obtaining \(\Sigma=\Delta^{-1}\). For example, for the Chicago street network in Figure [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"} there are 338 vertices and \(\Delta\) is sparse since it contains around 1.2% non-zero entries. In our implementation the inversion of \(\Delta\) took less than a 0.1 second using standard methods for inversion, so here sparse methods are not necessary, but for larger networks sparse methods may become relevant. The inequality [\[e:min\]](#e:min){reference-type="eqref" reference="e:min"} becomes useful when searching for point pairs \(u,v\in L\) with \(d_{\mathcal{R}}(u,v)\le r\) and \(v\in L_i\), since we need only to consider the cases where \(d_{\mathcal{R}}(u,a_i)\le r\) or \(d_{\mathcal{R}}(u,b_i)\le r\).
# Gaussian processes and isotropic covariance functions {#s:GP-cov-fct}
Let \((S,d)\) be a metric space as in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"} and recall that a function \(c:S\times S\mapsto\mathbb R\) is positive definite (sometimes called positive semi-definite) if \(\sum_{j,\ell=1}^n a_j{a_\ell}c(u_j,u_\ell)\ge0\) for all \(a_1,\ldots,a_n\in\mathbb R\), all pairwise distinct \(u_1,\ldots,u_n\in S\), and all \(n=1,2,\ldots\). Let \(Y=\{Y(u)\,|\,u\in S\}\) be a Gaussian process (GP) where each \(Y(u)\) is a real-valued random variable. The distribution of \(Y\) is specified by the mean function \(\mu(u)=\mathbb EY(u)\) and the covariance function \[c(u,v)=\mathbb C{\mathrm{ov}}(Y(u),Y(v))=\mathbb E[Y(u) {Y(v)}]-\mu(u) {\mu(v)}.\] The necessary and sufficient condition for a well-defined GP in terms of such two functions \(\mu\) and \(c\) is just that \(c\) is symmetric and positive definite.
## Isotropic covariance functions {#s:iso-cov-fct}
We are in particular interested in isotropic covariance functions \(c(u,v)=c_0(d(u,v))\) for all \(u,v\in S\), where with some abuse of terminology we also call \(c_0\) a covariance function. So \(c_0\) is required to be positive definite, that is, \(\sum_{j,\ell=1}^n a_j{a_\ell}c_0(d(u_j,u_\ell))\ge0\) for all \(a_1,\ldots,a_n\in\mathbb R\), all pairwise distinct \(u_1,\ldots,u_n\in S\), and all \(n=1,2,\ldots\). Henceforth, we assume that the variance \(\sigma^2=c_0(0)\) is strictly positive. Below in Sections [3.1.1](#s:classical){reference-type="ref" reference="s:classical"} and [3.1.2](#s:L-cov){reference-type="ref" reference="s:L-cov"} we pay attention to the correlation function \(r_0(t)=c_0(t)/\sigma^2\) in connection to the cases (a)--(c) in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"}.
### The classical cases of Euclidean spaces and spheres {#s:classical}
Let \(S=\mathbb R^k\) and \(d(u,v)=\|u-v\|\), cf. case (a) in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"}. and provided detailed studies of continuous isotropic covariance functions for this case. Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"} shows examples of popular parametric models for continuous isotropic correlation functions which are well-defined for every dimension \(k=1,2,\ldots\). Instead let \(S=\mathbb S^k\) and \(d(u,v)=d_\mathcal{G}(u,v)\), cf. case (b) in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"}. gave a complete characterization of continuous isotropic covariance functions in this case; see also. In particular, it was proven that \(r_0\) defines a continuous correlation function for any dimension \(k=1,2,\ldots\) if and only if \(r_0(t)=\sum_{\ell=0}^\infty\beta_\ell\cos^\ell(t)\) where the sequence \(\beta_0,\beta_1,\ldots\) is a probability mass function. Examples of such correlation functions are seen in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"}. Compared to the case \(S=\mathbb R^k\), the ranges of shape and smoothness parameters are more restrictive.
### The case of linear networks {#s:L-cov}
Suppose \(S=L\) is a linear network, cf. case (c) in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"}, where we let either \(d=d_{\mathcal G}\) or \(d=d_{\mathcal R}\). Many of the commonly used isotropic correlation functions, including those in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"}, are valid with respect to the resistance metric but not always with respect to the geodesic metric. The reason for this is discussed in this section. Typically (including all of our examples), \(r_0\) will be a completely monotone function. Recall that a function \(f: [0,\infty)\mapsto\mathbb R\) is completely monotonic if it is non-negative and continuous on \([0,\infty)\) and for \(j=0,1,\ldots\) and all \(u>0\), the \(j\)-th derivative \(f^{(j)}(u)\) exists and satisfies \((-1)^jf^{(j)}(u)\ge0\). By Bernstein's theorem, \(f\) is completely monotone if and only if it is the Laplace transform of a non-negative finite measure on \([0,\infty)\), meaning that for every \(t\ge0\), \[\label{e:bernstein} f(t)=f(0)\int\exp(-st)\,\mathrm dF(s)\] where \(F\) is a cumulative distribution function with \(F(s)=0\) for \(s<0\). We refer to \(F\) as the Bernstein CDF corresponding to \(f\). Thus, any non-negative valued distribution with a known Laplace transform can be used to produce a completely monotone function. This fact is used in the following example.
In the next theorem, which summarises Theorems 1 and 2 in, we need the following definition. We say that \(L\) is a 1-sum of \(\mathcal L_1=L_1\cup\ldots\cup L_j\) and \(\mathcal L_2=L_{j+1}\cup\ldots\cup L_m\) if \(\mathcal L_1\) and \(\mathcal L_2\) are (connected) linear networks where \(1\le j<m\), \(\mathcal L_1\cap\mathcal L_2=\{u_0\}\) consists of a single point \(u_0\), and \[d(u,v)=d(u,u_0)+d(v,u_0)\quad\mbox{whenever \(u\in\mathcal L_1\) and \(v\in\mathcal L_2\).}\] This property is possible if \(d=d_{\mathcal G}\) or \(d=d_{\mathcal R}\) but unless \(L\) is a straight line segment it is impossible if \(d\) is the given by the usual Euclidean distance. Using induction we say for \(n=3,4,\ldots\) that \(L=\mathcal L_1\cup\ldots\cup\mathcal L_n\) is a 1-sum of \(\mathcal L_1,\ldots,\mathcal L_n\) if \(L\) is a 1-sum of \(\mathcal L_1\cup\ldots\cup\mathcal L_{n-1}\) and \(\mathcal L_n\).
In Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"}, for each model, \(r_0\) is completely monotone for the ranges of the parameters. So for each example of \(r_0\) in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"} and for \(r_0\) given by \(f_1\), \(f_2\), or \(f_3\) in [\[e:f_1\]](#e:f_1){reference-type="eqref" reference="e:f_1"}--[\[e:f_3\]](#e:f_3){reference-type="eqref" reference="e:f_3"}, \(r_0(d_{\mathcal R}(u, v))\) is a valid correlation function, but by Theorem [\[t:1\]](#t:1){reference-type="ref" reference="t:1"} we only know that \(r_0(d_{\mathcal G}(u, v))\) is valid if \(L\) is a 1-sum of trees and loops. Note that in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"}, the ranges of the parameters are the same for the two cases \(S=\mathbb S^k\) and \(S=L\). Finally, [\[e:f_1\]](#e:f_1){reference-type="eqref" reference="e:f_1"} is the special case of the generalized Cauchy function when \(\alpha=1\) in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"}, whilst [\[e:f_2\]](#e:f_2){reference-type="eqref" reference="e:f_2"} and [\[e:f_3\]](#e:f_3){reference-type="eqref" reference="e:f_3"} are not covered by Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"}. For example, consider the Chicago street network in the left panel of Figure [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"}. Here \(L\) is not a 1-sum of trees and loops and therefore we cannot use the geodesic metric for the cases of covariance functions related to the Chicago street network. See also the counter examples in. On the other hand, the dendrite data shown in the right panel of Figure [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"} is observed on a tree, so here we can use the geodesic/resistance metric (by Theorem [\[t:1\]](#t:1){reference-type="ref" reference="t:1"}(C) the two metrics are equal in this case).
## Simulation of GPs on linear networks {#s:sim-GP}
This section discusses how to simulate a GP \(Y=\{Y(u)\,|\,u\in L\}\) on a linear network \(L=\cup_{i=1}^m L_i\) as given in case (c) in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"} using three different algorithms, which are all available in our package `coxln`. We assume without loss of generality that the mean function of \(Y\) is zero. The following is a straightforward algorithm applicable to any metric \(d\) and any linear network \(L\).
Specifically, we have chosen a grid \(D=V\cup D_1\cup…\cup D_m\) where each \(D_i\) is a fine discretization of \(L_i\) as described after the proof of Theorem [\[t:simalgotree\]](#t:simalgotree){reference-type="ref" reference="t:simalgotree"} below. The disadvantage of Algorithm [\[a:1\]](#a:1){reference-type="ref" reference="a:1"} is of course that the dimension of \(\Sigma_D\) can be large and hence eigenvalue decomposition (as well as other methods) can be slow. Algorithm [\[a:2\]](#a:2){reference-type="ref" reference="a:2"} below is much faster but requires \(L\) to be a tree and \(c(u,v) = \sigma^2\exp(-s d(u, v))\) to be an exponential covariance function with parameter \(s>0\) and \(d=d_{\mathcal G}=d_{\mathcal R}\) (the exponential correlation function \(r_0(t)=\exp(-ts)\) appears as two special cases in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"} with scale parameter \(\phi=1/s\), namely the powered exponential model with \(\alpha=1\) and the Mátern model with \(\alpha=\tfrac12\)). But we first need to establish a Markov property given in the following theorem, where we denote the shortest path between \(u,v\in L\) by \(p_{uv}\).
We use the following notation in Algorithm [\[a:2\]](#a:2){reference-type="ref" reference="a:2"}. Pick an arbitrary origin \(u_0\in V\) and set \(G_0(u_0)=\{u_0\}\). For \(j=1,2,\ldots\), if \(u\in V\setminus\cup_{i=0}^{j-1}G_i(u_0)\) and \(u\sim v\) for some \(v\in G_{j-1}(u_0)\), we call \(u\) a child of \(j\)-th generation to \(u_0\) and define \(G_j(u_0)\subset V\) as the set of all children of \(j\)-th generation to \(u_0\). Moreover, if \(u\in G_{j-1}(u_0)\), \(v\in G_j(u_0)\), and \(u\sim v\), we define a GP \(Y(u,v)=\{Y(w)\,|\,w\in(u,v]\}\) where \((u,v]\) is the half-open line segment with endpoints \(u\) and \(v\) so that \(u\) is excluded and \(v\) is included.
In practice, when using Algorithm [\[a:2\]](#a:2){reference-type="ref" reference="a:2"} for simulation, a discretization on each line segment is needed. For some integer \(n_j>0\) (which may depend on \(l_j\)) and \(i\in\{0,1,\ldots,{n_j}\}\), define \(s_i=i l_j/{n_j}\). Then, for each \(u\in L_j\), if \(u\) is the midpoint between \(s_{i-1}\) and \(s_i\), we approximate \(Y(u)\) by the average \((Y(u_j(s_{i-1}))+Y(u_j(s_i)))/2\), and otherwise if \(u\) is closest to \(u_j(s_i)\), we approximate \(Y(u)\) by \(Y(u_j(s))\). Further, to generate the \({n_j}-1\) normal variables \(Y(u_j(s_i))\), \(i=1,\ldots,{n_j}-1\), we start by generating the variable \(Y(u_j(s_1))\) in accordance to [\[e:M2\]](#e:M2){reference-type="eqref" reference="e:M2"} and [\[e:M22\]](#e:M22){reference-type="eqref" reference="e:M22"} with \(w=w_1=w_2=s_1\). Then we can add \(u_j(s_1)\) to the vertex set, whereby we split \(l_j\) into the two line segments given by this new vertex and the endpoints of \(l_j\). Hence, if \(n_j>1\), we can repeat the procedure when generating \(Y(u_j(s_2))\), and so on until all the \(n_j-1\) normal variables have been generated. Theorems [\[t:markov\]](#t:markov){reference-type="ref" reference="t:markov"} and [\[t:simalgotree\]](#t:simalgotree){reference-type="ref" reference="t:simalgotree"} do not hold if \(L\) is not a tree. Indeed, a GP on the circle \(\mathbb S^1\) (which is equivalent to a loop of length \(2\pi\)) with exponential covariance function, considering four arbitrary distinct points on \(\mathbb S^1\), it can be shown that the GP is not Markov by calculating the precision matrix of the GP at these four points. On the other hand, letting \(c(u,v) = a \cosh(b(d_\mathcal{G}(u,v)-\pi))\) for \(a,b>0\), verified that the GP on \(\mathbb S^1\) with covariance function \(c\) is Markov, but considering two arbitrary points on a tree together with a point on the path connecting them, it can be shown that the GP on the tree with covariance function \(c\) is not Markov. Consequently, we cannot have a covariance function only depending on the geodesic distance which makes an arbitrary linear network Markov. Moreover, in general, if we want to simulate a GP with an exponential covariance function on a linear network which is not a tree, we cannot rely on Markov properties. Instead we have to use the straightforward but slower Algorithm [\[a:1\]](#a:1){reference-type="ref" reference="a:1"}. For other covariance functions than the exponential, we may use the following theorem, which follows immediately from [\[e:bernstein\]](#e:bernstein){reference-type="eqref" reference="e:bernstein"} and the central limit theorem.
Theorem [\[t:3\]](#t:3){reference-type="ref" reference="t:3"} allows simulation of any GP with a covariance function of the form [\[e:c-ber\]](#e:c-ber){reference-type="eqref" reference="e:c-ber"}, if a simulation algorithm for \(F\) is available. For the case of a tree, Theorem [\[t:3\]](#t:3){reference-type="ref" reference="t:3"} combined with Algorithm [\[a:2\]](#a:2){reference-type="ref" reference="a:2"} gives the following simulation algorithm.
Conditioned on \(S_1...,S_n\), the output in Algorithm [\[a:3\]](#a:3){reference-type="ref" reference="a:3"} is a zero mean GP with covariance function \[\label{e:approxc} c(u,v|S_1...,S_n)=\frac{\sigma^2}{n}\sum_{i=1}^n\exp(-S_i d(u,v)).\] The running times of Algorithms [\[a:1\]](#a:1){reference-type="ref" reference="a:1"}--[\[a:3\]](#a:3){reference-type="ref" reference="a:3"} for simulating a GP on a tree are compared in the following example, and the choice of \(n\) in Algorithm [\[a:3\]](#a:3){reference-type="ref" reference="a:3"} needed to obtain that [\[e:approxc\]](#e:approxc){reference-type="eqref" reference="e:approxc"} is a good approximation of the covariance function [\[e:c-ber\]](#e:c-ber){reference-type="eqref" reference="e:c-ber"} is considered in the example thereafter.
If \(L\) is not a tree, Algorithms [\[a:2\]](#a:2){reference-type="ref" reference="a:2"} and [\[a:3\]](#a:3){reference-type="ref" reference="a:3"} cannot be used, so we use the straightforward Algorithm [\[a:1\]](#a:1){reference-type="ref" reference="a:1"} instead, provided of course that \(Y\) is specified by a valid covariance function, meaning that Algorithm [\[a:1\]](#a:1){reference-type="ref" reference="a:1"} may work for \(d=d_{\mathcal{R}}\) but not for \(d=d_{\mathcal{G}}\) as illustrated in the following example.
# Point processes and some of their characteristics {#s:3}
Throughout this section we consider a general metric space \((S,d)\) equipped with its Borel \(\sigma\)-algebra \(\mathcal B\) and a reference measure \(\nu\). Sometimes this is specified to be \(S\in\{\mathbb R^k,\mathbb S^k,L\}\) with \(d\) and \(\nu\) as given in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"}.
## Point process setting {#s:3.1}
We restrict attention to point processes whose realisations can be viewed as locally finite subsets \(x\) of \(S\): For \(x,A\subseteq S\), define \(x_A= x\cap A\) and let \(n(x_A)\) denote the cardinality of \(x_A\). We let the state space of a point process be \[\mathcal N= \{x\subset S\,|\,n(x_A)<\infty \mbox{ whenever \(A\subseteq S\) is bounded}\}\] and equip \(\mathcal N\) with the smallest \(\sigma\)-algebra \(\mathcal F\) such that the mapping \(x\mapsto n(x_A)\) is measurable whenever \(A\subseteq S\) is a bounded Borel set. So by a point process is meant a random variable \(X\) with values in \(\mathcal N\). Then, for any bounded Borel set \(A\subseteq S\), the count \(N(A)=n(X_A)\) is a random variable. We need the following notions. We say that \(X\) is a Poisson process with intensity function \(\rho:S\mapsto[0,\infty)\) if for any bounded Borel set \(A\subseteq L\), \(N(A)\) is Poisson distributed with finite mean \(\int_A\rho(u)\,\mathrm d\nu(u)\), and conditioned on \(N(A)\), the points in \(X_A\) are independent and each point has a density proportional to \(\rho\) with respect to \(\nu\) restricted to \(A\). For every \(u\in L\), we let \(X_u\) be the point process which follows the reduced Palm distribution of \(X\) at \(u\), that is, \[\mathbb E\sum_{u\in X}h(X\setminus\{u\},u)=\int\rho(u)\mathbb E h(X_u,u)\,\mathrm d\nu(u)\] for any non-negative measurable function \(h\) defined on \(\mathcal N\times S\) (equipped with the product \(\sigma\)-algebra of \(\mathcal F\) and \(\mathcal B\)). Intuitively, \(X_u\) follows the distribution of \(X\setminus\{u\}\) conditioned on that \(u\in X\). If \(\rho(u)=0\), \(X_u\) may follow an arbitrary distribution. If \(X\) is a Poisson process with intensity function \(\rho\), then \(X\) and \(X_u\) are identically distributed whenever \(\rho(u)>0\). Let \(X_1\) denote the Poisson process with intensity 1. Suppose \(S\in\{\mathbb S^k,L\}\) and the distribution of \(X\) is absolutely continuous with respect the distribution of \(X_1\) (this makes sense since \(S\) is bounded and hence \(X\) is finite). Let \(f\) be the density of (the distribution of) \(X\) with respect to (the distribution of) \(X_1\). If \(\rho(u)>0\), then \(X_u\) has a density \(f_u\) with respect to \(X_1\) such that \[\label{e:Palmdensity} f_u(\{u_1,\ldots,u_n\})=f(\{u,u_1,\ldots,u_n\})/\rho(u)\quad\mbox{for \(n=1,2,\ldots\) and pairwise distinct }u_1,\ldots,u_n\in S\setminus\{u\}.\]
## Moment and invariance properties of point processes {#s:3.2}
The point process \(X\) has \(n\)-th order intensity function \(\rho(u_1,\ldots,u_n)\) (with respect to the \(n\)-fold product measure of \(\nu\)) if this is a non-negative Borel function so that \[\label{e:1} \mathbb E\left[N(A_1)\cdots N(A_n)\right]=\int_{A_1}\cdots\int_{A_n}\rho(u_1,\ldots,u_n)\,\mathrm d\nu(u_1)\cdots\,\mathrm d\nu(u_n)<\infty\] for every pairwise disjoint, bounded, Borel sets \(A_1,\ldots,A_n\subset S\) (\(\rho(u_1,\ldots,u_n)\) is also called the \(n\)-th order product density for the \(n\)-th order reduced moments measure). Thus, \(\rho(u_1,\ldots,u_n)\) is a locally integrable function, which is almost everywhere unique on \(S^n\) (with respect to the \(n\)-fold product measure of \(\nu\)). In the following, for simplicity nullsets are ignored, so the non-uniqueness of \(\rho(u_1,\ldots,u_n)\) is ignored. Moreover, when we write \(\rho(u_1,\ldots,u_n)\) it is implicitly assumed that the \(n\)-th order intensity function exists. In particular, \(\rho(u)\) is the usual intensity function. The point process is said to be (first-order) homogeneous if \(\rho(u)=\rho_0\) is constant. For instance, this is the case if \(S=\mathbb R^k\) and \(X\) is stationary, i.e., its distribution is invariant under translations in \(\mathbb R^k\); or if \(S=\mathbb S^k\) and \(X\) is isotropic, i.e., its distribution is invariant under rotations on \(\mathbb S^k\). A similar example is not easy to specify if \(S=L\), since there is no natural (transitive) group action on a linear network. Instead of the second order intensity function, one usually considers the pair correlation function (pcf) given by \[g(u,v)=\frac{\rho(u,v)}{\rho(u)\rho(v)},\] setting \(\frac{a}{0}=0\) for \(a\ge0\), and one usually assumes that \[\label{e:iso-asump} g(u,v)=g_0(d(u,v))\] is isotropic. For a Poisson process, \[\label{e:Poisson-rho-n} \rho(u_1,\ldots,u_n)=\rho(u_1)\cdots\rho(u_n),\] so \(g=1\). One often interprets \(g_0(t)>1\) as repulsion/inhibition and \(g_0(t)<1\) as attraction/clustering for point pairs at distance \(t\) apart but one should be careful with this interpretation as \(t\) grows. The specific models in this paper are typically attractive (\(g>1\)) and satisfies the following stronger property: For \(n=2,3,\ldots\), \(i=1,\ldots,n-1\), and any pairwise distinct \(u_1,\ldots,u_n\in S\), \[\label{e:pos-cor} \rho(u_1,\ldots,u_n)\ge \rho(u_1,\ldots,u_i)\rho(u_{i+1},\ldots,u_n),\] or equivalently, for any pairwise disjoint, bounded, Borel sets \(A_1,A_2,\ldots\subset S\), \[\mathbb E\left[N(A_1)\cdots N(A_n)\right]\ge \mathbb E\left[N(A_1)\cdots N(A_i)\right]\mathbb E\left[N(A_{i+1})\cdots N(A_n)\right].\] So [\[e:pos-cor\]](#e:pos-cor){reference-type="eqref" reference="e:pos-cor"} means that the counts \(N(A_1),N(A_2),\ldots\) are positively correlated at all orders, and for brief we shall say that \(X\) is positively correlated at all orders. Non-parametric estimation of \(\rho\) and \(g\) are discussed in e.g. and the references therein; see also when \(S=\mathbb R^k\); when \(S=\mathbb S^k\); and and when \(S=L\). For non-parametric estimation of the pcf, usually kernel methods are used under the assumption [\[e:iso-asump\]](#e:iso-asump){reference-type="eqref" reference="e:iso-asump"} of isotropy (but we do not need to assume that \(X\) is (first-order) homogeneous). Since kernel methods are sensible to the choice of bandwidth, popular alternatives are given by estimators of the \(K\)-function defined in [\[e:K1new\]](#e:K1new){reference-type="eqref" reference="e:K1new"}--[\[e:K2new\]](#e:K2new){reference-type="eqref" reference="e:K2new"} below. On the other hand, for the specific parametric models in this paper, we have simple expressions for \(g_0\) but not for \(K\), and since the \(K\)-function is an accumulated version of \(g_0\), it may be harder to interpret. More details on these points are given in Sections [4.2.1](#s:3.2.1){reference-type="ref" reference="s:3.2.1"} and [4.2.2](#s:3.2.3){reference-type="ref" reference="s:3.2.3"} below.
### \(K\)-functions in the classical cases of Euclidean spaces and spheres {#s:3.2.1}
Let either \((S,d)=(\mathbb R^k,\|\cdot\|)\) and \(\nu\) be Lebesgue measure or \((S,d)=(\mathbb S^k,d_{\mathcal G})\) and \(\nu\) be surface measure, cf. cases (a)--(b) in Section [2.1](#s:new-setting){reference-type="ref" reference="s:new-setting"}. Suppose again that the pcf is isotropic. Following, , and, we say that \(X\) is second-order intensity reweighted stationary (SOIRS) if \(S=\mathbb R^k\) and second order intensity reweighted isotropic (or pseudo/correlation isotropic) if \(S=\mathbb S^k\). This property is satisfied if \(S=\mathbb R^k\) and \(X\) is stationary and isotropic or if \(S=\mathbb S^k\) and \(X\) is isotropic. Following again, , and, we define the (inhomogeneous) \(K\)-function for \(t\ge0\) by \[\label{e:K1new} K(t)=\int_{d(u,v)\le t} g_0(d(u,v))\,\mathrm d\nu(v)\] which does not depend on the choice of \(u\in S\). So \[K(t)={K_{\mathbb R^k}(t)=\sigma_{k-1} \int_0^t r^{k-1} g_0(r)\,\mathrm dr} \quad \text{if } S=\mathbb R^k \text{ and } t\ge0\] where \(\sigma_{k-1}=2\pi^{k/2}/\Gamma(k/2)\) is the surface area of the \((k-1)\)-dimensional unit sphere, and \[K(t)={K_{\mathbb S^k}(t)=\sigma_{k-1}\int_0^t g_0(\vartheta)\sin^{k-1}\vartheta\,\mathrm d\vartheta} \quad \text{if } S=\mathbb S^k \text{ and } 0\le t\le\pi.\] Thus, for \(k=1\), \(K_{\mathbb R}(t)=K_{\mathbb S}(t)\) agree for every \(t\in[0,\pi]\).
### \(K\)-functions in the case of linear networks {#s:3.2.3}
Suppose \(S=L\) and \(g(u,v)=g_0(\delta(u,v))\) where \(\delta\) is a metric on \(L\) (since we consider derivatives below, it is convenient to switch from the previous notation \(d\) to \(\delta\)). Then \(X\) is said to be \(\delta\)-correlated, cf. if \(\delta=d_{\mathcal G}\), \(X\) is also said to be second-order reweighted pseudostationary. Following and defining \(R=\inf_{u\in L}\sup_{v\in L}d(u,v)\), the \(K\)-function is given by \[\label{e:K2new} K(t)=K_L(t)=\int_0^t g_0(r)\,\mathrm dr,\quad 0\le t\le R.\] Note that \(K\) depends only on \(g_0\), but \(g\) depends on both \(g_0\) and \(\delta\). We have \(K_L=K_{\mathbb R}/2\) if \(L\) is a straight line segment which is broken into \(m\) line segments. If \(X\) is a Poisson process, then \(K(t)=t\). Non-parametric estimation of \(K\) was carefully studied in under the following technical assumption for the metric. Suppose \(\delta\) is regular, meaning that for every \(u\in L\), \(\delta(u,v)\) is a continuous function of \(v\in L\) and there is a finite set \(N\subset L\) such that for \(i=1,\ldots,m\) and all \(v\in L_i\setminus N\), the Jacobian \[\label{e:Jacobian} J_\delta(u,v)=|(\mathrm d/\mathrm dt) \delta(u,v)|\] exists and is non-zero where \(t=\|v-a_i\|\). Both \(d_\mathcal{G}\) and \(d_\mathcal{R}\) are regular, where \(J_{d_\mathcal{G}}=1\) and a useful expression for the calculation of \(J_{d_\mathcal{R}}\) is given in the following corollary which follows immediately from Proposition [\[p:calc\]](#p:calc){reference-type="ref" reference="p:calc"}.
Some final remarks are in order. For \(u\in L\) and \(0\le t\le R\), define \[w_\delta(u,t)=1\bigg/\sum_{v\in L:\, \delta(u,v)=t}1/J_\delta(u,v).\] This is a weight which accounts for the geometry of the linear network when shifting from arc length measure on \(L\) to Lebesgue measure on the positive half-line. For \(\delta=d_{\mathcal{G}}\), we have \(w_{\mathcal G}(u,t)=1/\#\{v\in L\,|\,\delta(u,v)=t\}\), since \(J_{d_\mathcal{G}}=1\), and for \(\delta=d_{\mathcal{R}}\), once the matrix \(\Sigma\) from Section [2.2](#s:res-met){reference-type="ref" reference="s:res-met"} has been calculated, \(w_{d_{\mathcal{R}}}\) is quickly calculated from [\[e:der-d-R\]](#e:der-d-R){reference-type="eqref" reference="e:der-d-R"}. It follows from [\[e:1\]](#e:1){reference-type="eqref" reference="e:1"}, [\[e:K2new\]](#e:K2new){reference-type="eqref" reference="e:K2new"}, and that for any Borel set \(A\subseteq L\) of positive arc length measure, \[K(t)= \frac{1}{\nu_L(A)}\mathbb E\sum_{u\in X_A}\sum_{v\in X\setminus\{u\}}\frac{1(\delta(u,v)\le t)w_\delta(u,\delta(u,v))}{\rho(u)\rho(v)}. \label{e:K3}\] Non-parametric estimators of \(K\) are based on omitting the expectation symbol in [\[e:K3\]](#e:K3){reference-type="eqref" reference="e:K3"}, possibly after elaborating on the right hand side in [\[e:K3\]](#e:K3){reference-type="eqref" reference="e:K3"} in order to realize how correction factors can be included in order to adjust for edge effects, cf.İn terms of Palm probabilities, for \(d_L\)-almost all \(u\in L\) with \(\rho(u)>0\), \[K(t)=\frac{1}{|L|}\mathbb E \sum_{v\in X_u}\frac{1(\delta(u,v)\le t)w_\delta(u,\delta(u,v))}{\rho(v)}.\] In general the weight makes it hard to interpret this expression of \(K\).
## Estimation and model checking {#s:est-check}
For parametric families of Cox point process models the most common estimation methods are based on the intensity, pair correlation, or \(K\)-functions using either minimum contrast estimation, composite likelihood, or Palm likelihoods, see and the references therein. For the analyses in Section [6](#s:appl){reference-type="ref" reference="s:appl"}, we use minimum contrast estimation for fitting the Cox process models in Section [5.2](#s:models){reference-type="ref" reference="s:models"} to various datasets in Section [6](#s:appl){reference-type="ref" reference="s:appl"} using the pcf, i.e., if \(g_0\) depends on a parameter, we estimate this parameter by minimizing the integral \[\label{e:contrast} D(g_0,\hat g_0) = \int_{a_1}^{a_2} \left|g_0(t)^q-\hat{g}_0(t)^q\right|^p \mathrm{d}t\] where \(\hat g_0\) is a non-parametric estimate of \(g_0\) and \(0\leq a_1 < a_2\) and \(p,q>0\) are user-specified values (see Section [6](#s:appl){reference-type="ref" reference="s:appl"}). The models in Section [5.2](#s:models){reference-type="ref" reference="s:models"} all have nice expressions of the pcf but not of the \(K\)-function. Moreover, the models used for the data analyses in Section [6](#s:appl){reference-type="ref" reference="s:appl"} include a parameter for the intensity function, which \(g_0\) does not depend on, and this parameter is estimated by a simple moment method; in the simplest case homogeneity is assumed and the intensity is estimated by \(\hat{\rho}={n(x)}/{|L|}\). For model checking other functional summary statistics are needed when \(\rho\), \(g\), or \(K\) and their corresponding non-parametric estimators have been used for estimation. For the case \(S=L\), suggested analogies to the so-called \(F,G,J\)-functions which account for the geometry of the linear network. showed that their \(F,G,J\)-functions make good sense under a certain condition called intensity reweighted moment pseudostationarity (IRMPS): \(X\) is IRMPS if \(\inf\rho>0\) and \(\delta\) is a regular metric on \(L\) such that for \(n=2,3,\ldots\), any pairwise distinct \(u_1,\ldots,u_n\in L\), and any \(u\in L\), \(g(u_1,\ldots,u_n)=\rho^{(n)}(u_1,\ldots,u_n)/[\rho(u_1)\cdots\rho(u_n)]\) is of the form \[\label{e:IRMPS} g(u_1,\ldots,u_n)=g_0(\delta(u,u_1),\ldots,\delta(u,u_n))\] for some function \(g_0\). This condition is satisfied if \(X\) is either a Poisson process or a log Gaussian Cox process (LGCP) with a stationary pair correlation function (which is usually not a natural assumption, cf. Section [5.2.1](#s:LGCP){reference-type="ref" reference="s:LGCP"}). Apart from these examples did not verify any other cases of models where IRMPS is satisfied. At least demonstrated the practical usefulness of their empirical estimator of the \(J\)-function for both a Poisson process, a simple sequential inhibition (SSI) point process, and a LGCP. We show in Section [5.2.1](#s:LGCP){reference-type="ref" reference="s:LGCP"} that IRMPS is in general not satisfied for the LGCP. For the SSI point process in it is hard to evaluate \(g(u_1,\ldots,u_n)\) and hence to check the assumption of IRMPS. Alternatively, introduced three purely empirical summary functions obtained by modifying the empirical \(F,G,J\)-functions for (inhomogeneous) point patterns on a Euclidean space to linear networks. Briefly, the modification consists of replacing the Euclidean space with the linear network, introducing the shortest path distance instead of the Euclidean distance, and adapting the notion of an eroded set to linear networks. demonstrated the usefulness of their empirical \(F,G,J\)-functions for model checking although underlying theoretical functions are missing. We also use these functions for the data analyses in Section [6](#s:appl){reference-type="ref" reference="s:appl"}. Specifically, we will consider a concatenation of the three functions and validate a fitted model using a 95% global envelope (i.e., a 95% confidence region for the concatenated function) and a \(p\)-value obtained by the global envelope test (based on the extreme rank length) as described in, , and. For the dendrite spine locations datasets analysed in it is concluded that the results based on such a global envelope test are consistent with the results obtained if instead the summary functions from are used.
# Cox processes driven by transformed Gaussian processes {#s:cox}
## General setting {#s:cox-general}
Let \((S,d)\) be a metric space with reference measure \(\nu\) and let \(\Lambda=\{\Lambda(u)\,|\,u\in S\}\) be a non-negative stochastic process so that with probability 1, for any bounded Borel set \(A\subseteq S\), the measure \(Usually in applications\[is unobserved, and so if only one point pattern dataset is observed the Cox process\)X\(is indistinguishable from the inhomogeneous Poisson process\)X\|\(. Henceforth, assume\)E(u)\(is a locally integrable function with respect to\]. Then the Cox process\)X\(is well-defined, since almost surely\) By conditioning on \(\Lambda\) it follows from [\[e:1\]](#e:1){reference-type="eqref" reference="e:1"} and [\[e:Poisson-rho-n\]](#e:Poisson-rho-n){reference-type="eqref" reference="e:Poisson-rho-n"} that \[\label{e:Cox-rho-n} \rho(u_1,\ldots,u_n)=\mathbb E\left[\Lambda(u_1)\cdots\Lambda(u_n)\right].\] Let \(W\subseteq S\) be a bounded Borel set with \(\nu(W)>0\); we think of \(W\) as an observation window. Then \(X_W\) is a Cox process driven by \(\Lambda\) restricted to \(W\), and \(X_W\) has a density given by \[\label{e:densityCox} f(\{u_1,\ldots,u_n\})=\mathbb E\left[\exp\left(\int_W (1-\Lambda(u))\,\mathrm d\nu(u)\right)\prod_{i=1}^n\Lambda(u_i)\right]\quad\mbox{for pairwise distinct }u_1,\ldots,u_n\in W\] with respect to \(X_1\cap W\), where \(X_1\) is the unit rate Poisson process, cf. Section [4.1](#s:3.1){reference-type="ref" reference="s:3.1"}. In particular, if \(S\in\{\mathbb S^k,L\}\), \(W=S\), and \(u\in S\) with \(\rho(u)>0\), then \(X_u\) has density \[\label{e:PalmdensityCox} f_u(\{u_1,\ldots,u_n\})=\mathbb E\left[\exp\left(\int_W (1-\Lambda(u))\,\mathrm d\nu(u)\right)\frac{\Lambda(u)}{\rho(u)}\prod_{i=1}^n\Lambda(u_i)\right]\] with respect to \(X_1\), cf. [\[e:Palmdensity\]](#e:Palmdensity){reference-type="eqref" reference="e:Palmdensity"}. In general the densities in [\[e:densityCox\]](#e:densityCox){reference-type="eqref" reference="e:densityCox"} and [\[e:PalmdensityCox\]](#e:PalmdensityCox){reference-type="eqref" reference="e:PalmdensityCox"} are intractable because the expected values are difficult to evaluate. Instead we exploit [\[e:Cox-rho-n\]](#e:Cox-rho-n){reference-type="eqref" reference="e:Cox-rho-n"} for calculating the intensity and pair correlation functions and for making inference as demonstrated in the following sections. As pointed out in it is useful to write \(\Lambda\) as \[\label{e:Lambda} \Lambda(u)=\rho(u)\Lambda_0(u)\] where \(\Lambda_0=\{\Lambda_0(u)\,|\,u\in S\}\) is a non-negative 'residual' stochastic process with \(\mathbb E\Lambda_0(u)=1\) whenever \(\rho(u)>0\). Then \[\rho(u_1,\ldots,u_n)=\rho(u_1)\cdots\rho(u_n)\mathbb E\left[\Lambda_0(u_1)\cdots\Lambda_0(u_n)\right]\] and \[\label{e:g-Lambda0} g(u,v)=\mathbb E\left[\Lambda_0(u)\Lambda_0(v)\right].\] Typically, it is only \(\rho(u)\) which is allowed to depend on covariate information, whilst \(\Lambda_0\) is considered to account for unobserved covariates or other effects which has not been successfully fitted by a Poisson process with intensity function \(\rho\), see e.g. and. Usually, assumptions on \(\Lambda_0\) are imposed so that \(g(u,v)=g_0(d(u,v))\) is isotropic; this is the case if \(S=\mathbb R^k\) and \(\Lambda_0\) is stationary and isotropic, or if \(S=\mathbb S^k\) and \(\Lambda_0\) is isotropic. Furthermore, for specific models of \(\Lambda_0\) it usually happens that \[\label{e:all-orders} \mathbb E\left[\Lambda_0(u_1)\cdots\Lambda_0(u_n)\right]\ge\mathbb E\left[\Lambda_0(u_1)\cdots\Lambda_0(u_i)\right]\mathbb E\left[\Lambda_0(u_{i+1})\cdots\Lambda_0(u_n)\right]\] for \(n=2,3,\ldots\), \(i=2,\ldots,n\), and all pairwise distinct \(u_1,\ldots,u_n\in S\), meaning that \(X\) is positively correlated at all orders, cf. [\[e:pos-cor\]](#e:pos-cor){reference-type="eqref" reference="e:pos-cor"}.
## Models {#s:models}
Consider a GP \(Y=\{Y(u)\,|\,u\in S\}\) with mean function \(\mu\) and covariance function \(c\), and let \(Y_1,\ldots,Y_h\) be independent copies of \(Y\). In the remainder of this paper we study the following models.
- \(X\) is a log Gaussian Cox process (LGCP) if \[\label{e:defLGCP} \Lambda_0(u)=\exp(Y(u))\] and \(\mu(u)=-c(u,u)/2\) for all \(u\in S\). The latter condition is required since we want \(\mathbb E\Lambda_0(u)=1\).
- \(X\) is an interrupted Cox process (ICP) if \(\mu=0\) and \[\label{e:defInter} \Lambda_0(u)=\Pi(u)(1+2c(u,u))^{h/2}\] for all \(u\in S\) where we define \(\Pi(u)=\exp(-\sum_{i=1}^h Y_i(u)^2)\) (this definition differs slightly from the one used in, which includes a factor \(1/2\) inside the exponential function). Since \(\mathbb E\Pi(u)=(1+2c(u,u))^{-h/2}\), we have \(\mathbb E\Lambda_0(u)=1\).
- \(X\) is a permanental Cox point process (PCPP) if \(\mu=0\), \(c(u,u)=1\), and \[\label{e:defPCPP} \Lambda_0(u)=\frac{1}{h}\sum_{i=1}^h Y_i(u)^2\] for all \(u\in S\). Since the sum in [\[e:defPCPP\]](#e:defPCPP){reference-type="eqref" reference="e:defPCPP"} is \(\chi^2(h)\)-distributed, \(\mathbb E\Lambda_0(u)=1\).
In all cases, the distribution of \(X\) is completely specified by \((\rho,c)\) and in the case of an ICP or PCPP the value of \(h\). Note that the intensity function \(\rho\) can be any non-negative locally integrable function with respect to \(\nu\). LGCP, ICP, and PCPP models are well-studied when \(S\in\{\mathbb R^k,\mathbb S^k\}\), and most of their properties immediately extend to the case \(S=L\) as discussed in the following.
### Log Gaussian Cox processes {#s:LGCP}
Let \(X\) be a LGCP, cf. [\[e:defLGCP\]](#e:defLGCP){reference-type="eqref" reference="e:defLGCP"}. studied LGCPs in the case \(S=\mathbb R^k\), and in the case \(S=\mathbb S^k\). As in and, we obtain for the general setting of the space \(S\in\{\mathbb R^k,\mathbb S^k,L\}\) the following results. For any integer \(n\ge2\) and any pairwise distinct \(u_1,\ldots,u_n\in S\), \[\label{e:n-LGCP} \rho(u_1,\ldots,u_n)=\rho(u_1)\cdots\rho(u_n)\exp\bigg(\sum_{1\le i<j\le n}c(u_i,u_j)\bigg).\] In particular, the LGCP is determined by \(\rho\) and \[\label{e:pairLGCP} g=\exp(c),\] i.e., by its first and second order moment properties. Note that \(g\) is isotropic if and only if \(c\) is isotropic, and then \(\rho(u_1,\ldots,u_n)\) depends only on the inter-point distances \(d(u_i,u_j)\), \(1\le i<j\le n\). Moreover, in most specific models, including those in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"}, \(c\ge0\) or equivalently \(X\) is positively correlated at all orders, cf. [\[e:pos-cor\]](#e:pos-cor){reference-type="eqref" reference="e:pos-cor"} and [\[e:n-LGCP\]](#e:n-LGCP){reference-type="eqref" reference="e:n-LGCP"}. For any \(u\in S\) with \(\rho(u)>0\), the reduced Palm distribution of \(X_u\) is a LGCP with intensity function \(\rho(v|u)=\rho(v)\exp(c(u,v))\) but the pcf is still \(g(v,w|u)=g(v,w)=\exp(c(v,w))\) for \(v,w\in S\). This follows from [\[e:densityCox\]](#e:densityCox){reference-type="eqref" reference="e:densityCox"} and [\[e:PalmdensityCox\]](#e:PalmdensityCox){reference-type="eqref" reference="e:PalmdensityCox"} in the cases \(S=W=\mathbb S^k\) and \(S=W=L\); if \(S=\mathbb R^k\), see. Consequently, if \(c\) is isotropic, the \(K\)-functions of \(X\) and \(X_u\) agree. We return to this result in Section [7.4](#s:7.4){reference-type="ref" reference="s:7.4"}. Assuming \(S=L\) is a linear network, let us return to the concept of IRMPS as defined by [\[e:IRMPS\]](#e:IRMPS){reference-type="eqref" reference="e:IRMPS"}. noticed that IRMPS is satisfied for the LGCP if \(\inf\rho>0\) and for all \(u_1,u_2,u\in L\), \[\label{e:strong-cond} c(u_1,u_2)=c_1(\delta(u,u_1),\delta(u,u_2))\] for some function \(c_1\). This statement is true due to [\[e:n-LGCP\]](#e:n-LGCP){reference-type="eqref" reference="e:n-LGCP"}, however, in our opinion [\[e:strong-cond\]](#e:strong-cond){reference-type="eqref" reference="e:strong-cond"} is a very strong condition, since we are not aware of any good examples where it is satisfied unless \(L\) is isometric to a closed interval and \(\delta\) is usual (Euclidean/geodesic/resistance) distance. Incidentally, in the metric \(\delta\) is assumed to be origin independent; they did not define the meaning of 'origin independent' but we have been informed (by personal communication) that they had in mind that [\[e:strong-cond\]](#e:strong-cond){reference-type="eqref" reference="e:strong-cond"} should be satisfied and when they let \(\delta=d_{\mathcal R}\) be the resistance metric, they admit that they misunderstood the meaning of origin independent as used in. Moreover, the proof of is incorrect (they claim that \(\delta(u',u_1)=\delta(u'',u_1)\) for any \(u',u'',u_1\in L\), which is obviously not correct if \(\delta=d_{\mathcal R}\), and which seems wrong in general for another choice of metric).
### Interrupted Cox processes
Let \(X\) be an ICP, cf. [\[e:defInter\]](#e:defInter){reference-type="eqref" reference="e:defInter"}. Then \(X\) conditioned on \(\Pi\) is obtained by an independent thinning of a Poisson process \(Z\) on \(S\) with intensity function \(\rho_Z(u)=\rho(u)\left(1+c(u,u)\right)^{h/2}\), where the selection probabilities are given by \(\Pi\). In the terminology of, \(X\) is an interrupted point process. studied the ICP (as well as many other examples of interrupted point processes) when \(S=\mathbb R^k\). For our general state space setting, assuming for ease of presentation that \(c(u,v)=\sigma^2r_0(d(u,v))\) is isotropic with \(r_0(0)=1\), we obtain in a similar way as in that the mean selection probability is constant and given by \[\label{e:pms} p_{\mathrm{ms}}=(1+2\sigma^2)^{-h/2}\] and the pcf is given by \[\label{e:g0ICP} g_0(t)=\left(\frac{(1+\sigma^2)^2}{(1+\sigma^2)^2-\sigma^4 r_0(t)^2}\right)^{h/2}.\] Third and higher-order moment results are less simple to express, but it can be proven that \(X\) is positively correlated at all orders. As \(\sigma\) increases from 0 to infinity, then \(p_{\mathrm{ms}}\) decreases from 1 to 0, whilst \(g_0(t)\) increases from 1 to \((1-r_0(t)^2)-{h/2}\) if \(r_0(t)\not=0\), which shows a trade-off between the degree of thinning and the degree of clustering. To understand how \(g_0(t)\) depends on \(h\) it is natural to fix the value of \(p_{\mathrm{ms}}\in(0,1)\). Then \[g_0(t)=\left(1+\left(1-p_{\mathrm{ms}}^{h/2}\right)^2r_0(t)^2\right)^{-h/2}\] is a strictly decreasing function of \(h\) whenever \(r_0(t)\not=0\). Consequently, taking \(h=1\) is natural if we wish to model as much clustering as possible when keeping the degree of thinning fixed. For the case \(S=L\) and \(r_0(t)=\exp(-t/\phi)\) specifying an isotropic exponential covariance function, expressions of \(K_L\) when \(h=1,2,\ldots,5\) were given in. Although these expressions were given for \(d=d_{\mathcal G}\), they remain true for a general metric \(d\) because the pcf depends only on \(h\), \(\sigma^2\), and \(r_0(t)=\exp(-t/\phi)\), cf. [\[e:g0ICP\]](#e:g0ICP){reference-type="eqref" reference="e:g0ICP"} and our comment after [\[e:K2new\]](#e:K2new){reference-type="eqref" reference="e:K2new"}.
### Permanental Cox point processes
Let \(X\) be a PCPP, cf. [\[e:defPCPP\]](#e:defPCPP){reference-type="eqref" reference="e:defPCPP"}. The case \(S=\mathbb R^k\) was studied in and; using the parametrization in the latter paper, \(X\) is a PCPP with parameters \(\alpha=h/2\) and \(C(u,v)=\sqrt{\rho(u)\rho(v)}c(u,v)/\alpha\). The process can also be defined in the general state space case, see. To stress that \(c\) is a correlation function, we shall write \(c=r\). For \(n=1,2,\ldots\) and \(u_1,\ldots,u_n\in S\), define the \(\alpha\)-weighted permanent by \[{\mathrm{per}}_\alpha[r](u_1,\ldots,u_n)=\sum_{\pi}\alpha^{\#\pi}r(u_1,u_{\pi_1})\cdots r(u_n,u_{\pi_n})\] where the sum is over all permutations \(\pi=(\pi_1,\ldots,\pi_n)\) of \((1,\ldots,n)\) and \(\#\pi\) is the number of cycles. The usual permanent corresponds to \(\alpha = 1\), in which case \(X\) is also called a Boson process. We have \[\label{e:uha} \rho(u_1,\ldots,u_n)=\rho(u_1)\cdots\rho(u_n)\mathrm{per}_\alpha[r](u_1,\ldots,u_n)/\alpha^n\] from which it can be verified that \(X\) is positively correlated at all orders. It also follows that the degree of clustering is a decreasing function of \(\alpha\). In particular, \[\label{e:pairPCPP} g(u,v)=1+r(u,v)^2/\alpha.\] Thus \(g\le 1+1/\alpha\le 3\), which reflects the limitation of modelling clustering by a PCPP. showed that exact computation of permanents of general matrices is a \(\#\)P (sharp P) complete problem, so no deterministic polynomial time algorithm is available. For most statistical purposes, approximate computation of permanent ratios is sufficient, and analytic approximations are available for large \(\alpha\). However, as \(\alpha\rightarrow\infty\), \(X\) tends to a Poisson process and the process becomes less and less interesting for the purpose of modelling clustering. The PCPP can be extended to the case where \(\alpha\ge0\) and \(c\) is not necessarily a covariance function (in which case we loose the connection to Gaussian and Cox processes), see and. Indeed the process also extends to the case where \(\alpha\) is a negative integer whereby a (weighted) determinantal point process (DPP) is obtained. We return to DPPs in Section [7.1](#s:7.1){reference-type="ref" reference="s:7.1"}.
## An index of cluster strength
Consider again a LGCP, ICP, or PCPP \(X\) defined on the metric space \(S\) and with an isotropic covariance function \(c(u,v)=\sigma^2r_0(d(u,v))\) where \(r_0(0)=1\), \(h=1\) if \(X\) is an ICP, and \(\sigma=1\) if \(X\) is a PCPP. To quantify how far \(X\) is from a Poisson process we follow in defining \[\label{e:varphi} \varphi=g_0(0)-1.\] If \(X\) has constant intensity \(\rho\), the total variation distance between the distribution of \(X\) and that of a Poisson process with intensity \(\rho\) is at most \(\rho\nu(S)\sqrt\varphi\). This follows since immediately extends to our setting. used \(\varphi\) to describe the degree of clustering for another class of Cox processes, namely Neyman-Scott point processes (with the cluster size following a Poisson distribution): the degree of clustering increases as \(\varphi\) increases. Although our Cox processes are not cluster point processes, \(\varphi\) has a similar interpretation since realizations of the processes may look more or less clustered. More precisely, it follows from [\[e:pairLGCP\]](#e:pairLGCP){reference-type="eqref" reference="e:pairLGCP"}, [\[e:g0ICP\]](#e:g0ICP){reference-type="eqref" reference="e:g0ICP"}, and [\[e:pairPCPP\]](#e:pairPCPP){reference-type="eqref" reference="e:pairPCPP"} that \(\varphi\) is equal to \[\label{e:varphiICP} \exp(\sigma^2)-1\mbox{ if \(X\) is a LGCP},\quad \frac{1+\sigma^2}{\sqrt{1+2\sigma^2}}-1\mbox{ if \(X\) is an ICP},\quad 2/h \mbox{ if \(X\) is a PCPP,}\] and \(\varphi+1\) is the maximal value of the pcf. Thus, if \(X\) is a LGCP or an ICP, \(\varphi\) is a strictly increasing function of \(\sigma>0\), where \(X\) approaches a Poisson process as \(\sigma\rightarrow0\), while realizations of \(X\) become more and more clustered as \(\sigma\) grows. If instead \(X\) is a PCPP, the degree of clustering decreases as \(h\in\{1,2...\}\) increases, where \(X\) approaches a Poisson process as \(h\rightarrow\infty\). The index \(\varphi\) for 'cluster strength' is used in Sections [6.3](#s:simstudy){reference-type="ref" reference="s:simstudy"} and [7.3](#s:7.3){reference-type="ref" reference="s:7.3"}.
# Application of statistical inference procedures {#s:appl}
For the analyses of the real and simulated datasets in this section, we have used the functions `mincontrast` and `linearpcf` from `spatstat` for calculating the minimum contrast estimates and the non-parametric estimate of the pcf (the latter is only available in the case of the geodesic metric), respectively, as well as a range of other minor functions from `spatstat` for handling point processes on linear networks. Furthermore, we have used the function `global_envelope_test` from the `GET` package for global envelope tests. The rest of the code (such as the estimation of the pcf in the case of the resistance metric, or the estimation of the \(F\), \(G\), and \(J\)-functions) we have implemented ourselves, and it is available in our package `coxln`.
## Analysis of Chicago crime dataset
used the empirical \(K\)-function together with simulations to show that the Chicago crime dataset (shown in the left panel of Figure [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"} and available in `spatstat`) was more clustered than a homogeneous Poisson process while an inhomogeneous Poisson process with log-quadratic intensity fitted by maximum likelihood provided a better fit. They remarked that the latter model was only provisional and shown for demonstration. For the analysis of the street crimes in this paper we fitted instead all three model classes (LGCP, ICP, and PCPP) given in Section [5](#s:cox){reference-type="ref" reference="s:cox"}. For the Gaussian process underlying the Cox process models, we used for simplicity a constant mean and an isotropic exponential covariance function with metric \(d=d_{\mathcal R}\) (since the network is not a 1-sum of trees and cycles, the geodesic metric will not give well-defined models). Thus we have (up to) four unknown parameters in each model: \(\rho\), the intensity of the point process; \(\sigma^2\), the variance of the Gaussian process (this is not a parameter in the PCPP model); \(s\), the scaling of the exponential covariance function; and \(h\), the number of Gaussian processes used in the model (this is not a parameter in the LGCP model). To estimate the parameters of all three models, we used the unbiased estimate \(\hat{\rho} ={n(x)}/{|L|}= 0.00372\) for the intensity and estimated the remaining parameters by the minimum contrast method based on the pair correlation function, cf. [\[e:contrast\]](#e:contrast){reference-type="eqref" reference="e:contrast"}. We estimated \(g_0\) non-parametrically on the interval \(r\in[0,100]\), but since the shape of the non-parametric estimate mostly resembled the shapes of the theoretical pair correlation functions for the three models on the interval \(r\in[20,100]\), in [\[e:contrast\]](#e:contrast){reference-type="eqref" reference="e:contrast"} we let \(a_1=20\) and \(a_2=100\) while \(p\) and \(q\) were given by the default values in `spatstat`. The estimated parameter values are shown in Table [3](#tab:chicago-estimates){reference-type="ref" reference="tab:chicago-estimates"}. Figure [\[fig:chicago-pcf\]](#fig:chicago-pcf){reference-type="ref" reference="fig:chicago-pcf"} shows the non-parametric estimate together with the estimated pair correlation functions for each of the three models, where the estimated pair correlation functions are very similar for the LGCP and the ICP, while the pair correlation function for the PCPP deviates from the other two at low distances.
The left column of Figure [\[fig:chicago-sim\]](#fig:chicago-sim){reference-type="ref" reference="fig:chicago-sim"} shows a simulation from each of the three fitted models. None of the simulations show strong deviation from the data, although the ICP does show a tendency to have small densely packed clusters of points that are not present in the data. Moreover, the right column of Figure [\[fig:chicago-sim\]](#fig:chicago-sim){reference-type="ref" reference="fig:chicago-sim"} shows the 95% global envelopes based on a concatenation of the empirical \(F,G,J\)-functions (based on 999 simulations) discussed at the end of Section [4.3](#s:est-check){reference-type="ref" reference="s:est-check"}. The LGCP provides the best fit with a \(p\)-value of 0.155 for the global envelope test. The ICP provides a rather bad fit to the data with a \(p\)-value of just 0.004, and the envelope shows a clear discrepancy between data and the model at distances below 50, where the discrepancy indicates more clustering in the fitted ICP model than in the data. The PCPP model shows a decent fit with a \(p\)-value just below 0.05 due to the \(F\)-function going outside the envelope at distance 100.
## Dendrite data {#s:den-an}
analysed several point pattern datasets given by spine locations on different dendrite trees which were identified by linear networks. For each dataset they fitted an inhomogeneous ICP with \(h=1\), \(c\) an isotropic exponential covariance function with metric \(d=d_{\mathcal G}=d_{\mathcal R}\), and an inhomogeneous intensity function given by a constant intensity \(\rho_{\mathrm{mb}}>0\) on the main branch and a different constant intensity \(\rho_{\mathrm{sb}}>0\) on the side branches. In this section we restrict attention to the dataset for dendrite number five in (the right panel of Figure [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"}) modelled by an ICP with \(h=1\) and considering different covariance function models. Hence we estimated the intensity parameters by \(\hat\rho_{\mathrm{mb}}= {n(x_{\mathrm{mb}})}/{|L_{\mathrm{mb}}|}=0.119\) and \(\hat\rho_{\mathrm{sb}} = {n(x_{\mathrm{sb}})}/{|L_{\mathrm{sb}}|}=0.184\) where \(L_{\mathrm{mb}}\) denotes the main branch, \(L_{\mathrm{sb}}\) the union of side branches, and \(x_{\mathrm{mb}}\) and \(x_{\mathrm{sb}}\) the point patterns restricted to these two subsets of \(L\). First, we let \(c(u,v)=\sigma^2\exp(-sd(u,v))\) be an isotropic exponential covariance function with \(d=d_{\mathcal G}=d_{\mathcal R}\). Then we estimated \((\sigma^2,s)\) following the recommendations in, i.e., we used the minimum contrast method with \(a_1=0\), \(a_2=50\), \(p=2\), and \(q=1\) in [\[e:contrast\]](#e:contrast){reference-type="eqref" reference="e:contrast"}. The obtained estimate \((\hat\sigma^2,\hat s)\) is given in Table [4](#tab:neu-estimates){reference-type="ref" reference="tab:neu-estimates"} and the corresponding estimated pcf is shown in Figure [\[fig:neu-pcf\]](#fig:neu-pcf){reference-type="ref" reference="fig:neu-pcf"} (the dotted curve). The estimate deviates a bit from that obtained in, which is due to a different implementation of the non-parametric estimation of the pair correlation function. Figure [\[fig:neu-FGJ\]](#fig:neu-FGJ){reference-type="ref" reference="fig:neu-FGJ"} (left panel) shows the 95% global envelope based on a concatenation of the empirical \(F,G,J\)-functions (based on 999 simulations). This shows a satisfactory fit with a \(p\)-value of \(0.26\) for the global envelope test. The \(p\)-value deviates substantially from the value obtained in, which may be due to the different estimates of \((\sigma^2,s)\) and a different number of simulations used in the global envelope test.
Second, we let \(c\) be one of the three covariance functions given in Example [\[ex:0\]](#ex:0){reference-type="ref" reference="ex:0"}, i.e., those with a gamma, an inverse gamma, and a generalised inverse Gaussian Bernstein CDF. We used the same minimum contrast procedure as above for estimating the parameters, thereby obtaining the estimates in Table [4](#tab:neu-estimates){reference-type="ref" reference="tab:neu-estimates"}. For all three cases of the estimated Bernstein CDFs, the variance is close to zero meaning that the distributions are almost degenerate and the corresponding covariance functions are very close to exponential covariance functions. Since the mean values of these distributions are very close to \(\hat s\) from the estimated exponential covariance function, all three estimated covariance functions are very close to the estimated exponential covariance function, thus leading to the same estimated model. This suggests possible problems with the estimation procedure, which we will explore in a simulation study in Section [6.3](#s:simstudy){reference-type="ref" reference="s:simstudy"}, or with unidentifiability of the parameters of the Bernstein CDF. To try out a model which do not become almost identical to the model using the exponential covariance function, we considered a covariance function with an inverse gamma Bernstein CDF with \(\tau=2\) fixed (corresponding to an inverse gamma distribution with an infinite variance). Table [4](#tab:neu-estimates){reference-type="ref" reference="tab:neu-estimates"} shows the minimum contrast estimate and Figure [\[fig:neu-pcf\]](#fig:neu-pcf){reference-type="ref" reference="fig:neu-pcf"} shows the corresponding estimated covariance function. This covariance function has a larger variance parameter \(\sigma^2\) and a heavier tail than the estimated exponential covariance function. The 95% global envelope in Figure [\[fig:neu-FGJ\]](#fig:neu-FGJ){reference-type="ref" reference="fig:neu-FGJ"} (right panel) and the \(p\)-value of \(0.243\) for the global envelope test reveal that the covariance function with inverse gamma density and fixed parameter \(\tau=2\) provides a similar good fit as the exponential covariance function. This observation is further discussed in Section [7.3](#s:7.3){reference-type="ref" reference="s:7.3"}.
## Simulation study {#s:simstudy}
The analysis of the dendrite dataset showed the problem that we in practice get an exponential covariance function when we fit the covariance functions given in Example [\[ex:0\]](#ex:0){reference-type="ref" reference="ex:0"}. It should be noted that all these covariance functions have the exponential covariance function as a limiting case, which makes it possible to get arbitrarily close to an exponential covariance function in the estimation procedure, while this was not the case when the \(\tau\) parameter was fixed in the analysis of the dendrite data. To explore whether this was a general problem or simply applied to the dendrite dataset, we made a number of simulations using covariance functions which were different than the exponential case to see if we still obtained exponential covariance functions from the estimation procedure. Specifically we took the network used in the dendrite dataset and simulated 1000 simulations of an ICP with a homogeneous intensity function and a covariance function with inverse gamma Bernstein CDF. We did this for various combinations of parameters, where \(\sigma^2\in\{10^{-1},10^0,10^1,10^2,10^3,10^4\}\) and \(\tau\in\{1.1,1.5,2,5\}\) while the rest of the parameters were given by \((\rho, h, \phi) = (1, 1, 0.02)\). The mean number of points is \(\mathbb E N(L) = \rho|L| = 736\) for all the chosen parameter settings. For each simulation we fitted two models using minimum constrast estimation (using the same values of \(a_1,a_2,q,p\) as in Section [6.2](#s:den-an){reference-type="ref" reference="s:den-an"}): an ICP with an inverse gamma Bernstein CDF with \(h\) and \(\tau\) fixed at their true values, and an ICP with an exponential covariance function with \(h=1\) fixed. For each simulation, we calculated the non-parametric estimate of the pcf and used \(D\) in [\[e:contrast\]](#e:contrast){reference-type="eqref" reference="e:contrast"} for calculating its distance to each of the two pcfs for the estimated models; using an obvious notation, these distances are denoted \(D_\text{Exp}\) and \(D_\text{IG}\). For each combination of parameters \(\sigma^2\) and \(\tau\), Table [5](#tab:comparison){reference-type="ref" reference="tab:comparison"} shows the percentage of simulations where \(D_\text{IG}>D_\text{Exp}\), i.e., the cases where the non-parametric estimate of the pcf resembles an exponential covariance function more than a covariance function with inverse gamma Bernstein CDF with \(\tau\) equal to the value used in the simulations. To quantify how far the ICP is from a Poisson process, the values of \(\varphi\) given by [\[e:varphiICP\]](#e:varphiICP){reference-type="eqref" reference="e:varphiICP"} and \(p_\textrm{ms}\) given by [\[e:pms\]](#e:pms){reference-type="eqref" reference="e:pms"} are also shown in Table [5](#tab:comparison){reference-type="ref" reference="tab:comparison"}. The row in the table corresponding to \(\sigma^2=10^{-1}\) contains values close to \(50\%\), which makes sense since \(g_0(0)-1=0.00416\) or \(p_\mathrm{ms}=0.912\) reveal that the simulated processes are almost Poisson. In the rest of the table the non-parametric estimate are typically closest to an exponentiel covariance function, showing that there indeed is a tendency for getting an estimated model corresponding to an exponential covariance function (although the values seem to decrease in the last row).
We also made similar simulation studies for LGCPs and PCPPs. These showed similar values to those in Table [5](#tab:comparison){reference-type="ref" reference="tab:comparison"} for the LGCP, while the values where significantly smaller for the PCPP (typically in the range \(30\%\)-\(40\%\) when \(\varphi=2\) is as large as possible in this model). The results of this section are further discussed in Section [7.3](#s:7.3){reference-type="ref" reference="s:7.3"}.
# Concluding remarks {#s:conclusion}
Our results and considerations on point processes on linear networks in the previous sections give rise to various conclusions and open research problems which we discuss in Sections [7.1](#s:7.1){reference-type="ref" reference="s:7.1"}--[7.4](#s:7.4){reference-type="ref" reference="s:7.4"} below.
## New point process models {#s:7.1}
mentioned the lack of repulsive models on linear networks. An interesting case is a determinantal point process (DPP) which to the best of our knowledge has yet not been investigated in connection to linear networks. For any given covariance function \(c\) on \(L\), a well-defined DPP will be specified by the density \[\label{e:dppdens} f(\{u_1...,u_n\})\propto \mathrm{det}\left(\{c(u_i,u_j)\}_{i,j=1...,n}\right)\] with respect to the unit rate Poisson process on \(L\). A DPP is a model for repulsiveness (e.g. \(g\le1\)) and it possesses a number of appealing properties. See (and the references therein) who considered DPPs with state space \(S=\mathbb R^k\) but many things easily are modified to the case \(S=L\). Several facts are interesting: If \(c\) is isotropic then the pcf of the DPP is isotropic, and e.g. \(c\) could be specified by one of the isotropic covariance functions given in Table [1](#tab:covariances-sphere){reference-type="ref" reference="tab:covariances-sphere"} and Example [\[ex:0\]](#ex:0){reference-type="ref" reference="ex:0"}. The normalizing constant which is omitted in [\[e:dppdens\]](#e:dppdens){reference-type="eqref" reference="e:dppdens"} can be expressed in terms of the eigenvalues of a spectral decomposition of \(c\). This spectral decomposition is also needed when specifying the \(n\)-th order intensity functions and an efficient simulation algorithm of the DPP. However, it is an open problem how to determine the eigenvalues and eigenfunctions of the spectral representation, in particular if \(c\) is required to be isotropic. Until this problem has been solved, we need to work with the unnormalized density (in the right hand side of [\[e:dppdens\]](#e:dppdens){reference-type="eqref" reference="e:dppdens"}) and to use MCMC methods for approximating the normalizing constant. The recent paper on a new class of Whittle--Matérn GPs on compact metric graphs is interesting for several reasons including the following. The model class is flexible and contains differentiable GPs; the precision matrix at the vertices can be quickly calculated; and Markov properties of the GP can be used for computationally efficient inference. Thus the Whittle--Matérn GPs may serve as an interesting alternative to the GPs considered in the present paper. However, we believe the Whittle--Matérn GPs are less useful when constructing Cox processes driven by transformed GPs because the Whittle--Matérn covariance function is not isotropic (which is a drawback for several reasons, cf. items (I)--(V) in Section [1](#s:intro){reference-type="ref" reference="s:intro"}). Specifically, the variance function of a Whittle--Matérn GP is not constant and used the geodesic (rather than the resistance) metric.
## The choice of metric for isotropic covariance and pair correlation functions {#s:7.2}
Isotropic covariance functions are available for larger classes of linear networks with respect to the resistance metric than the geodesic metric, cf. Theorem [\[t:1\]](#t:1){reference-type="ref" reference="t:1"}. Therefore, we are able to obtain LGCPs, ICPs, and PCPPs for larger classes of linear networks with respect to the resistance metric than the geodesic metric. Suppose \(L\) is not a tree and that both \(c_0(d_{\mathcal G}(u,v))\) and \(c_0(d_{\mathcal R}(u,v))\) are well-defined isotropic covariance functions. (Recall that \(d_{\mathcal G}\ge d_{\mathcal R}\) with equality if and only if \(L\) is a tree, cf. (C) in Theorem [\[t:res-m\]](#t:res-m){reference-type="ref" reference="t:res-m"}.) Suppose also that \(c_0> 1\) and \(c_0\) is a decreasing function, so \(c_0(d_{\mathcal G}(u,v))\le c_0(d_{\mathcal R}(u,v))\). This is typical for the covariance function of the GP underlying a LGCP, ICP, or PCPP (an exception is a LGCP if \(c_0\) can be negative but usually \(c_0\ge0\)). Consider the two LGCPs, ICPs, or PCPPs obtained by using the same type of transformed isotropic GP (or GPs) with the same mean and \(c_0\)-function but using the different metrics \(d=d_{\mathcal G}\) and \(d=d_{\mathcal R}\). Using an obvious notation, let us denote these two Cox processes by \(X_{\mathcal G}\) and \(X_{\mathcal R}\) and their corresponding pair correlation functions by \(g_{\mathcal G}(u,v)=g_0(d_{\mathcal G}(u,v))\) and \(g_{\mathcal R}(u,v)=g_0(d_{\mathcal R}(u,v))\). It follows from [\[e:pairLGCP\]](#e:pairLGCP){reference-type="eqref" reference="e:pairLGCP"}, [\[e:g0ICP\]](#e:g0ICP){reference-type="eqref" reference="e:g0ICP"}, and [\[e:pairPCPP\]](#e:pairPCPP){reference-type="eqref" reference="e:pairPCPP"} that \(g_{\mathcal G}(u,v)\le g_{\mathcal R}(u,v)\), indicating that \(X_{\mathcal R}\) is able to model a higher degree of clustering than \(X_{\mathcal G}\). On the other hand, for a DPP, we have always that \(g_0\le 1\), and typically \(g_0< 1\) and \(g_0\) is an increasing function. So for two such DPPs with the same \(g_0\)-function but given by the different metrics \(d_{\mathcal G}\) and \(d_{\mathcal R}\), respectively, the DPP using \(d_{\mathcal R}\) is able to a higher degree of repulsiveness than the DPP using \(d_{\mathcal G}\). In conclusion, using the resistance metric rather than the geodesic metric allows more well-defined LGCP, ICP, and PCPP models as well as more flexibility for modelling clustering and repulsiveness.
## Estimation {#s:7.3}
For parameter estimation in this paper we used minimum contrast estimation based on the pcf which worked well when we used an exponential covariance function, but this approach was not able to identify the parameters in the covariance functions given by the Bernstein CDFs in Example [\[ex:0\]](#ex:0){reference-type="ref" reference="ex:0"} unless we fixed a subparameter, cf. Sections [6.2](#s:den-an){reference-type="ref" reference="s:den-an"} and [6.3](#s:simstudy){reference-type="ref" reference="s:simstudy"}. We are not sure if this is a problem of unidentifiability or caused by the choice of estimation procedure for the following reasons. On the one hand, concluded that for the dendrite data minimum contrast estimation based on the pcf performs better than if it is based on the \(K\)-function, and the alternative method based on composite likelihood estimation is less reliable than minimum contrast estimation for the exponential covariance function, thus suggesting that this approach will not improve estimation for more complex covariance functions either. This is also in line with the comments in on estimation procedures. In particular, concluded that Neyman-Scott point process models are poorly identified under weak clustering, irrespective of the model parametrization, and this is not due to faults in currently available fitting methods (they considered the Euclidean state space case but we do not expect the choice of space to be of importance for this conclusion). On the other hand, if the problem is unidentifiability, one idea for estimating the parameters of the Bernstein CDFs could be to to use shrinkage estimators involving a penalty on cluster scale as was done in using the parameter \(\varphi\) in [\[e:varphi\]](#e:varphi){reference-type="eqref" reference="e:varphi"}. However, we also remade the simulation studies behind Table [5](#tab:comparison){reference-type="ref" reference="tab:comparison"} in the case where we fixed the parameter \(\sigma^2\) at its true value, which did not improve the estimation of the parameters in the Bernstein CDF, thus suggesting this approach might not work. A more careful study of whether the problem is the choice of estimation procedure or unidentifiability, extending the methods of to our setup, is left for future research. More ambitiously than minimum contrast, composite likelihood, and other moment based estimation procedures, future work may also investigate likelihood-based inference for parametric point process models on linear networks. This could include the adaptation of missing data MCMC methods for maximum likelihood and Bayesian approaches to linear networks.
## Miscellaneous {#s:7.4}
The special results in Section [5.2.1](#s:LGCP){reference-type="ref" reference="s:LGCP"} for LGCPs and any choice of the state space \(S\in\{\mathbb R^k,\mathbb S^k,L\}\) could possibly be exploited for developing techniques of model fitting or checking, where we have the following results in mind and assume the covariance function of the underlying GP is isotropic. The \(K\)-functions based on \(X\) and \(X_u\) agree, so if \(X\) is observed within a bounded window \(W\subseteq S\), empirical estimators \(\hat K\) and \(\hat K_u\) for \(u\in X_W\) are expected to be close. Moreover, the special structure in [\[e:n-LGCP\]](#e:n-LGCP){reference-type="eqref" reference="e:n-LGCP"} implies that \[\frac{\rho(u_1,u_2,u_3)}{\rho(u_1)\rho(u_2)\rho(u_3)}=g_0(d(u_1,u_2))g_0(d(u_1,u_3))g_0(d(u_2,u_3)).\] This structure was exploited in to construct a third-order moment characteristic which is useful for model checking when \(S=\mathbb R^k\), but it remains to consider the case \(S\in\{\mathbb S^k,L\}\). We have given various references for non-parametric estimation of the intensity, pair correlation, and \(K\)-functions on linear networks, cf. Sections [4.2](#s:3.2){reference-type="ref" reference="s:3.2"} and [4.2.2](#s:3.2.3){reference-type="ref" reference="s:3.2.3"}. In the inhomogeneous case these estimators depend 'locally' on the intensity function at the individual observed points. Recently, when \(S=\mathbb R^k\), demonstrated the advantages of introducing new global estimators over the existing local estimators. Again it remains to consider the case \(S\in\{\mathbb S^k,L\}\). | {'timestamp': '2022-12-19T02:10:07', 'yymm': '2212', 'arxiv_id': '2212.08402', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08402'} |
null | null |
null | null |
null | null |
# Introduction {#Sec:1}
In the last few decades, quantum metrology has become an important area of research and its application in various arenas of physics has been delved into. In particular, it aims at improving the measurement precision of unknown parameters governing a physical system, over the "classical limit\", often called the shot-noise limit, attained without any "quantumness\" in the encoding. The classical limit can be surpassed with the aid of quantum resources , notably with entanglement . The advances in the discipline have far-reaching implications in many areas of physics, including high-resolution optical imaging , quantum thermometry , gravitational wave detection , magnetometers , etc. See also in this regard. The measurement precision of a parameter to be estimated, of a physical system, can be enhanced by incorporating some quantumness in the system like entanglement or squeezing , but the precision cannot be increased indefinitely. The minimum error of estimation is lower bounded by the well-known Cramér-Rao bound which is inversely proportional to the square root of the relevant Fisher information . The quantum version of the Cramér-Rao bound is derived using the quantum Fisher information which involves a maximization over all feasible measurements . The measurement precision in any measurement process evidently depends on the initial encoding of the probes and the final measurement strategy. For optimal choice of encoded probes and decoding scheme, the error in measurement can be minimized, and that minimum error is achievable by repeating the experiment a large number of times . Due to erroneous realization of the probes or the presence of impurities in the measuring device, there may arise fluctuations or deformations which can affect the efficiency of the measurement. Also, there may appear imperfections while tuning the circuit components of the measurement device. Such impurities and fluctuations give rise to effective disorder in the system parameters. Also, a quantum system is generally susceptible to noise from the surrounding environments. Noise coming from the environment usually causes a decoherence effect on the system during time evolution, which degrades the metrological efficiency and is regarded as one of the major challenges in entanglement-enhanced sensing . See also in this regard. While the presence of imperfections in a system in the form of noise or disorder are normally predicted to deteriorate the efficiency of quantum devices, or reduce the advantages provided by the ideal system over their classical counterparts, in some situations, it can boost the usefulness of the system . Also, in a metrological setting, the disadvantages provided by a decoherence effect can be overcome by using a Markovian collective dephasing channel under certain conditions , implementing non-Markovianity in the system or incorporating external fields and interaction between the system particles . Our aim in this paper is to study the effect of various imperfections present in a system in the form of a disorder, on the metrological precision of measuring a parameter of the system. We identify instances where disorder, more precisely, glassy disorder, provides an enhancement in the efficiency of estimation of a parameter, over the ideal system. We construct disorder models by introducing disordered parameters into the ideal one, where the disorder distributions are chosen from a clutch of paradigmatic continuous and discrete distributions. In such a disordered system, we find scenarios where the presence of disorder provides an improvement in the efficiency of estimation of an unknown system parameter, over the ideal situation. These situations of attaining better efficiency of measurement in presence of disorder over the ideal scenario can be termed as "disorder-induced enhancements of metrological precision\" or instances of "order from disorder\". The disordered system parameters considered are assumed to be "glassy\". Moreover, we find scenarios where for an ideal system, the best metrological precision is obtained for maximally entangled initial probes, whereas insertion of disorder into the system provides a double advantage: an enhancement of metrological precision over the ideal one, and lower entanglement content in the probe state. The rest of the paper is presented as follows. In Sec. [2](#Sec:2){reference-type="ref" reference="Sec:2"}, we briefly discuss certain aspects of quantum metrology. The concept of glassy disorder and disorder averaging is discussed in Sec. [3](#Sec:3){reference-type="ref" reference="Sec:3"}. The disorder distributions, that provide disorder-induced enhancement of metrological precision, are also presented in this section. Sections [4](#Sec:4){reference-type="ref" reference="Sec:4"} and [5](#Sec:5){reference-type="ref" reference="Sec:5"} constitute evidence of order from disorder scenarios for copies of uncorrelated and two-qubit maximally entangled initial probes, for different glassy disorder models. The reduction of the requirement of entanglement content of the initial probes in presence of glassy disorder is shown in Sec. [6](#sec:6){reference-type="ref" reference="sec:6"}. A conclusion is presented in Sec. [7](#Sec:7){reference-type="ref" reference="Sec:7"}.
# Estimation of an unknown parameter {#Sec:2}
The enhancement of measurement precision in the frequency estimation protocol using the concepts of quantum metrology has been studied some time back in . The problem initiated with a single-qubit system described by the Hamiltonian \(\mathcal{H}_1\), given by \[\label{H0} \mathcal{H}_1=-\hbar\omega \ket{1}\bra{1},\] where \(\omega\) is the frequency to be estimated. For an \(n\)-qubit system, the dynamics is governed by the \(n\)-qubit Hamiltonian, \[\begin{aligned}
\mathcal{H}_{n} &=&\mathcal{H}_1\otimes \mathcal{I}_2\otimes\mathcal{I}_3\cdots \otimes \mathcal{I}_{n}+\mathcal{I}_1\otimes \mathcal{H}_1\otimes\mathcal{I}_3\cdots \otimes \mathcal{I}_{n}\nonumber\\ &&\phantom{aji dhaner}+\cdots +\mathcal{I}_1\otimes\mathcal{I}_2\cdots \otimes\mathcal{I}_{n-1} \otimes \mathcal{H}_1. \end{aligned}\] For the evolution of \(n\)-qubit initial probes by the unitary \(U_n=e^{-\frac{i}{\hbar} \mathcal{H}_n t}\), the minimum standard deviation in the estimation of \(\omega\), for uncorrelated and copies of two-qubit maximally entangled probes are given by respectively \[\label{nl_prod} \Delta \tilde{\omega}_0^{(p)}(t) = \frac{1}{t\sqrt{\nu}} \quad \text{and} \quad \Delta \tilde{\omega}_0^{(e)}(t) = \frac{1}{2t\sqrt{\nu}}.\] The superscript "\((p)\)\" stands for the initial probes in which, each copy is a single-qubit state and "\((e)\)\" represents the same, containing copies of two-qubit maximally entangled states. \(\nu\) is the number of measurements done on the encoded probes. So, for the former case \(\nu=n\), and \(\nu=n/2\) for the latter. For both the cases, the uncertainty in measuring \(\omega\) is optimal with respect to best measurement strategy and best input states. From Eq. ([\[nl_prod\]](#nl_prod){reference-type="ref" reference="nl_prod"}), we can see that the minimum error in frequency estimation, using two copies of product inputs is \(\frac{1}{t\sqrt{2}}\). The same minimum error using a single copy of two-qubit maximally entangled state is \(\frac{1}{2t}\), which is \(< \frac{1}{t\sqrt{2}}\). So, using maximally entangled initial probes offers a benefit in measurement precision over the same for uncorrelated initial probes. This enhancement of metrological precision by introducing quantumness in the initial states is a key idea of quantum metrology. The minimum deviations in frequency estimation, given in Eq. ([\[nl_prod\]](#nl_prod){reference-type="ref" reference="nl_prod"}), are not due to imprecise measurements of \(\omega\) in an experiment. It is the lower bound of the deviation of the estimated parameter, which directly comes from the quantum Cramér-Rao bound and is attainable for a large number of repetitions of an experiment corresponding to an optimal strategy. Below we discuss the prescription of obtaining the quantum Cramér-Rao bound. Suppose an unknown parameter, \(\theta\), is encoded in a state, \(\rho(\theta)\), of a physical system. To estimate \(\theta\), a measurement of elements \(\{\Pi_x\}\) is done on the system, and let the probability distribution for a measurement outcome \(x\) be \(f(x|\theta)\). Now, let us take the outcome of a single measurement to be \(x_1\). Based on this outcome, the predicted value of \(\theta\) can be represented by the estimator function \(g(x_1)\) . For a large number of measurements, our prediction is assumed to be correct on an average for a fixed \(\theta\). Therefore, \[\langle g \rangle_\theta = \int dx f(x|\theta) g(x) = \theta.\] Such a \(g(x)\) is called an unbiased estimator. So, for a given \(\theta\), we get a distribution of the estimate \(g(x)\) with probability \(f(x|\theta)\). The variance of the estimator function, while estimating \(\theta\), is lower bounded by the Cramér-Rao inequality given by \[\label{ClCramerRao1} \Delta^2 \theta \geq \frac{1}{F(\theta)},\] where \(F(\theta)\) is the Fisher Information (FI) . Here we use the traditional notation \(\Delta^2 \theta\) for \(\Delta^2 g(x)|_\theta\) as the variance of the distribution of the estimator function \(g(x)\). The Fisher information for a single measurement is defined in terms of the probability distribution \(f(x|\theta)\), through the relation \[\begin{aligned}
F(\theta) &=& \int dx \big[\frac{\partial}{\partial \theta}\log f(x|\theta)\big]^2. \label{FI} \end{aligned}\] From the additivity of FI, for \(\nu\) measurements, we have the FI, \(F^{\nu}(\theta)=\nu F(\theta)\). Therefore for \(\nu\) measurements, ([\[ClCramerRao1\]](#ClCramerRao1){reference-type="ref" reference="ClCramerRao1"}) has the general form \[\label{ClCramerRao} \Delta^2 \theta \geq \frac{1}{\nu F(\theta)},\] which attains the equality for large \(\nu\). The choice of measurement strategy has been fixed so far. To make the FI independent of such a choice, one has to maximize the Fisher information with respect to all possible measurements. As a consequence of the maximization, the Cramér-Rao bound becomes \[\label{QCRbound} \Delta \theta \geq \frac{1}{\sqrt{\nu \big(\max_{\Pi_x} F(\theta)\big)}} = \frac{1}{\sqrt{ \nu F_Q(\theta)}}.\] \(F_Q\) is the quantum Fisher information, obtained by maximizing the Fisher information \(F(\theta)\) with respect to all possible \(\{ \Pi_x\}\), where \(f(x|\theta)=\text{Tr}[\Pi_x \rho(\theta)]\). Here \(\{ \Pi_x\}\) is a positive operator valued measurement and \(\rho(\theta)\) is the quantum state where \(\theta\) is enoded. The quantum Fisher information can also be directly calculated for a given \(\rho(\theta)\) as \[\label{QFI} F_Q(\theta)=\text{Tr}\Big[\rho(\theta) {L_s\big[ \rho(\theta) \big]}^2 \Big].\] \(L_s\big[ \rho(\theta) \big]\) is known as the symmetric logarithmic derivative (SLD) of \(\rho(\theta)\). If the eigen-decomposition of \(\rho(\theta)\) is \(\rho(\theta)=\sum_i \lambda_i(\theta) \ket{e_i(\theta)}\bra{e_i(\theta)}\), the SLD is given by \[L_s \big[ \rho(\theta) \big] = \sum_{i,j}\frac{2 \bra{e_i(\theta)} \frac{\partial \rho(\theta)}{\partial \theta} \ket{e_j(\theta)}}{\lambda_i(\theta)+\lambda_j(\theta)} \ket{e_i(\theta)}\bra{e_j(\theta)},\] with the sum being over all pairs of \(i\) and \(j\) for which \(\lambda_i(\theta)+\lambda_j(\theta)\ne 0\). The maximum of \(F(\theta)\) in ([\[FI\]](#FI){reference-type="ref" reference="FI"}) is obtained by considering the projective measurement in the eigenbasis of the SLD . For the parameter \(\theta\) encoded in a pure state \(\ket{\psi(\theta)}\), QFI simplifies to \[\label{pure_fq} F_Q(\theta)=4 \big[\braket{\dot{\psi}(\theta)|\dot{\psi}(\theta)}-|\braket{\dot{\psi}(\theta)|\psi(\theta)}|^2 \big],\] where \(\ket{\dot{\psi}({\theta})}\) refers to the derivative of \(\ket{\psi({\theta})}\) with respect to \(\theta\) . The minimum uncertainty in measuring a parameter, obtained so far, is optimal with respect to all possible measurements, but still depends on initial choice of inputs \(\rho_0^{\otimes \nu}\). So, an optimization over the initial states also has to be done in order to possibly increase the accuracy of the measurement . Hence, the minimum deviation in the estimation of \(\theta\) takes the form \[\Delta \theta_{opt} = \min_{\rho_0^{\otimes\nu}} \frac{1}{\sqrt{ \nu F_Q(\theta)}}. \label{opt}\] In some situations, the optimal state is easy to determine. For example, let the parameter \(\theta\) be encoded in the state through a unitary operator of the form \(U(\theta)=e^{-i\tilde{H}\theta}\), with \(\tilde{H}\) being the generator of the unitary, independent of \(\theta\). In this situation, it can be shown that the best input state for attaining the minimum error in \(\theta\) is of the form \[\label{best_input} \ket{\psi({\theta})}_{opt}=\big( \ket{\lambda_{\max}} + \ket{\lambda_{\min}} \big)/\sqrt{2},\] where \(\ket{\lambda_{\max}}\) and \(\ket{\lambda_{\min}}\) are the respective eigenvectors associated with the highest and lowest eigenvalues of the generator \(\tilde{H}\) . Additionally, for the unitary \(U(\theta,\phi)=e^{-i(\tilde{H}_1\theta+\tilde{H}_2\phi)}\), where \([\tilde{H}_1,\tilde{H}_2]=0\), while estimating \(\theta\) or \(\phi\) independently, the best choice of input states remains the same as in Eq. ([\[best_input\]](#best_input){reference-type="ref" reference="best_input"}), with the eigenvectors corresponding to \(\tilde{H}_1\) and \(\tilde{H}_2\) respectively. For a situation where \(\tilde{H}_1\) and \(\tilde{H_2}\) do not commute, the best input state cannot be obtained following this scheme. In this paper, we focus on the minimum error, as given by the right hand side of ([\[opt\]](#opt){reference-type="ref" reference="opt"}), in the estimation of a parameter, say \(\theta\), by obtaining the QFI directly from ([\[QFI\]](#QFI){reference-type="ref" reference="QFI"}) and choosing the optimal input states following the prescription around Eq. ([\[best_input\]](#best_input){reference-type="ref" reference="best_input"}). This actually presents the Fisher information-based lower bound for the deviation in estimation of an unknown parameter \(\theta\), where the measurement basis is taken to be the eigenbasis of SLD, which is the optimal choice of basis. In further discussions, we will only concentrate on this Fisher information-based lower bound, with optimal choice of basis and inputs, and this quantity will be denoted as \(\Delta \tilde{\theta}\), for the parameter \(\theta\).
# Glassy disordered models and disorder averaging {#Sec:3}
For acquiring better precision in metrological estimations, an experiment should be repeated a large number of times to attain the Cramér-Rao lower bound. The high number of repetitions of the measurements, in turn, can cause some fluctuations or imperfections in the system parameters which may drive the system some "distance\" away from its ideal nature. Also, as a consequence of the failure to construct an ideal experimental setup, disorder can appear in a system while tuning the parameters of the setup. Irrespective of the origin of the disorder, it may affect the efficiency of a measurement scheme, while estimating an unknown parameter. In this paper, we consider such scenarios, where some types of disorders are present in a system in the form of randomness, incorporated into the parameters of the system. We take the disorder parameter to be glassy, which means that the equilibration time of the disorder in the system is several orders of magnitude larger than the relevant observation time. This indicates that, for a specific realization of the disorder, the disordered parameters virtually do not alter during the time of investigation. They may change after a long time, but that time span is not in our study of interest . Such systems are usually referred to in the literature as "glassy\" . They have also been called "quenched\" disordered systems.
## Models of disorder {#sec:3-a}
We first consider the insertion of a disorder parameter, into a single qubit system described by the Hamiltonian \(\mathcal{H}_1\). We incorporate the disorder parameter in two ways into the "ideal\" Hamiltonian. The altered forms of the Hamiltonian, with disorder, are taken as \[\begin{aligned}
&&H_1 =-\hbar\omega(1+\epsilon)\ketbra{1}{1} \\ &&\text{and} \quad H_2 =-\hbar(\omega+\epsilon)\ketbra{1}{1}, \end{aligned}\] where the disorder is modeled by the parameter \(\epsilon\), selected randomly from several paradigmatic distributions, both discrete and continuous. We will find that the metrological estimation provides order from disorder for certain cases.\
\
**\(\mathbf{\;\;D_1}\): *a discrete disorder.*** For this instance, the disorder parameter \(\epsilon\) is distributed as \[\begin{aligned}
\label{eq:D1} P(\epsilon) &=& \frac{(1-p)}{2}\quad \text{when} \quad \epsilon=\pm\alpha, \nonumber \\ &=& p \quad \quad \phantom{tum} \text{when} \quad \epsilon=0, \nonumber \\ &=& 0 \; \; \; \; \; \;\;\;\;\;\;\;\;\; \text{otherwise}, \end{aligned}\] with \(\alpha>0\) and \(0 \le p < 1\). The "strength\" of the disorder is ascertained by the standard deviation of the probability distribution, given by \[\sigma_{D_1}=\alpha \sqrt{1-p}.\]\
\
**\(\;\mathbf{D_2}\): *a discrete disorder.*** This is another discrete distribution which is somewhat different from \(D_1\). Here the \(\epsilon\) is distributed as \[\begin{aligned}
P(\epsilon) &=& \frac{1}{2} \quad \text{when} \quad \epsilon=\pm\alpha, \nonumber \\ &=& 0 \quad \; \text{otherwise}, \end{aligned}\] for \(\alpha > 0\). This distribution has a standard deviation \(\sigma_{D_2}=\alpha\), which provides the "strength\" of this disorder.\
\
**\(\;\;\mathbf{G_1}\): *a modulated Gaussian disorder.*** This is a continuous distribution with mean zero. The corresponding probability distribution function is given by \[\label{func1} P(\epsilon) = \frac{1}{A_1}|\epsilon^2-1|e^{-(\epsilon^2/a^2)} \;\;\;\;\;-\infty \le \epsilon \le \infty,\] where \[A_1 = \int_{-\infty}^{\infty} |\epsilon^2-1|e^{-(\epsilon^2/a^2)} d\epsilon,\] and \(a>0\). The "strength\" of this disorder is characterized by the standard deviation of the distribution, \[\begin{aligned}
\sigma_{G_1} &=& \frac{1}{2} \Big[a^3 \Big(2 \sqrt{\pi } (2-3 a^2)\, \text{erf}\big(\frac{1}{a}\big) \nonumber \\ &+& a e^{-\frac{1}{a^2}} \Big(\sqrt{\pi } \sqrt{\frac{1}{a^2}} e^{\frac{1}{a^2}} (3 a^2-2)+12\Big)\Big) \Big]^{1/2},\;\;\; \end{aligned}\] where \(\text{erf}(z)=\frac{2}{\sqrt{\pi}}\int_0^z e^{-y^2} dy.\)\
\
**\(\;\;\mathbf{G_2}\): *a modulated Gaussian disorder.*** Another continuous distribution with mean zero and a steeper gradient, in comparison to \(G_1\), is represented by the distribution function, \[\label{func2} P(\epsilon) = \frac{1}{A_2}|\epsilon^2-1|^2e^{-(\epsilon^2/a^2)} \;\;\;\;\;-\infty \le \epsilon \le \infty,\] where \[A_2 = \int_{-\infty}^{\infty} |\epsilon^2-1|^2 e^{-(\epsilon^2/a^2)} d\epsilon,\] and \(a>0\). The standard deviation in this case is \[\sigma_{G_2} = \Big[\frac{\sqrt{\pi }a^3}{8} \left(15 a^4-12 a^2+4\right)\Big]^{1/2}.\]\
\
**\(\;\;\mathbf{G_3}\): *a modulated Gaussian disorder.*** This represents a distribution sharper than \(G_2\). The disorder parameter \(\epsilon\) is, in this case, distributed as \[\label{func3} P(\epsilon) = \frac{1}{A_3}|\epsilon^2-1|^4e^{-(\epsilon^2/a^2)} \;\;\;\;\;-\infty \le \epsilon \le \infty,\] where \[A_3 = \int_{-\infty}^{\infty} |\epsilon^2-1|^4 e^{-(\epsilon^2/a^2)} d\epsilon,\] and \(a>0\). Again we define the "strength\" of the disorder as the standard deviation of the probability distribution, given by \[\sigma_{G_3} = \Big[ \frac{\sqrt{\pi }}{32 a^3} \left(945 a^8-840 a^6+360 a^4-96 a^2+16\right) \Big]^{1/2}.\] These five disorder models provide advantages in the precision of metrological estimation of the relevant quantities, over the same for the ideal scenario with uncorrelated initial probes, in case of evolution by the disordered Hamiltonian \(H_1\). The Gaussian and uniform disorder models are found to be non-beneficial in this case, but they provide significant enhancement in measurement precision by incorporating some external fields and interaction between the system particles, when copies of two-qubit states are taken as initial probes. These two distribution functions are also given below, for completeness.\
\
**\(\mathbf{\;\;G}\): *Gaussian disorder.*** In this scenario, \(\epsilon\) is chosen from a Gaussian distribution with a mean of zero and a standard deviation of \(\sigma_G\), and the corresponding probability density function is given by \[P(\epsilon) = \frac{1}{\sigma_G\sqrt{2\pi}} e^{-(\epsilon/\sigma_G\sqrt{2})^2}, \;\;\;\;-\infty \le \epsilon \le \infty.\] **\(\mathbf{\;\;U}\): *Uniform disorder.*** In this circumstance, \(\epsilon\) is distributed as \[\begin{aligned}
\label{uniform} P(\epsilon) &=& \frac{1}{s}, \;\;\;\; \text{when} \;\;\;\;-\frac{s}{2} \le \epsilon \le \frac{s}{2} \nonumber \\ &=& 0 \;\;\;\;\;\;\;\; \text{otherwise}. \end{aligned}\] The "strength\" of the disorder, quantified by the standard deviation of the distribution, is given by \(\sigma_U = \frac{s}{2\sqrt{3}}\).
## Disorder averaging
When the disorder parameters of a system are glassy, the relevant physical quantities are the glassy disorder-averaged versions of the corresponding quantities, which we will simply refer to as disorder-averaged quantities. For glassy disorder, the averaging must be done after the relevant physical quantity has been already computed. Suppose, we want to compute the Fisher information-based lower bound of the minimum uncertainty in measuring frequency, in presence of glassy disorder in the system. Then, we first need to evaluate the uncertainty in measuring frequency, \(\Delta \tilde{\omega}_{\epsilon}(t)\), encoded in a physical system, for an arbitrary but fixed time \(t\), and for an arbitrary but fixed configuration of the disorder. The disorder-averaged deviation in frequency is then given by \[\Delta \tilde{\omega}_{\mathcal{Q}}(t)=\int_{-\infty}^{\infty} \Delta \tilde{\omega}_{\epsilon}(t) P(\epsilon) d\epsilon.\] For discrete probability distributions, the integral is substituted by a summation. If this type of integral or summation cannot be addressed analytically, the averaging has to be handled by a numerical method as follows. We have to generate \(N\) instances of the disorder \(\epsilon\) Haar-uniformly, and let us assume that the instances are denoted as \(\epsilon_i\). So, we have \(\Delta \tilde{\omega}_{\epsilon_i}\) for each instance. Then we have to take the average of these \(N\) instances. This \(N\) should be large enough to reach the convergence with a certain precision. The disorder-averaged uncertainty in frequency estimation will take the form \[\Delta \tilde{\omega}_{\mathcal{Q}}(t)=\frac{1}{N} \sum_i^N \Delta\tilde{\omega}_{\epsilon_i}(t).\] This approach of determining the averaged quantities assumes that the integrals and summations converge to finite values. In presence of disorder, a system naturally undergoes inherent fluctuations. As a result, there may occur oscillations in the variance of the estimated parameter, about a mean value with a standard deviation. Therefore, if the "advantage", \(\beta=\Delta \tilde{\omega}_0(t)-\Delta \tilde{\omega}_{\mathcal{Q}}(t)\), of the disordered model over the ideal scenario in a parameter estimation protocol is greater than the standard deviation, \(\tilde{\sigma}\), in \(\Delta \tilde{\omega}_{\epsilon}(t)\), due to the disorder-induced distribution of the variance of the estimated parameter, then the situation can be referred to as a disorder-induced enhancement over the ideal situation. Here \(\Delta \tilde{\omega}_0(t)\) represents the Fisher information-based lower bound of minimum error in estimating \(\omega\) in an ideal scenario. In the paper, in all cases, we consider the relative advantage and the corresponding standard deviation, and rechristen them as \(\beta\) and \(\tilde{\sigma}\). This is attained by dividing the original quantities by \(\Delta \tilde{\omega}_0(t)\). In particular, this has the effect of making the derived quantities dimensionless. Suffixes of the quantities will be used to refer to the different disorder distributions used.
# Measurement precision in presence of glassy disorder for uncorrelated initial probes {#Sec:4}
We begin with finding the disorder-averaged minimum uncertainty in frequency \(\omega\), in presence of the disorder parameter \(\epsilon\), incorporated in the Hamiltonian \(H_1\). For a single-qubit system, with the encoding being effected by a unitary evolution with the unitary \(U_1=e^{-\frac{i}{\hbar}H_1 t}\), the optimal input state for measuring \(\omega\) is \(\ket{\psi_0}=(\ket{0}+\ket{1})/\sqrt{2}\), where \(\ket{0}\) and \(\ket{1}\) are eigenvectors of \(H_1\). The QFI evaluated using ([\[pure_fq\]](#pure_fq){reference-type="ref" reference="pure_fq"}) in this case is found to be \(F_{Q}^\epsilon=t^2(1+\epsilon)^2\). As FI is additive, we can extend this situation to \(\nu\) copies of \(\ket{\psi_0}\), \(\ket{\psi_0}^{\otimes \nu}\), and the disorder-averaged minimum deviation in the estimation of the frequency is given by \[\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}}(t)=\int_{\text{lb}}^{\text{ub}}\frac{1}{|1+\epsilon|t\sqrt{\nu}} \; P(\epsilon) d\epsilon,\] where \(\text{lb}\) and \(\text{ub}\) denote respectively the lower bound and upper bound of the corresponding probability distribution function from which \(\epsilon\) is chosen. We now compare the disorder-averaged quantities for the different probability distributions, discussed in Sec. [3.1](#sec:3-a){reference-type="ref" reference="sec:3-a"}, with the ideal minimum uncertainty in \(\omega\), i.e., \(\Delta \tilde{\omega}^{(p)}_0\), given in Eq. ([\[nl_prod\]](#nl_prod){reference-type="ref" reference="nl_prod"}).\
\
**Discrete glassy disorder of type \(\mathbf{D_1}\):** For the disorder parameter \(\epsilon\) chosen from the discrete distribution, \(D_1\), the disorder-averaged frequency deviation, for \(\nu\) copies of uncorrelated probes, is given by \[\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{D_1}}(t) = \frac{1}{t\sqrt{\nu}}\Big[\frac{1-p}{2}\Big(\frac{1}{|1-\alpha|}+\frac{1}{1+\alpha}\Big)+p \Big].\] At \(\alpha=1\), the function \(\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{D_1}}(t)\) diverges. We find that for \(\alpha \ge \frac{1+\sqrt{5}}{2} \approx 1.618\), the error in frequency in presence of glassy disorder of type \(D_1\), has an advantage over the ideal scenario i.e., \(\Delta \tilde{\omega}_{\mathcal{Q}_{D_1}}^{(p)}(t)<\Delta \tilde{\omega}^{(p)}_0(t)\), \(\forall t\). The blue surface of Fig. [\[1q_discrete1\]](#1q_discrete1){reference-type="ref" reference="1q_discrete1"}-(a) presents the relative advantage, \(\beta_{D_1}=\frac{\Delta \tilde{\omega}^{(p)}_0(t)-\Delta \tilde{\omega}_{\mathcal{Q}_{D_1}}^{(p)}(t)}{\Delta \tilde{\omega}^{(p)}_0(t)}\), with respect to \(\alpha\) and \(p\). \(\beta_{D_1}\) exhibits positive values, which signifies a possibility of obtaining an order from disorder situation for the discrete quenched disorder \(D_1\). But, for further assurance, we have to compare \(\beta_{D_1}\) with the standard deviation in \(\Delta \tilde{\omega}_{\epsilon_{D_1}}^{(p)}(t)/\Delta \tilde{\omega}_0^{(p)}(t)\). Suppose, the standard deviation is denoted as \(\tilde{\sigma}_{D_1}\). The behavior of \(\tilde{\sigma}_{D_1}\) is shown by the yellow surface of Fig. [\[1q_discrete1\]](#1q_discrete1){reference-type="ref" reference="1q_discrete1"}-(a), which is much smaller than \(\beta_{D_1}\) across a wide range of the \(\alpha-p\) space for \(\alpha \gtrsim 1.618\). This implies that we can have disorder-induced enhancements in efficiency of the frequency measurement for \(\alpha \gtrsim 1.618\) for the disorder model \(D_1\).\
\
\
**Discrete glassy disorder of type \(\mathbf{D_2}\):** If \(\epsilon\) is chosen randomly from the discrete distribution \(D_2\), the disorder-averaged minimum deviation in estimation of frequency, for \(\nu\) copies of uncorrelated probes, is given by \[\Delta \tilde{\omega}_{\mathcal{Q}_{D_2}}^{(p)}(t) = \frac{1}{2t\sqrt{\nu}}\Big[\frac{1}{|1-\alpha|}+\frac{1}{1+\alpha} \Big].\] Like in the case of \(D_1\), here also \(\Delta \tilde{\omega}_{\mathcal{Q}_{D_2}}^{(p)}(t)\) diverges for \(\alpha=1\) and the advantage, i.e., \(\Delta \tilde{\omega}_{\mathcal{Q}_{D_2}}^{(p)}(t)<\Delta \tilde{\omega}^{(p)}_0(t)\), is obtained only for the region, \(\alpha \ge \frac{1+\sqrt{5}}{2} \approx 1.618\) The behavior of the relative advantage, \(\beta_{D_2}=\frac{\Delta \tilde{\omega}^{(p)}_0(t)-\Delta \tilde{\omega}_{\mathcal{Q}_{D_2}}^{(p)}(t)}{\Delta \tilde{\omega}^{(p)}_0(t)}\), along with the standard deviation in \(\Delta \tilde{\omega}_{\epsilon_{D_2}}^{(p)}(t)/\Delta \tilde{\omega}_0^{(p)}(t)\), represented by \(\tilde{\sigma}_{D_2}\), with respect to \(\alpha\) are depicted in Fig [\[1q_discrete1\]](#1q_discrete1){reference-type="ref" reference="1q_discrete1"}-(b), for \(\alpha \gtrsim 1.618\). It should be noted that \(\tilde{\sigma}_{D_2}\) is significantly lower than \(\beta_{D_2}\). This suggests that there occurs the order from disorder effect using this distribution for \(\alpha \gtrsim 1.618\).\
\
**Modulated Gaussian glassy disorder of type \(\mathbf{G_1}\):** For the modulated Gaussian disorder of type \(G_1\), the minimum error in estimation of frequency for \(\nu\) uncorrelated initial probes is \[\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{G_1}} = \frac{1}{t\sqrt{\nu}}\frac{2a+e^{1/a^2}\sqrt{\pi}+2e^{1/a^2}\sqrt{\pi}\;\text{erf}(\frac{1}{a})-1}{\big[-4a + \sqrt{\pi} e^{1/a^2} (2-a^2)(1-2 \text{erf}(\frac{1}{a}))\big]}.\;\;\;\;\] From the nature of the quantity \(\frac{\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{G_1}}(t)}{\Delta\tilde{\omega}_0^{(p)}(t)}\), with the variation of \(a\), depicted in Fig. [\[modulated_gauss\]](#modulated_gauss){reference-type="ref" reference="modulated_gauss"}-(a) (solid purple line), it is evident that this glassy system allows \(\Delta \tilde{\omega}_{\mathcal{Q}_{G_1}}^{(p)}(t) < \Delta \tilde{\omega}_0^{(p)}(t)\). For this particular distribution, the standard deviation in \(\Delta\tilde{\omega}_{\epsilon_{G_1}}^{(p)}(t)/\Delta \tilde{\omega}_0^{(p)}(t)\) is non-existent.\
\
**Modulated Gaussian glassy disorder of type \(\mathbf{G_2}\):** For this glassy disorder situation, the disorder parameter, \(\epsilon\), is chosen from the distribution \(G_2\). The disorder-averaged frequency deviation for this case for \(\nu\) copies of initial uncorrelated probes is \[\Delta \tilde{\omega}_{\mathcal{Q}_{G_2}}^{(p)} = \frac{1}{t\sqrt{\nu}}\Big[\frac{4 \big[ (a+a^3)\frac{1}{\sqrt{e}}-\frac{1}{2}(a^2-2)\sqrt{\pi}\;\text{erf}(\frac{1}{a}) \big]}{(4-4a^2+3a^4)\sqrt{\pi}}\Big].\] In Fig. [\[modulated_gauss\]](#modulated_gauss){reference-type="ref" reference="modulated_gauss"}-(a), we present the quantity \(\frac{\Delta\tilde{\omega}^{(p)}_{\mathcal{Q}_{G_2}}(t)}{\Delta\tilde{\omega}_0^{(p)}(t)}\) with the green dashed line, which shows that this type of glassy disorder provide better precision in a larger region of \(a\) than the same of \(G_1\). Fig. [\[modulated_gauss\]](#modulated_gauss){reference-type="ref" reference="modulated_gauss"}-(b) shows that the relative advantage, \(\beta_{G_2}=\frac{\Delta\tilde{\omega}_0^{(p)}(t)-\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{G_2}}(t)}{\Delta\tilde{\omega}_0^{(p)}(t)}\), is much greater than \(\tilde{\sigma}_{G_2}\), which is the standard deviation in \(\Delta\tilde{\omega}^{(p)}_{\epsilon_{G_2}}(t)/\Delta\tilde{\omega}_0^{(p)}(t)\). So, the occurrence of an order from disorder situation is prominently present in this situation.\
\
**Modulated Gaussian glassy disorder of type \(\mathbf{G_3}\):** The disorder-averaged frequency deviation for \(\nu\) copies of initial uncorrelated probes is given in this case by \(\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{G_3}}\), whose analytical form is given in Appendix [\[appen1-natun\]](#appen1-natun){reference-type="ref" reference="appen1-natun"}. Here the range of the distribution parameter \(a\), providing an enhancement in frequency estimation, is better than the previous two continuous distributions, \(G_1\) and \(G_2\). Compare the three curves of Fig. [\[modulated_gauss\]](#modulated_gauss){reference-type="ref" reference="modulated_gauss"}-(a). To ensure that this advantage is a disorder-induced enhancement in precision of measurement, we have plotted the relative advantage, \(\beta_{G_3}=\frac{\Delta\tilde{\omega}_0^{(p)}(t)-\Delta\tilde{\omega}^{(p)}_{\mathcal{Q}_{G_3}}}{\Delta\tilde{\omega}_0^{(p)}}\), and the standard deviation of \(\Delta\tilde{\omega}^{(p)}_{\epsilon_{G_3}(t)}(t)/\Delta\tilde{\omega}_0^{(p)}(t)\), given by \(\tilde{\sigma}_{G_3}\), in Fig [\[modulated_gauss\]](#modulated_gauss){reference-type="ref" reference="modulated_gauss"}-(c). It can be seen that \(\beta_{G_3}\) is positive, and \(\tilde{\sigma}_{G_3}\) is substantially lower than \(\beta_{G_3}\). So here also, we attain the phenomenon of order from disorder.\
The disorder-averaged quantities for the Gaussian (\(G\)) and Cauchy-Lorentz (\(CL\)) (see Appendix [\[appen1\]](#appen1){reference-type="ref" reference="appen1"}) distributions, represented as \(\Delta \tilde{\omega}_{\mathcal{Q}_G}^{(p)}\) and \(\Delta \tilde{\omega}_{\mathcal{Q}_{CL}}^{(p)}\) respectively, do not converge. Due to this difficulty, we try truncated Gaussian (\(TG\)) and truncated Cauchy-Lorentz (\(TCL\)) distributions (distribution functions are given in Appendix [\[appen1\]](#appen1){reference-type="ref" reference="appen1"}). A truncated distribution is one, in which the domain of definition of the random variable is reduced to a finite range. If the disorder parameter \(\epsilon\) is chosen randomly from a truncated Gaussian or truncated Cauchy-Lorentz distribution, no benefit over the ideal scenario is achieved i.e., \(\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{TG}}(t)/\Delta \tilde{\omega}_0^{(p)}(t)>1\) and \(\Delta \tilde{\omega}^{(p)}_{\mathcal{Q}_{TCL}}(t)/\Delta \tilde{\omega}_0^{(p)}(t)>1\). For uniform quenched disorder (\(U\)) also, no disorder-induced enhancement is achieved. See Appendix [\[appen1\]](#appen1){reference-type="ref" reference="appen1"} for more discussions in this regard. An important point to be noted is that while an introduction of disorder parameter in the scale of energy, as in \(H_1\), leads to a definite change in the frequency estimation, there is no change in the same if we introduce the same disorder parameter in the form of an energy shift (see \(H_2\)). The reason is the following. For a unitary evolution with the unitary \(U_2=e^{-iH_2 t/\hbar}\), the variance in estimating \(\omega\) is given by \(\frac{1}{4\nu\Delta^2\mathcal{H}_1}\), where \(\mathcal{H}_1\) is the ideal-system Hamiltonian. Hence, in this case, the error is independent of \(\epsilon\), and so there is no effect, on metrological precision of frequency estimation, of having this type of disorder in the system.
# Measurement precision in presence of glassy disorder for copies of two-qubit maximally entangled initial probes {#Sec:5}
We now study whether there is an enhancement in the efficiency of measurement precision, in presence of glassy disordered parameters, for \(\nu\) copies of initial probes, of which, each copy is a two-qubit maximally entangled state. Moreover, we study the effect on the precision of incorporating interaction between the system particles and external fields in a disordered system. We consider a two-qubit system acted upon by the Hamiltonian, \[H_3=-\hbar\omega\big[(1+\epsilon_1)\ketbra{1}{1}\otimes \mathcal{I} + \mathcal{I} \otimes (1+\epsilon_2)\ketbra{1}{1}\big],\] where \(\epsilon_1\) and \(\epsilon_2\) are disorder parameters introduced in the ideal system. The optimal input state can be constructed using Eq. ([\[best_input\]](#best_input){reference-type="ref" reference="best_input"}) as \(\ket{\phi_0}=(\ket{00}+\ket{11})/\sqrt{2}\) and the encoded final state after evolving with the unitary \(U_3=e^{-i H_3 t}\), comes out to be \(\ket{\phi_f}=(\ket{00}+e^{-i(2+\epsilon_1+\epsilon_2)\omega t}\ket{11})/\sqrt{2}\). The quantum Fisher information for this case has the form, \(F_{Q}^{\epsilon_1,\epsilon_2}=t^2(2+\epsilon_1+\epsilon_2)^2\), obtained from ([\[pure_fq\]](#pure_fq){reference-type="ref" reference="pure_fq"}). The single-copy scenario of the two-qubit maximally entangled initial state can be extended to \(\nu\) copies of the same initial probes and the error in frequency for that case turns out to be \[\Delta \tilde{\omega}_{\mathcal{Q}}^{(e)}=\int_{\text{lb}_1}^{\text{ub}_1}\int_{\text{lb}_2}^{\text{ub}_2}\frac{1}{t\sqrt{\nu}|2+\epsilon_1+\epsilon_2|} \; P_1(\epsilon_1) P_2(\epsilon_2) d\epsilon_2 d\epsilon_1,\] where \(\text{lb}_i\) and \(\text{ub}_i\), for \(i=1,2\), refer to the lower and upper bounds of the probability distribution functions \(P_i(\epsilon_i)\) respectively. Here \(P_1(\epsilon_1)\) and \(P_2(\epsilon_2)\) are independent of each other. In absence of disorder, the deviation in frequency estimation is \(\Delta\tilde{\omega}_0^{(e)}=\frac{1}{2t\sqrt{\nu}}=\frac{1}{2}\Delta\tilde{\omega}_0^{(p)}\). Therefore to attain an advantage over the ideal case, we need \(\frac{\Delta\tilde{\omega}_{\mathcal{Q}}^{(e)}}{\Delta\tilde{\omega}_0^{(p)}} < \frac{1}{2}\). Following the line pursued in case of uncorrelated initial probes, we now investigate the effect of discrete disordered systems by choosing \(\epsilon_1\) and \(\epsilon_2\) randomly from the discrete disorder of type \(D_1\). The corresponding probability distribution functions \(P_i(\epsilon_i)\) are the same as in Eq. ([\[eq:D1\]](#eq:D1){reference-type="ref" reference="eq:D1"}), with the parameters \(\alpha_i\) and probability \(p_i\) for \(i=1,2\) respectively. The minimum deviation in estimation of \(\omega\) is given by \[\begin{aligned}
\Delta\tilde{\omega}_{\mathcal{Q}_{D_1}}^{(e)} &=&\frac{1}{t\sqrt{\nu}}\Big[\frac{1}{2} \sum_{i,j}p_i \left(1-p_j\right) \left(\frac{1}{\left| \alpha _j+2\right| }+\frac{1}{\left| 2-\alpha _j\right| }\right)\nonumber\\ &+&\frac{1}{4}\sum_{i,j}\Big\{p_i p_j+(1-p_i)(1-p_j)\Big(\frac{1}{|\alpha_i-\alpha_j+2|}\nonumber\\ &&\phantom{check korte}+\frac{1}{|(-1)^j(\alpha_i+\alpha_j)+2|}\Big)\Big\}\Big]. \end{aligned}\] Here \(i\) and \(j\) both run from \(1\) to \(2\). In Fig. [\[projection\]](#projection){reference-type="ref" reference="projection"}, we plot the projection of the dimensionless quantity \(\frac{\Delta\tilde{\omega}_{\mathcal{Q}_{D_1}}^{(e)}}{\Delta\tilde{\omega}_0^{(p)}}\), with respect to \(\alpha_1\) and \(\alpha_2\), and observe that there is a certain range of \(\alpha_1,\alpha_2\) for a fixed \(p_1\) and \(p_2\), in which there is an advantage in the measurement precision over the ideal scenario. The advantage is highest in two regions, viz. for small \(\alpha_1\) and high \(\alpha_2\), and for small \(\alpha_2\) and high \(\alpha_1\) (see the light blue parts of Fig. [\[projection\]](#projection){reference-type="ref" reference="projection"}). We have also compared the standard deviation in \(\frac{\Delta\tilde{\omega}_{{{(\epsilon_1)_{D_1} (\epsilon_2)_{D_1}}}}^{(e)}}{\Delta\tilde{\omega}_0^{(p)}}\), with the relative advantage, \(\frac{\Delta\tilde{\omega}_0^{(p)}-\Delta\tilde{\omega}_{\mathcal{Q}_{D_1}}^{(e)}}{\Delta\tilde{\omega}_0^{(p)}}\), and notice that there exists a region in the \((\alpha_1,\alpha_2)\)-plane for which the standard deviation is less than the relative advantage, which tells that in this case also, we obtain a disorder-induced enhancement in the efficiency of measurement over the ideal one. Similar to the single-qubit case, the truncated Cauchy-Lorentz \((TGL)\) and truncated Gaussian \((TG)\) distributions never give an advantage over the disorder-free scenario.
# In presence of Ising interaction between the probe qubits and introduction of external fields {#sec:6}
We now apply an external field on each of the qubits of a two-qubit single copy of the initial probe, in which the two qubits are also interacting through an Ising interaction. The disorder parameters are introduced in the coupling strengths of the interaction and the fields. The Hamiltonian of the composite system takes the form \[\begin{aligned}
H_4 &=& \overline{H}(\hat{A},\epsilon) + H_I, \;\;\;\; \text{where} \nonumber \\ H_I &=& \hbar Jk\big[ (1+\epsilon_1)\sigma_z^1 + (1+\epsilon_2)\sigma_z^2 \big] + \hbar J \epsilon_3 (\sigma_z^1\otimes\sigma_z^2), \nonumber \end{aligned}\] with \(k\) being a dimensionless constant and \(\epsilon_1\), \(\epsilon_2\) and \(\epsilon_3\) are the disorder parameters. \(\overline{H}(\hat{A},\epsilon)\) is an arbitrary hermitian operator which may be dependent on some other operators like \(\hat{A}\), or even some disorder parameter \(\epsilon\), but is independent of \(J\), and also \([\overline{H}(\hat{A},\epsilon),H_I] = 0\). The parameter \(J\) is encoded into the system by evolving each copy of the the initial states by the Hamiltonian \(H_4\) and the best input state is evaluated using Eq. ([\[best_input\]](#best_input){reference-type="ref" reference="best_input"}). Then the Fisher information-based lower bound, \(\Delta \tilde{J}_{\mathcal{Q}}^{(2)}\), in estimation of the parameter \(J\), in presence of various types of glassy disorders in the system, is evaluated numerically. Here "\((2)\)\" in the superscript is to remind us that the copies of two-qubit initial probes, which can be entangled or product, are used for encoding. An important point to be noted here is that while measuring the parameter \(J\) of a similar Hamiltonian containing only the interaction term, the best input state becomes a product state, instead of an maximally entangled one . In this situation also, the entanglement content of the initial probes depends on the relative values of disorder parameters. We will discuss about this matter after a while. For the interacting two-qubit probes, governed by \(H_4\), we study the effects of Gaussian and uniform disorder on the measurement precision of \(J\). We evaluate the relative advantage in estimated \(J\), viz. \(\beta^{\prime}_{X}=\frac{\Delta \tilde{J}_0^{(2)}-\Delta \tilde{J}_{\mathcal{Q}_X}^{(2)}}{\Delta \tilde{J}_0^{(2)}}\), for \(X\) being \(G\) or \(U\) (for Gaussian and uniform disorders respectively), and compare \(\beta^{\prime}_{X}\) with \(\sigma^{\prime}_X\), which is the standard deviation in \(\frac{\Delta \tilde{J}_{(\epsilon_1)_X(\epsilon_2)_X(\epsilon_3)_X}^{(2)}}{\Delta \tilde{J}_0^{(2)}}\). Here \(\Delta \tilde{J}_0^{(2)}\) is the error in \(J\), estimated in absence of any type of disorder in the interaction Hamiltonian \(H_I\), i.e., for \(\epsilon_1=\epsilon_2=\epsilon_3=0\). The standard deviation, \(\sigma_X^{\prime}\), for both \(X=G\) and \(U\), are significantly less than the corresponding \(\beta^{\prime}_{X}\). Precisely, \(\sigma_G^{\prime}\) oscillates around an average value \(\approx 0.33\), and the oscillation of \(\beta^{\prime}_{G}\) is about \(\approx 0.53\). For uniform distribution, \(\sigma_U^{\prime}\) fluctuates around \(\approx 0.27\), whereas the fluctuation of \(\beta^{\prime}_{U}\) is around \(\approx 0.44\). In these numerical investigations the strength of the Gaussian disorder, \(\sigma_G\), is taken to be \(\frac{1}{2}\) and for uniform distribution, \(\sigma_U=\frac{1}{2\sqrt{3}}\), which corresponds to \(s=1\). We also set \(J=1.5\) and \(k=0.08\). So there are instances of "order from disorder\" for both Gaussian and uniform glassy disorders. Interestingly, the presence of various glassy disorders in the system provide advantages over the ideal scenario, also with respect to the entanglement content of the initial states. It can be easily checked using the optimal input state, described around Eq. ([\[best_input\]](#best_input){reference-type="ref" reference="best_input"}), that for the estimation of \(J\), in absence of disorder in the fields and interaction term (\(\epsilon_1=\epsilon_2=\epsilon_3=0\)), the best choice of initial probes are maximally entangled states, viz. \(\ket{\phi_0}\). But, in presence of disorder, depending upon the relative values of \(\epsilon_1\), \(\epsilon_2\) and \(\epsilon_3\), the optimal choice of the initial states may become a product one. It is observed that, the best initial inputs become copies of \(\ket{0+}\) and \(\ket{+0}\), when \(\epsilon_3 < 1+\epsilon_1\) and \(\epsilon_3 < 1+\epsilon_2\), respectively. Here \(\ket{0}\) and \(\ket{1}\) are eigenstates of \(\sigma_z\), and \(\ket{+}=(\ket{0}+\ket{1})/\sqrt{2}\). So, the optimal choice of input state may switch from a maximally entangled state to a product one, in presence of disorder. The explicit form of the Hamiltonian \(H_I\) shows that in absence of disorder, the Ising interaction between the two qubits disappear. This interaction comes into play in presence of disorder, which can be considered to be responsible for this peculiarity in the entanglement content of the two-qubit inputs. Thus, we find a situation when, even with a lower entanglement content of the initial probes, a certain metrological advantage may be obtained in glassy disordered systems over the ideal one. We numerically investigate the average entanglement of the optimal input states and see that it is much less the same of the ideal scenario. Since we take into account two-qubit pure states, the measure of entanglement considered is the von Neumann entropy of the local subsystem . Precisely, the average entanglement for the Gaussian disorder is \(\approx 0.09\), and the same for uniform disorder is \(\approx 0.13\).
# Conclusion {#Sec:7}
In this paper, we examine the effect of glassy disorders present in the system, on the precision in estimation of system parameters. We consider copies of uncorrelated as well as two-qubit maximally entangled initial probes and identify instances where relative advantages in the efficiency of estimations are obtained in presence of glassy disorder in the system over the ideal cases. In looking for disorder-induced enhancement, we compare the disorder-averaged minimum Fisher information-based error with the corresponding quantity in the ideal, i.e., disorder-free limit. However, we also check whether the difference between the disorder-averaged and ideal cases is higher than the disorder-induced fluctuations in the disorder-average. We also find cases where a glassy disorder, present in the system, reduces the requirement of initial quantum correlation, entanglement, for attaining better precision in the estimation of a parameter over an ideal-disorder free-situation. We acknowledge computations performed using Armadillo , and QIClib on the cluster computing facility of Harish-Chandra Research Institute, India. We also acknowledge partial support from the Department of Science and Technology, Government of India through the QuEST grant (grant number DST/ICPS/QUST/Theme-3/2019/120). | {'timestamp': '2022-12-19T02:13:30', 'yymm': '2212', 'arxiv_id': '2212.08523', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08523'} |
null | null |
# One-Stage Cascade Refinement Networks
## All Scale Label Assignment with Pseudo-Boxes
Let \(F_{i} \in \mathbb{R}^{H \times W \times C}\) be the feature maps at layer \(i\) of a feature pyramid and \(s_i\) be the total down-sampling stride of this layer. A ground-truth bounding box is defined as \(B_\mathrm{g}=\left(c_x, c_y, h, w\right)\), where \(\left(c_x, c_y\right)\) denotes the center of the bounding box, and \(h, w\) are its height and width. To avoid using hyper-parameters related to anchor boxes, we choose a one-stage anchor-free method as the baseline, which generally regards the center (e.g., the center point or part) of the object as foreground to define positives. However, such center-based label assignment has an implicit assumption that the positive area inside the bounding box of size \(h \times w\) must be larger than the square of the stride \(s_i\), namely, \(h \times w \ge s_i^2\), so that at least one feature point would be labeled as positive. This is a valid assumption for generic object detection, but not for infrared small targets, whose sizes generally vary from \(2\times 2\) to \(12 \times 12\). For example, in FCOS, the stride \(s\) of the finest layer is \(8\). If \(hw < s^2\) for a given target, the bounding box of that target may not cover any feature points, leading to it being mislabeled and resulting in low detection performance, as shown in [\[fig:asap\]](#fig:asap){reference-type="ref" reference="fig:asap"}. This is because a large portion of infrared small objects in a certain scale are not being trained. This motivates us to *re-examine the label assignment for infrared small targets*. In fact, label assignment can be further divided into scale assignment and spatial assignment. In generic object detection, the scale assignment assigns targets of different sizes to the appropriate feature pyramid level for detection, and the spatial assignment then selects positive sample locations at the selected level (e.g., center sampling in FCOS ). However, infrared small targets do not exhibit large scale variation. Instead, many targets are missed if the stride of the prediction layer is larger than the target. Motivated by these observations, we propose an All-Scale Pseudo-Box (ASPB) based scheme to customize the label assignment for anchor-free infrared small target detection to improve detection performance. The main idea is to decouple the ground-truth target size from the label assignment. Specifically, for scale assignment, we no longer consider the small targets predicted by each scale as competitors, but rather as coarse-to-fine approximations under different computational budgets. To this end, we relax the scale constraints on the label assignment and treat all target boxes at all scales as positive samples. For the spatial assignment, we introduce a pseudo-box with the same center and size of \(p \times p\) (\(p \ge s\)) to replace the ground-truth bounding boxes of the tiny targets whose sizes are smaller than the feature map stride, *i.e.*, \(h \times w < s^2\). This allows us to assign positive labels to more feature points, improving detection performance for infrared small targets. Given a ground-truth bounding box \((c_x, c_y, h, w)\), the pseudo-box \(B_\mathrm{p}\) is defined as \[B_\mathrm{p}=\left\{\begin{array}{lc} (c_x, c_y, h, w), & \text { if } hw > p^2 \\ (c_x, c_y, p, p), & \text { otherwise } \end{array}\right.\] As illustrated in [\[fig:asap\]](#fig:asap){reference-type="ref" reference="fig:asap"}, when \(p = s\), at least one feature point will be labeled as positive and the target mislabeling issue is resolved. As \(p\) increases, *i.e.*, \(p > s\), more neighboring feature points will be labeled as positive samples, which further relieves the sample imbalance issue but at the cost of introducing more low-quality detected bounding boxes.
## Top-Down Cascade Refinement Head
The performance of a one-stage detector is also affected by the severe sample imbalance and inaccuracy of bounding box regression. The tricky problem is that the solutions to these two issues have opposite requirements in terms of scale. With our ASPB label assignment scheme, making predictions on the high-level feature maps of the feature pyramid is a straightforward way to solve the class imbalance problem because the number of negative samples will drop exponentially as the down-sampling stride increases. However, this solution is not ideal for improving bounding box regression accuracy because high-level feature maps have low spatial resolution and are not suitable for precise localization. Instead, because it is simpler to regress a local box than a remote one, the bounding box regression task favors low-level fine feature maps over high-level coarse feature maps. Simply dividing the classification and localization tasks into different scales, namely classifying the feature points at the high level and regressing the bounding box at the low level, is not a workable solution because the one-to-many mapping relationship between multi-scale feature points can lead to high rankings for low-quality bounding boxes, degrading the final detection performance. The proposed OSCAR Head is a solution to the issue of combining the efficiency of a one-stage detector with the merits of region proposal and cascaded bounding box refinement from two-stage detectors. The OSCAR Head consists of a high-level soft proposal head and a low-level refinement head, as shown in [\[fig:oscar-arch\]](#fig:oscar-arch){reference-type="ref" reference="fig:oscar-arch"}. The soft proposal head predicts a classification score \(\boldsymbol{p}^{h}{x, y}\) and a coarse bounding box \(\boldsymbol{b}^{h}{x, y}\) for each feature point \((x,y)\). These coarse boxes are then passed as initial guesses with top-down modulation scores to the refinement head, where they are further refined by predicting a category score \(\boldsymbol{p}^{l}{x, y}\) and a refined bounding box \(\boldsymbol{b}^{l}{x, y}\).
### Top-Down Score Modulation
To address the extreme imbalance between small targets and background during training, we propose a top-down score modulation mechanism as a soft region proposal module. This module is designed to improve the accuracy of the detector by using the more accurate score of the coarse layer to modulate the less accurate score of the fine layer. Compared to the serialized region proposal operation in two-stage detectors, which involves cropping feature maps and performing ROI Pooling, our soft region proposal can be implemented in a single forward pass and is therefore more efficient. This allows for parallel processing, which can improve the overall performance of the detector. Given a coarse feature map \(X^h\) of size \((H^h, W^h)\) and a fine feature map \(X^l\) of size \((h^l, w^l)\), both are the last feature maps before prediction in corresponding layers. The top-down score modulation weight is defined as follows: \[T(X^h) = \sigma(\mathrm{Conv}_2(\delta(\mathrm{Conv}_1(\mathrm{Up}(X^h)))))\] where \(\mathrm{Conv}_1\) and \(\mathrm{Conv}_2\) are convolutional layers whose kernel sizes are \(\frac{C}{4} \times C \times 3 \times 3\) and \(C \times \frac{C}{4} \times 3 \times 3\), respectively. The \(\delta\) denotes the Rectified Linear Unit (ReLU) and \(\sigma\) is a Sigmoid function. \(\mathrm{Up}(\cdot)\) denotes the up-sampling operation to interpolate the high-level feature maps to the same size of low-level feature maps. Then the modulated low-level classification scores can be obtained via \[S^l = \sigma(\mathrm{Conv}(T(X^h) \odot X^l))\] where \(\odot\) is the element-wise multiplication. By modulating the scores of the fine layer with the more accurate scores from the high-level features, the top-down modulation approach effectively eliminates false negatives and reduces the impact of class imbalance on the detector's performance.
### Top-Down Cascaded Regression
One-stage methods typically rely on a single step of regression using different feature layers to predict the locations and sizes of objects with varying scales. However, this approach can be ineffective in certain challenging scenarios, particularly for infrared small targets. To address this limitation, we propose a cascaded regression strategy for more accurate prediction of target locations and sizes. To improve the regression accuracy in the fine-level feature maps, we use the high-level layer to first predict coarse locations and sizes of underlying infrared small targets. This initial coarse prediction serves as a learnable anchor for the fine-level layer to further refine the predictions. The decoded coarse bounding box is denoted as \(B^{\mathrm{h}} = \left(x_0^{h}, y_0^{h}, x_1^{h}, y_1^{h}\right)\). To align with the next-level refined feature maps, we first need to up-sample the coarse bounding boxes. Here, we denote the interpolated bounding box as \(\tilde{B}^{\mathrm{h}} = \left(\tilde{x}_0^{h}, \tilde{y}_0^{h}, \tilde{x}_1^{h}, \tilde{y}_1^{h}\right)\). To further produce precise object locations and sizes, we feed both the coarse features and the coarse boxes to the refinement head. Formally, if a feature point is associated to a groundtruth box \(B_\mathrm{g}=\left(x_0, y_0, x_1, y_1\right)\), the regression targets for this feature point can be formulated as \[\begin{aligned} \Delta x_0 & =(x_0-\tilde{x}_0^{h})/s, \quad \Delta x_1=(x_1-\tilde{x}_1^{h})/s \\ \Delta y_0 & =(y_0-\tilde{y}_0^{h})/s, \quad \Delta y_1=(y_1-\tilde{y}_1^{h})/s. \end{aligned}\] where \(s\) is the stride of the low-level feature map. During inference, the regressed target \(\hat{B}^{l}\) is predicted via concatenating the up-sampled feature maps, denoted as \[\hat{B}^{l} = \mathrm{Conv}(\mathrm{Concat}(\mathrm{Up}(X^h), X^l))\] where \(\mathrm{Concat}\) is the concatenating operation and \(\mathrm{Up}\) denotes the feature map up-sampling. To some extent, the coarse boxes can be viewed as dynamic anchors.
## Normalized Contrast Prediction
The proposed ASPB label assignment method significantly improves the class imbalance issue. However, it also introduces many low-quality predicted bounding boxes by assigning positive labels to feature points that are far away from the center of an infrared small target. To suppress these low-quality bounding boxes, we propose to add a new branch to the refinement head that predicts the localization quality score of a detected bounding box. It is important to note that the calculation of center-ness requires that the feature point be within the ground-truth bounding box. However, for infrared small targets, this requirement may not always be satisfied. As shown in [\[fig:asap\]](#fig:asap){reference-type="ref" reference="fig:asap"}, when the target size is smaller than the feature map stride, the center-ness cannot be calculated. In contrast, our normalized contrast is calculated using the original input image, which can be thought of as a softened semantic segmentation mask. No matter how the feature map is down-sampled, the normalized contrast can always be calculated. **Loss Function.** The overall loss is defined as follows: \[\begin{aligned} L &=\frac{1}{N^{h}_{\text {pos }}} \sum_{x,y} L_{\mathrm{cls}}\left(\boldsymbol{p}^{h}_{x,y}, c_{x,y}^{h,*}\right) + \mathbbm{1}_{\left\{c_{x,y}^{h,*}>0\right\}} L_{\mathrm{reg}}\left(\boldsymbol{t}^{h}_{x,y}, \boldsymbol{t}_{x,y}^{h,*}\right)\\ & + \frac{1}{N^{l}_{\text {pos }}} \sum_{x,y} L_{\mathrm{cls}}\left(\boldsymbol{p}^{l}_{x,y}, c_{x,y}^{l,*}\right) + \mathbbm{1}_{\left\{c_{x,y}^{l,*}>0\right\}} L_{\mathrm{reg}}\left(\boldsymbol{t}^{l}_{x,y}, \boldsymbol{t}_{x,y}^{l,*}\right)\\ & + \lambda L_{\mathrm{noco}}\left(n^{l}_{x,y}, n_{x,y}^{l,*}\right) \end{aligned}\] where \(L_{\mathrm{cls}}\) is focal loss, \(L_{\mathrm{reg}}\) the IOU loss, and \(L_{\mathrm{noco}}\) is the quality focal loss as in. \(N_{\mathrm{pos}}\) denotes the number of positive samples and \(\lambda\) being 1 in this paper is the balance weight for the regression loss. \(\mathbbm{1}_{\left\{c_{i}^{*}>0\right\}}\) is the indicator function, being 1 if the \(c_{i}^{*}>0\) and 0 otherwise. **Inference.** During testing, the predicted normalized contrast is combined with the two-level classification scores to calculate the final score for each detected bounding box. As a result, bounding boxes that are far from the center of an infrared small target may be given lower scores due to the effect of the normalized contrast. Therefore, the non-maximum suppression is likely to filter out these low-quality bounding boxes, which can greatly improve the detection performance.
# Related Work {#sec:related-work}
## Infrared Small Target Detection
### Traditional Methods
Infrared small target detection is typically formulated as either a target enhancement problem or a background suppression problem. These methods can be further classified based on the modeling assumptions they use, such as local contrast approaches and sparse plus low-rank decomposition methods. Local contrast methods model infrared small targets as outliers in a slowly-changing background of highly correlated pixels. Different variants of local contrast measures have been proposed to enhance small targets and suppress pixel pulse noise and edges simultaneously. Sparse plus low-rank decomposition methods exploit the non-local self-correlation property of background patches by assuming that all background patches come from a single subspace or a mixture of low-rank subspace clusters. This allows target-background separation to be achieved by decomposing the original data into a sparse target image and a low-rank background image. However, in real-world scenarios, there are many distractors that stand out as outliers in the background alongside true targets. Traditional approaches lack the ability to differentiate between true targets and similar distractors at a high semantic level, because they only use grayscale values or low-level handcrafted features such as central differences and entropy. When given an infrared image without any targets, traditional approaches often yield false positive detections because the contrasts they predict are not semantically normalized. Another problem is that the contrast values of different targets in multi-target settings cannot be normalized to the same scale, which makes it more likely to miss targets with low contrast values and results in a high number of missed detections. These modeling issues can cause traditional methods to perform poorly when dealing with complicated scenarios.
### Deep Learning Methods
Infrared small target detection can now be modeled as a supervised machine-learning problem due to the availability of publicly accessible datasets like SIRST-V1. Currently, most studies focus on designing multi-scale feature fusion models that can produce more refined target representations because of the problem of insufficient intrinsic features. To exchange high-level semantics and low-level details, Dai *et al.* proposed an asymmetric contextual modulation module that incorporates a bottom-up modulation pathway based on point-wise channel attention and a top-down global contextual feedback. To enhance features that were matched with prior knowledge, Dai *et al.* proposed a specialized nonlinear feature transformation layer, which they embedded into a deep network. Meanwhile, Li *et al.* designed an interactive module that is densely nested, allowing for gradual interaction between high-level and low-level elements. Another significant challenge in infrared small target detection is achieving a balance between miss detections and false alarms. To address this, Wang *et al.* proposed splitting the task into two smaller tasks, each of which would be handled by an adversarially trained model that focuses on reducing either false alarms or miss detections. Although these methods significantly outperform traditional methods, formulating the detection problem as a semantic segmentation or image translation requires computationally expensive high-resolution feature maps to recover pixel-level details. In contrast to the pixel-wise segmentation methods mentioned above, we formulate the infrared small target detection as a bounding box regression task, which is less affected by labeling ambiguity. Furthermore, in the proposed OSCAR, the top-down modulation is applied to the detection head on the label space rather than the feature space in the neck, specifically the feature pyramid.
## Generic Object Detection
### Two-Stage Methods
The two-stage detector, such as Faster R-CNN, is a paradigm of object detection algorithm that first generates a set of region proposals and then classifies the objects within those proposals (see [\[fig:abbr-arch\]](#fig:abbr-arch){reference-type="ref" reference="fig:abbr-arch"} (a)). These algorithms can be conceptualized as cascades, where the first stage removes a significant number of background samples and the second stage classifies the remaining regions. In addition, Cascade R-CNN extends the two-stage paradigm by using multiple sub-networks to progressively improve the quality of the region proposals (see [\[fig:abbr-arch\]](#fig:abbr-arch){reference-type="ref" reference="fig:abbr-arch"} (b)). Our OSCAR follows the idea of a cascade refinement, but it differs in two important aspects: (1) Instead of using a sparse set of candidate object boxes for the next stage, OSCAR performs soft region proposal on dense predictions by down-weighting easy background samples; (2) Although it performs progressive refinement on the predicted results, OSCAR remains a one-stage detector, not a two-stage method, avoiding the additional computational overhead of general cascading pipelines.
### Single-Stage Methods
One-stage approaches, such as DenseBox and CenterNet, bypass the region proposal step and directly predict classification scores and bounding box regression offsets, aiming for real-time performance while maintaining high accuracy. There are some works that try to transplant the merits of two-stage methods to one-stage approaches. For example, RefineDet uses a refinement module to mimic the second regression of two-stage approaches (see [\[fig:abbr-arch\]](#fig:abbr-arch){reference-type="ref" reference="fig:abbr-arch"} (c)), and Consistent Optimization reduces the gap between the training and testing phases by tying consecutive classification targets to the regressed anchors. To some extent, our proposed OSCAR can be viewed as a continuation of this concept, but it differs in two ways: (1) OSCAR abandons the commonly used divide-and-conquer strategy, which assigns high-level feature maps to large objects and low-level feature maps to small objects, and instead treats multi-scale predictions as progressive approximations of the target. (2) to the best of our knowledge, this is the first time that top-down cascade bounding box refinement has been applied.
# Conclusion {#sec:conclusion}
We have proposed a comprehensive solution to the challenges of infrared small target detection, including a new research benchmark consisting of the SIRST-V2 dataset, the normalized contrast evaluation metric, and the DeepInfrared toolkit. Furthermore, our proposed one-stage cascade refinement network outperforms existing methods through three improvements, namely, an all-scale pseudo-box-based label assignment scheme, the use of the high-level head as soft proposals for the low-level refinement head, and a normalized contrast branch for better localization quality. Extensive experiments and analyses on the SIRST-V2 dataset further validate the effectiveness and efficiency of OSCAR. We hope it can serve as a simple yet effective baseline for the community.
# Introduction {#sec:introduction}
Infrared imaging systems have become critical optoelectronic components in surveillance systems due to their ability to detect objects at any time of day, their long range, and their ability to capture heat sources. Long-distance targets of interest often appear as small targets in the field of vision because of the requirements of early warning detection systems and the resolution of infrared imaging systems. As a result, infrared small target detection is essential for many applications and has been a focus of research in recent years. Since the seminal work of Miss Detection vs. False Alarm (MDvsFA) and Asymmetric Contextual Modulation Networks (ACMNet), many Convolutional Neural Network (CNN) based methods have been proposed and model the infrared small target detection as a semantic segmentation problem. Despite their success, it has been difficult to further improve the detection performance even with deeper backbones and more powerful attention modules. We argue that the performance bottleneck in infrared small target detection originates in the way the task is modeled as semantic segmentation, which is unreasonable for infrared small targets in three ways. First, for real-world applications, semantic segmentation is only an intermediary representation; the ultimate objective is to locate the centroids of infrared small targets so they can be used as input for the tracking task. Therefore, segmentation integrity is only a rough approximation of detection accuracy. Second, due to the dispersion effect of long-distance imaging, it is impossible for humans to label infrared small targets with pixel-level accuracy, and ground-truth annotations include inherent ambiguities. The impact of these uncertain pixels is significant because they make up the majority of positive samples and their loss dominates the training process. Third, in order to segment infrared small targets accurately, inference must be performed on high-resolution feature maps, making it computationally intensive. Therefore, considering both detection accuracy and computational overhead, it is better to model infrared small target detection as an object detection problem rather than a semantic segmentation problem. However, the performance of infrared small target detection still lags behind that of normal-scale object detection by a significant margin. The performance gap in infrared small target detection is mainly caused by three factors. First, current label assignment paradigms, whether anchor-based or anchor-free, tend to mistakenly label tiny groundtruth targets as background, leading detectors to pay less attention to small targets. Second, for infrared small targets, a small perturbation in the bounding box can cause a significant disturbance in the Intersection over Union (IoU) metric, making both bounding box regression and localization evaluation more challenging than for large objects. Third, collecting real-world infrared small target images is more challenging than collecting generic images, resulting in a lack of non-synthetic data to train machine learning models. In this paper, we propose a comprehensive solution to address the challenges faced in infrared small target detection. We first introduce a new research benchmark, consisting of (1) SIRST-V2, a dataset of real-world, high-resolution single-frame infrared small targets, (2) normalized contrast, a reliable measure of localization quality that accounts for the inherent label ambiguity in long-range imaging, and (3) DeepInfrared, an open source toolkit specifically designed for infrared small target detection. We hope that our benchmark and toolkit will facilitate the development of novel algorithms for this task. To address the performance gap in infrared small target detection, we also propose a new label assignment scheme called All-Scale Pseudo-Box (ASPB). This scheme decouples the ground-truth target size from the spatial assignment by using scale-adaptive pseudo-boxes and also relaxes the scale constraints by treating all target boxes at all scales as positive samples. This new scheme is motivated by the observation that infrared small targets, such as those with bounding boxes of \(3\times3 \sim 5\times5\), may not contain any feature points when using a common feature map stride, such as \(8\). This is a notable difference from generic object detection. However, infrared small target detection is actually a sparse target detection task with an occlusion-free prior and a known scale. In the highly structured feature pyramids, the multi-scale predictions can be viewed as multiple approximations of the target in a coarse-to-fine manner, which allows for more accurate localization. Finally, motivated by this cross-scale spatial correspondence, we propose the **O**ne-**S**tage **Ca**scade **R**efinement (OSCAR) network, a new data-driven framework for single-frame infrared small target detection that aims to perform cascaded bounding box regression while still maintaining the efficiency of one-stage methods. In contrast to generic object detection, which adopts a divide-and-conquer strategy to represent different scales of objectives with different layers, OSCAR reformulates a feature pyramid's multi-level predictions as approximations of cascaded soft region proposals in a coarse-to-fine order towards the same ground-truth bounding box. To address the class imbalance issue and down-weigh well-classified negative samples, a top-down cross-head modulation mechanism is used as a soft region proposal module. We also incorporate a normalized contrast branch in the refining head, which significantly improves performance by suppressing low-quality predicted bounding boxes caused by pseudo-boxes. Our main contributions in this work are:
1. We propose a real-world high-resolution dataset for infrared small target detection, called *SIRST-V2*.
2. We design *normalized contrast*, a novel evaluation metric for infrared small targets, which takes into account centroid localization and robustness under the ambiguities of long-range imaging.
3. We provide an open-source toolkit, called *DeepInfrared*, to support researchers in developing algorithms for the SIRST-V2 benchmark and normalized contrast metric.
4. We propose *OSCAR*, a new framework for single-frame infrared small target detection that aims to perform cascaded bounding box regression in a top-down cascade manner. By utilizing the conclusion-free prior in highly structured feature pyramids, our method achieves state-of-the-art detection performance while maintaining the efficiency of one-stage methods.
# Experiments {#sec:experiment}
## Experimental Settings
We have compared the OSCAR network to several model-driven methods for infrared small target detection. These methods include the method based on facet kernel and random walker (FKRW), LCM, MPCM, WLDM, infrared patch-image model (IPI), non-negative IPI model via partial sum minimization of singular values (NIPPS), reweighted infrared patch-tensor model (RIPT), and stable multi-subspace learning (SMSL). The detailed hyper-parameter settings for these methods are listed in [\[tab:params\]](#tab:params){reference-type="ref" reference="tab:params"}, which shows the parameters that were determined by an exhaustive search on the trainval set of the SIRSTv2 dataset.
### Reasonableness of Top-Down Refinement
We further investigate the reasons for the performance gains of top-down modulation in our OSCAR head. Specifically, we examine two questions: (1) Is it regression or classification that requires higher resolution feature maps more? (2) How do feature map resolutions affect classification and regression separately? To study how feature map resolution affects the classification sub-task, we employ a simple network called SimpleGrid, whose architecture is shown in [\[fig:reasonable\]](#fig:reasonable){reference-type="ref" reference="fig:reasonable"}(a). We endow it with the ability of oracle localization so that the performance can still be expressed in terms of mNoCoAP. Specifically, we extract the groundtruth coordinates and use them to identify the target centroid as long as the feature points are accurately classified. The label assignment scheme for the SimpleGrid is shown in [\[fig:reasonable\]](#fig:reasonable){reference-type="ref" reference="fig:reasonable"}(b). We divide an input image into a regular grid of equally sized patches. A patch is considered a positive sample if the target centroid lies within it. We keep the input image size constant, but use different down-sampling rates for the network to obtain the prediction results with different feature map sizes, as shown in [\[tab:downsample\]](#tab:downsample){reference-type="ref" reference="tab:downsample"}.
The results are shown in [\[fig:reasonable\]](#fig:reasonable){reference-type="ref" reference="fig:reasonable"}(c). As can be seen, with perfect oracle localization, the performance of SimpleGrid, which only needs to classify feature points, increases as the feature map resolution decreases. In contrast, the performance of the baseline, which performs both bounding box regression and classification, decreases as the down-sampling rate increases. This answers our previous question: it is bounding box regression, not classification, that requires high-resolution feature maps. In fact, the classification accuracy drops as the feature map size increases due to severe class imbalance. Based on these observations, we can draw the following conclusions. First, low-resolution feature maps are sufficient for preserving infrared small targets. Therefore, if the goal is to determine the presence or absence of a target in a certain region rather than its precise location, low-resolution feature maps are more appropriate. Second, since the high-level layers can classify feature points more accurately, they can be used to modulate the features of the lower layers, which is the source of the performance gains from our top-down modulation.
### Necessity of Joint Classification and Regression
In our previous ablation study, we adopted a top-down modulation approach to jointly predict the target category and location. This raises the question of whether the classification and bounding box regression sub-tasks can be directly conducted on different layers, using coarse high-level feature maps for classification and refined low-level feature maps for regression. To investigate this, we implemented a decoupled head, as shown in [\[fig:ablation-arch\]](#fig:ablation-arch){reference-type="ref" reference="fig:ablation-arch"} (b), and compared it to the anchor-free baseline. This allowed us to evaluate the necessity of joint classification and regression.
In [\[tab:decoupled\]](#tab:decoupled){reference-type="ref" reference="tab:decoupled"}, we can see that decoupling the classification and localization sub-tasks into different layers leads to a significant drop in the final detection performance, from 75.6% to 55.7%. We believe that this decline is due to the one-to-many relationship between high and low-level feature points, which allows low-quality bounding boxes predicted by redundant feature points to rank highly, thus significantly reducing the final average precision. This sharp decline highlights the importance of jointly predicting the target category and location.
### Impact of Top-Down Cascade Refinement
The proposed OSCAR model has two major differences compared to other cascaded detectors: top-down score modulation (TDSM) and top-down cascaded regression (TDCR). TDSM views the high-level predictions as coarse estimates of the infrared small targets in the low-level refinement head, while TDCR uses a multi-stage bounding box regression to predict the final bounding box more accurately than the single-stage method. In this part, we investigate the impact of TDSM and TDCR on the detection performance of OSCAR. The results are shown in [\[tab:oscar\]](#tab:oscar){reference-type="ref" reference="tab:oscar"}. As can be seen from the table, both TDCR and TDSM alone improve the detection performance of the baseline. Combining the two gives the best performance, improving the baseline from 75.6% to 79.2%. We attribute the performance gain from TDSM to the fact that more accurate classification information at higher levels suppresses the scores of target-like background interferences. The performance gain from TDCR comes from its ability to predict the final bounding box more accurately using a multi-stage approach.
### With or Without Normalized Contrast
In this part, we investigate the impact of the normalized contrast branch. As mentioned previously, this branch is designed to suppress the low-quality detected bounding boxes produced by feature points that are far from the infrared small targets. These false alarms are introduced by the pseudo-boxes in our ASPB label assignment scheme. The results of this investigation are shown in [\[tab:noco\]](#tab:noco){reference-type="ref" reference="tab:noco"}. In this table, \"None\" indicates that no normalized contrast is used, \"NoCo\" denotes that the normalized contrast is used, \"NoCo\(^\dag\)\" stands for that the normalized contrast is multiplied with the interpolated classification scores from the soft proposal head as the final score, and \"Centerness\(^\dag\)\" means that the centerness is used instead of the normalized contrast. As can be seen from the table, the augmented branch effectively suppresses the false alarms introduced by the pseudo-boxes, which improves the performance from 79.2% to 80.3%. Additionally, the normalized contrast branch achieves a higher score than the centerness, which is due to the fact that in our mNoCoAP metric, the centroid of an infrared small target is annotated with the pixel of the highest signal-to-clutter ratio, while the centerness is merely a sub-optimal proxy since the center pixel does not necessarily have the highest local contrast.
### Choice of backbone depth and prediction level
In generic computer vision, the detection performance is usually improved gradually as the depth of the network increases. However, this is not the case for infrared small targets. The features that highlight small objects tend to be extinguished because of the down-sampling operations in the backbone, which can result in the contamination of the features of infrared small targets by noise in the background in deep networks. In addition, different prediction levels result in very different computational effort. For example, the computational load of our OSCAR head on the P2 layer is 300% more than predicting on the P3 layer. Therefore, it is necessary to explore the optimal network depth and prediction level to balance the performance and efficiency. In [\[tab:depth\]](#tab:depth){reference-type="ref" reference="tab:depth"}, we provide the mNoCoAP values and GFLOPS of our OSCAR network on different backbones and prediction levels. Here, \(C\) is the output channel number of the feature pyramid and \(s\) is the feature map stride of the refinement layer. As shown in the table, the detection performance of the network with the same parameter settings decreases as the network depth increases. Therefore, the ResNet-18 is already sufficient for the infrared small target detection task, and it is more important to focus on how to more accurately regress targets than to use a more powerful backbone to improve detection performance. When we predict the target at a shallower layer, the detection performance is better, but at the cost of a 2-3 times higher computational effort. Finally, we can significantly reduce the computation on top of the P2 layer by reducing the number of channels in the head. As shown in the table, by halving the number of channels, we can achieve better performance (83.52% vs 82.77%) while maintaining less computation than the P3 layer prediction (25.36 vs 25.55 GFLOPS).
## Comparison with State-of-the-Arts {#subsec:sota}
In this subsection, we compare our OSCAR model with several model-driven and data-driven methods, including some state-of-the-art deep networks. The mNoCoAP values and inference times are listed in [\[tab:sota\]](#tab:sota){reference-type="ref" reference="tab:sota"}. The inference speed comparison is provided in terms of FPS. Traditional methods are evaluated on the original size of the images, while deep learning methods are evaluated on resized images with a specific width and height, *e.g.*, \((1000, 600)\).
### Model-driven versus Data-driven
It can be seen that data-driven methods outperform in terms of both detection performance and detection speed. This large performance gap shows that the global sparsity and contrast priors no longer hold for model-driven methods when dealing with complicated images, yet data-driven methods can still produce satisfying detection results. We attribute this performance gain to two main factors. First, data-driven methods use deep networks to extract high-level semantics from the images, which allows them to better distinguish real targets from false alarms that look similar to infrared small targets. Second, data-driven methods have fewer hyper-parameters than model-driven methods, making them more robust to variations in the scene. In addition, traditional methods, particularly low-rank approaches, require dozens of iterations to make predictions, whereas deep networks can make predictions in a single forward pass. This makes data-driven methods faster and more efficient. Furthermore, the computational complexity of traditional methods grows exponentially with the size of the input image, while data-driven methods maintain a linear relationship with the image size. As infrared imaging technology advances, infrared images of high resolutions, even higher than \(1280 \times 1024\) in SIRST-V2, will become more common in the future. Therefore, data-driven methods should receive more attention from researchers, both in terms of detection performance and detection speed.
:::
### Among Data-Driven Methods
Then we compare the proposed OSCAR method with other deep learning-based approaches. Our results show that: (1) Semantic segmentation approaches are outperformed by all bounding box regression methods. We believe the reason for this lies in the impact of uncertainty in the labels of pixels caused by the dispersion effect of long-distance imaging, which makes up the vast majority of positive samples and predominates the training loss. Since the objective of this task is to locate the centroids of infrared small targets rather than their entire contours, bounding box regression is a more reasonable detection pipeline for infrared small targets. (2) Our proposed OSCAR method outperforms FCOS by 7.2%. This performance gain demonstrates the effectiveness of our approach in achieving the primary goal of this paper, which is to improve detection results by regressing the target bounding box coordinates in a two-step cascade. (3) Compared to the two-stage Faster R-CNN, which also regresses the final coordinates based on initial guesses, our proposed OSCAR method achieves a much better result. We believe this is because OSCAR avoids using anchor-related hyper-parameters, which are known to be sensitive to the final detection performance. Additionally, Faster R-CNN assigns multiple samples per feature point, whereas OSCAR only assigns one, which makes the class imbalance more severe for Faster R-CNN and leads to inferior classification accuracy. In addition to outperforming FCOS, our proposed OSCAR method also outperforms both RFLA and QueryDet. It is important to note that, although both RFLA and QueryDet are designed for small objects, the small objects they assume still have a scale gap with the infrared small targets used in this study. RFLA is generally designed for objects with sizes ranging from \(16 \times 16\) to \(32 \times 32\), while OSCAR is typically used for objects smaller than \(12 \times 12\), mostly in the range of \(3 \times 3\) to \(5 \times 5\). Therefore, the poorer performance of RFLA is predictable because it predicts small targets in the P3 head at a stride of \(8\), which is too large for detecting infrared small targets. As for QueryDet, our OSCAR method differs in the way it utilizes more information from the high-level head. We believe this is the reason why OSCAR achieves better performance than QueryDet. In QueryDet, only the existence of small objects is passed down from the P3 head to the P2 head, whereas OSCAR fully uses the predictions (both classification and regressed bounding boxes) produced by the coarse head.
## Visual Analysis
To showcase the benefits of the proposed OSCAR, we provide visualizations of the experimental results in two groups, [\[fig:vis-target\]](#fig:vis-target){reference-type="ref" reference="fig:vis-target"} and [\[fig:vis-background\]](#fig:vis-background){reference-type="ref" reference="fig:vis-background"}, depending on whether the test images contain real targets. We use **real circles** to denote **True Positive (TP)** detections, **green circles** for **False Positive (FP)** detections, and **yellow circles** for **False Negative (FN)** detections. The better the performance, the more red circles there are and the fewer green and yellow circles there are. From [\[fig:vis-target\]](#fig:vis-target){reference-type="ref" reference="fig:vis-target"}, it can be seen that the traditional methods, such as FKRW, IPI, NIPPS, and SMSL, are able to detect the targets in the images, but may generate false alarms. In particular, the IPI method performed better because it did not generate false alarms, which is consistent with its good performance among traditional methods in the quantitative metrics, as shown in [\[tab:sota\]](#tab:sota){reference-type="ref" reference="tab:sota"}. It can be concluded that low-rank methods typically outperform local contrast methods in terms of detection accuracy. Among the deep learning approaches, OSCAR is able to detect real targets while suppressing false alarms. It should be noted that FCOS often misses the tiny real targets in the chosen test images. This is likely because FCOS's label assignment ignores real targets whose bounding boxes do not include any feature points, preventing the smallest scale targets from being trained. In contrast, our proposed OSCAR solves this problem by using the pseudo-box for label assignment and performs better in detecting very small scale targets. Overall, this visualization of the results highlights the importance of semantic understanding and careful label assignment for accurate infrared small target detection. As shown in [\[fig:vis-background\]](#fig:vis-background){reference-type="ref" reference="fig:vis-background"}, the second group of input images primarily consist of urban scenes, which are newly added in the SIRST-V2 dataset and contain many cranes, street lights, and other non-target background distractions that are similar in scale to the infrared small targets. This set of experiments focuses on the ability of different methods to suppress false alarms using semantic understanding of the images. The traditional methods shown in [\[fig:vis-background\]](#fig:vis-background){reference-type="ref" reference="fig:vis-background"} all produce a high number of false alarms, because in their models the detection of infrared small targets is simplified to the detection of salient regions of the image that match the scale of the targets. However, in complex backgrounds not everything that fits the model's appearance prior is a real target, and it is likely to be a background distractor. In contrast, deep learning methods can extract the semantics of the images to produce significantly fewer false alarms. The performance gap between traditional and deep learning methods is largely due to their ability to suppress false alarms. Additionally, OSCAR outperforms FCOS in terms of not generating any false alarms for these four images. We believe this is due to the usage of normalized contrast branch, which helps to suppress false alarms. In conclusion, our experiments demonstrate that the OSCAR network is a promising approach for detecting infrared small targets. It achieves superior performance on the SIRST-V2 dataset and is efficient in terms of both detection performance and inference speed. We hope that the proposed OSCAR network will serve as a reliable baseline for future research in this area.
# DeepInfrared Eco-System: Fundamental Building Blocks
# DeepInfrared Eco-System {#sec:ecosystem}
## SIRST-V2: Pay More Attention to Urban Scenarios
First, we introduce our new SIRST-V2 dataset, which is the largest highest-resolution dataset for single-frame infrared small target detection that we are aware of. It is a milestone update of our previous SIRST-V1 dataset and contains 1024 representative images, mostly at a resolution of \(1280 \times 1024\). These images have been extracted from real-world videos of various scenarios and provide a challenging testbed for infrared small target detection in complex real-world scenarios. [\[fig:dataset\]](#fig:dataset){reference-type="ref" reference="fig:dataset"} illustrates some representative images from our new SIRST-V2 dataset. In these images, it is evident that most infrared small targets are dim and difficult to distinguish from their backgrounds due to the presence of clutter. Additionally, the newly added urban scenes in SIRST-V2 contain many background interference elements such as cranes and street lights that are similar in appearance to the targets. This makes it challenging for traditional approaches that rely on target saliency or low-rank and sparse decomposition to effectively identify the targets in these complex scenarios. A more effective approach would require the use of a network with a high-level semantic understanding of the entire image to accurately distinguish targets from non-target interferences.
## Normalized Contrast: A Measure for Locating Accuracy
The IoU metric is highly skewed to large objects, very sensitive for infrared small targets due to their tiny size (see [\[fig:ambiguity\]](#fig:ambiguity){reference-type="ref" reference="fig:ambiguity"} (a)). A small perturbation in either predicted box or groundtruth box can cause a significant change in the IoU value, which leads to unreliable results. However, in real-world applications, the main objective is to accurately locate the centroids of the infrared small targets so that they can be used for tracking. Therefore, IoU is only an intermediate representation and does not accurately reflect the overall detection performance. As such, it is necessary to design a new evaluation metric specifically for this task. As far as we are concerned, a good localization quality measure for infrared small targets should have two key properties. First, it should be robust to label ambiguity, meaning that it can accurately reflect the location of the target even when the labels are not perfectly aligned with the target. Second, it should be able to directly reflect the accuracy of the target centroid, meaning that it should provide a clear indication of the precision of the target's location. In our opinion, if the issues of computational efficiency and semantic normalization are addressed, the local contrast map has the potential to be a good quality measure for infrared small targets, as it offers a soft distribution that is adaptable to the target's shape and appearance. Despite the numerous definitions of local contrast, their computational efficiency is very low, as can be seen in [\[tab:sota\]](#tab:sota){reference-type="ref" reference="tab:sota"} for local contrast measure (LCM), multi-scale patch-based contrast measure (MPCM) and weighted local difference measure (WLDM) in terms of frame per second (FPS). This is because these methods do not have groundtruth information, and therefore require complex nonlinear filtering operations to suppress background noise. In fact, due to the availability of bounding box information, using only the linear normalization of the target bounding box region can also generate high-quality local contrast maps. As a result, we propose to construct a new representation, called *normalized contrast*, by normalizing the local contrast map according to the bounding box annotation, as shown in [\[fig:noco-generation\]](#fig:noco-generation){reference-type="ref" reference="fig:noco-generation"}. Since the groundtruth box is known, we only need to calculate the normalized contrast for the bounding box area and its surroundings, and set the rest of the values to 0. Given a groundtruth box \(T = (c_x, c_y, h_t, w_t)\), we can obtain a larger region \(R = (c_x, c_y, h_t+2h_b, w_t+2w_b)\) by extending a specific number of pixels in all directions (\(h_b\) in Y-axis, \(w_b\) in X-axis). \(\mathcal{C}_i = R_i-\min(R)\) is the unnormalized local contrast. \(R_i\) denotes the grayscale of the \(i\)-th pixel in the region \(R\) and \(\min(R)\) indicates the minimum grayscale in region \(R\). Here we set \(h_t = \gamma h_b, w_t = \gamma w_b\), where \(\gamma \in (0, 1]\) is a constant. The normalized contrast is defined as \[n_i = \frac{C_i \odot G_\sigma-\min(C_i \odot G_\sigma)}{\max(C_i \odot G_\sigma)-\min(C_i \odot G_\sigma)}\] The Gaussian distribution, denoted by \(G_\sigma\), is used to impose a central preference on the map, which helps to suppress the background clutters in the surrounding region. The normalized contrast map depicts the normalized distance from a given location to the center of the infrared small target, providing a clear indication of the precision of the target's location. In other words, the groundtruth normalized contrast map is actually a look-up table. Given a center point of a predicted bounding box, we can obtain a normalized contrast value for evaluating the localization accuracy of the predicted bounding box. Note that our normalized contrast measure differs from the center-ness measure used in the FCOS in several key ways.
1. First, the centroid of infrared small targets is typically annotated with the pixel of the highest signal-to-clutter ratio, rather than the geometric center of bounding boxes. The requirement for centroid locating accuracy is often at the sub-pixel level (smaller than 1 pixel), and using the center of the bounding box as the centroid may result in an error of several pixels. To reflect this domain knowledge, the normalized contrast is designed to directly reflect the signal-to-clutter ratio of each pixel, which aligns with the definition of the centroid.
2. Second, the center-ness is purely a measure of coordinate distances, and a small perturbation of the bounding box can cause a significant disturbance. In contrast, our normalized contrast is a soft distribution adaptive to the appearance of infrared small targets. By replacing the delta distribution (step function) of the bounding box with a slowly-changing soft distribution, we are able to provide a more stable measure of target localization.
The mean Normalized Contrast Average Precision (mNoCoAP) is a metric that we propose to fairly and precisely assess the performance of algorithms for detecting infrared small targets. It is similar to the mean Average Precision (mAP) metric, which is widely used in generic object detection. The mNoCoAP is calculated based on the Average Precision, which is the area under the Precision-Recall curve. The key difference between the mAP and our mNoCoAP lies in the definition of the True Positive criterion. In the mNoCoAP, we replace the IoU with the NoCo value of the predicted bounding box. This means that a prediction will be classified as a True Positive if the predicted NoCo value is above a certain threshold. Following the common practice in the COCO dataset, the mNoCoAP is an average of nine levels of \(\delta\), which controls the desired centroid locating precision, ranging from 0.1 to 0.9 with a step size of 0.1 We have chosen the mNoCoAP as the main evaluation metric for our DeepInfrared benchmark because it offers several advantages over existing metrics.
1. First, the mNoCoAP is a uniform evaluation metric that supports fair and preference-free evaluation of the performance of different detection paradigms, such as bounding box regression, semantic segmentation, and background suppression.
2. Second, the mNoCoAP is different from the receiver operating characteristic (ROC) curve in that it can not only be used to evaluate the performance of producing as few false positives or false negatives as possible, but it can also be used to evaluate the centroid locating accuracy of the tested method. This is an important consideration when dealing with infrared small targets, as the precision of the target's location is often critical for successful detection.
3. Third, the mNoCoAP is more resistant to label ambiguity than other metrics because it softens the harsh truncated boundary of an ambiguous ground-truth bounding box. This is especially important in the case of long-range imaging, where label ambiguity can be a significant issue.
## DeepInfrared Toolkit: An Open-Source Benchmark
In the field of generic computer vision, there are many open-source object detection toolboxes, such as MMDetection and GluonCV, that implement and integrate state-of-the-art detection algorithms. These toolboxes are widely used by researchers to reproduce experiments and improve algorithms. However, in the case of infrared small targets, there is currently no open-source toolbox available. This lack of a common platform makes it difficult for researchers to reproduce experiments and compare different algorithms, which can hinder progress in this field. To this end, we have developed DeepInfrared, a deep learning toolkit specially developed for the infrared small target detection task based on PyTorch. Compared with its counterparts in generic computer vision, our DeepInfrared toolkit differs in the following two points:
1. DeepInfrared provides well-trained models, training scripts, and logs of state-of-the-art infrared small target detection algorithms. This allows researchers to easily reproduce experiments and compare different algorithms.
2. DeepInfrared offers more flexible choices for backbones and necks, and has implemented various dataset loaders, attention blocks, and data augmentation pipelines. It also includes evaluation metrics specifically designed for infrared small target detection, and supports multiple annotation formats, such as semantic segmentation, bounding boxes, and normalized contrast. | {'timestamp': '2023-01-03T02:04:12', 'yymm': '2212', 'arxiv_id': '2212.08472', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08472'} |
null | null |
null | null |
null | null |
# Introduction
Low-dimensional Heisenberg systems are a fertile testing ground which offer well-recognised model materials, exact analytical solutions, multiple experimental predictions, newly proposed quantum simulations and technological applications, such as quantum spin transistors. Many low-dimensional Heisenberg magnets are \(d^1\) V\(^{4+}\) spin\(-1/2\) (s\(-1/2\)) oxides. In these compounds, the crystal field environment, with a distinctive short V-O bond, lifts the degeneracy of the t\(_{2g}\) orbitals, hence quenching the orbital magnetic moment. The lowest-energy half-filled \(d_{xy}\) orbital lies in the plane perpendicular to the short bond, where it can overlap with oxygen orbitals, creating superexchange paths that give rise to a dominant planar antiferromagnetic interaction, and hence to low-dimensional spin systems. \(\beta\)-VOSO\(_4\) was first synthesized in 1928 by A. Sieverts and E. L. Müller by the reduction of V\(^{5+}\) in sulfuric acid. Later, in 2007, it was discovered among volcanic products of the Tolbachik volcano in Kamchatka, Russia. Vanadyl based compounds like \(\beta\)-VOSO\(_4\) and its hydro-oxosulfate modifications attract the interest of chemists and biologists due to their potential applications in catalysis and the management of diabetes. \(\beta\)-VOSO\(_4\) crystallizes in the \(Pnma\) orthorhombic group and its structure represents a three-dimensional framework built up of distorted V\(^{4+}\)O\(_{6}\) octahedra and (SO\(_{4}\))\(^{2-}\) tetrahedra. Recent crystallographic studies have revealed 2% of correlated defects in the here investigated synthetic crystals. These are linked to an inversion of the short-long V-O distance pairs and form thin layers with a negative inter-layer correlation, randomly destroying the alternation of the V-O bonding pattern along the \(b\)-axis. Fig. [\[fig:AtomicStruc\]](#fig:AtomicStruc){reference-type="ref" reference="fig:AtomicStruc"}(c) shows a representation of the crystal structure of \(\beta\)-VOSO\(_4\) following Ref. showing a \(98\%\) occupation of the vanadium site at \((0.334, 0.25, 0.2675)\). Previous studies of the magnetism in \(\beta\)-VOSO\(_4\) through magnetic susceptibility measurements report long-range antiferromagnetic order and ferrimagnetic behaviour. Here, we revise these previously published interpretations in light of data from newly grown mm-sized single crystals. We characterize the magnetic properties, through magnetic susceptibility, high field magnetization measurements and powder inelastic neutron scattering (INS). The INS data shows clear gapless spinons with a continuum spanning up to 12 meV. These data are compared to calculations of the spinon continuum of the one-dimensional (1D) Heisenberg chain. Reasonable agreement is found for \(J_1=3.83(2)\) meV, with the dominant exchange running along the \(b\)-direction within the exchange path predicted by our \(ab\)-initio calculations. We find the effect of interchain couplings to be negligible and confirm the system to be a 1D s-\(1/2\) Heisenberg chain magnet.
# Experimental and Calculation Details
Single crystal samples of \(\beta\)-VOSO\(_4\) were synthesized through the reaction: V\(_2\)O\(_5\)+2H\(_2\)SO\(_4\) \(\rightarrow\) 2\(\beta\)-VOSO\(_4\) + 2H\(_2\)O + 0.5O\(_2\), following Ref.. A solution of 3 g of V\(_2\)O\(_5\) and 100 ml H\(_2\)SO\(_4\) was heated up just below the boiling point of H\(_2\)SO\(_4\) at 290\(^\circ\)C in a conical flask for a maximum period of \(3\) months. A fractionating column was used to condense all vapors (see Fig. [\[fig:AtomicStruc\]](#fig:AtomicStruc){reference-type="ref" reference="fig:AtomicStruc"}(a)). Slow heating and cooling rates of 135\(^\circ\)C/h and 30\(^\circ\)C/h were used, respectively. All products as seen in Fig. [\[fig:AtomicStruc\]](#fig:AtomicStruc){reference-type="ref" reference="fig:AtomicStruc"}(b) were washed multiple times in an ice-water bath and dried at 110\(^\circ\)C for 12 hours. Thermogravimetric analysis was performed to ensure no water was left in the samples. These results (not shown here) were in agreement with those of Paufler \(et\) \(al.\),. Our longest reaction of 3 months allowed the growth of needle-like single crystals as long as 6 mm of a green colour. X-ray Laue diffraction (Fig. [\[fig:AtomicStruc\]](#fig:AtomicStruc){reference-type="ref" reference="fig:AtomicStruc"}(d)) was used to confirm the crystallinity of the samples. Powder samples were produced by grinding the single crystals. Indications of a small amount of the impurity phase VH\(_6\)SO\(_8\) (space group \(P2_1/c\)) were found in X-ray and neutron powder diffraction data (not shown here). This impurity is contained in the powder matrix where the crystals grow (not in the single crystals). The content varies in the powder sample but it amounts to less than \(1\%\) of the total mass. A single crystal sample of mass \(7.021\) mg was used to measure magnetic susceptibility using a Quantum Design Superconducting Quantum Interference Device magnetometer. These measurements were carried out in both zero-field and field-cooled modes with a magnetic field of 0.5T applied along the three main crystallographic directions. The temperature range for the measurements was 2-300K. \[might be a These measurements were carried out in both zero-field and field-cooled modes with a magnetic field of \(0.5\) T applied along the three main crystallographic directions. The temperature range for the measurements was \(2–300\) K. This magnetometer was also used to measure magnetization in a single crystal of mass \(27.97\) mg with magnetic fields up to 5 T applied along the \(a\)-axis at 4.2 K. This measurement was performed in order to calibrate the magnetic moment measured in the pulsed high magnetic fields using an induction method at the International MegaGauss Science Laboratory. This measurement was also performed at 4.2 K on the same sample and fields up to 66.3 T were achieved. Additionally, a home-made high sensitivity magnetometer was used to measure the magnetic susceptibility in the temperature range \(0.085\) K and \(2.5\) K with a magnetic field of \(0.5\) T applied along the \(c\)-axis. Inelastic neutron scattering (INS) experiments were performed on a polycrystalline sample of mass 2.614 g in an annular shape Al-foil holder, at 3 K using the direct geometry neutron ToF instrument NEAT at HZB. Measurements were performed with neutrons of incident wavelengths of 2, 3 and 5\(\rm{\AA}\) (incident energies \(20.45\) meV, \(9.09\) meV and \(3.27\) meV respectively). Density-functional (DFT) band-structure calculations were performed in the FPLO code using the generalized gradient approximation (GGA) for the exchange correlation potential. A dense **k**-mesh with up to 343 points in the symmetry-irreducible part of the Brillouin zone was used.
# Results and Calculations
## Static Susceptibility and Magnetization
The magnetic static susceptibility of \(\beta\)-VOSO\(_4\) with magnetic field of \(0.5\) T applied along the main three crystallographic axes is shown in Fig. [\[fig:susceptibility\]](#fig:susceptibility){reference-type="ref" reference="fig:susceptibility"}. The three curves are slightly different, suggesting a weakly anisotropic g-tensor. The Curie-Weiss temperatures and Curie constants obtained from a fit to the Curie-Weiss law above 150 K are listed in Table [\[table:suscept\]](#table:suscept){reference-type="ref" reference="table:suscept"}. The former temperatures are between \(-34.5\) K and \(-39.4\) K, suggesting dominant AFM interactions. Zero field cooled and field cooled measurements were done, and showed no sign of any hysteresis processes in the measured ranges (not shown here). All three susceptibility plots show a broad maximum centred at \(\approx29\) K. Such a susceptibility curve is characteristic of low dimensional magnets. The inset in Fig. [\[fig:susceptibility\]](#fig:susceptibility){reference-type="ref" reference="fig:susceptibility"} shows the magnetic susceptibility for magnetic fields applied along the \(c\)-axis and temperatures down to \(0.085\) K. No sign of a transition to a long range magnetic ordered state was detected within this experimental range. Assuming the hypothesis that \(\beta\)-VOSO\(_4\) is a s\(-1/2\) antiferromagnetic Heisenberg chain system, we use the expression calculated by Johnston, et al., (eqs. 30 and 31), for the characteristic temperature of the maximum of the magnetic susceptibility T\(_{\chi_{max}} = 0.640851|J_{1D}|/k_B\). From this expression, we can estimate an intra-chain interaction (\(J_{1D}\)) of \(3.9\) meV for fields along the \(a\) and \(c\)-directions and \(3.78\) meV for fields along the \(b\)-axis. Furthermore, by setting the \(g\) factors to those reported in Ref., we obtain \(\chi_{max}\)T\(_{\chi_{max}}=0.035477 g^2\frac{cm^3K}{mol}\) for fields along the \(a\)-axis, \(\chi_{max}\)T\(_{\chi_{max}}=0.037857 g^2\frac{cm^3K}{mol}\) for the \(b\)-axis and \(\chi_{max}\)T\(_{\chi_{max}}=0.035744 g^2\frac{cm^3K}{mol}\) for the \(c\)-axis. These values are very close to the theoretical value for a \(s-1/2\) AF uniform Heisenberg chain where \(\chi_{max}\)T\(_{\chi_{max}}=0.0353229(3) g^2\frac{cm^3K}{mol}\). Giving these similarities, we proceed to fit the data to the mean-field approximation for the magnetic susceptibility of a s\(-1/2\) uniform (\(\alpha=1\)) Heisenberg chain with an intra-chain interaction (\(J_{1D}\)), following eq. \(50\) in Ref.. The total fitting function also includes a Curie Weiss law term to account for the contribution of paramagnetic impurities \([C_{imp}/(T-\theta_{CW-imp})]\) which dominate the data below \(7\) K and a background susceptibility (\(\chi_0\)) to account for core diamagnetism and van Vleck contributions. All resultant parameters are listed in Table [\[table:suscept\]](#table:suscept){reference-type="ref" reference="table:suscept"}. The impurity Curie constant amounts to 1-1.45% of the \(\beta\)-VOSO\(_4\) Curie constant values. These paramagnetic impurities could be well linked to the 2% defect planes reported in Ref.. The g values were also extracted using the fit to eq. \(50\) in Ref.. These values are very close to 2 but are slightly bigger than those typically reported for V\(^{4+}\). In comparison the g-values extracted from the \(\chi_{max}\)T\(_{\chi_{max}}\) expressions are \({1.991, 1.947, 1.998}\) for fields parallel to each main crystallographic axis.
:::
The high field magnetization of \(\beta\)-VOSO\(_4\) at 4.2 K for fields applied along the \(a\)-axis is shown in Fig. [\[fig:magn\]](#fig:magn){reference-type="ref" reference="fig:magn"}. The magnetization increases with \(H\) following the function: \[\begin{aligned}
\label{eq:magn} M=M_{sat}(1-\sqrt{1-(H/H_{sat})}) \end{aligned}\] where \(M_{sat}\) is the saturation magnetization per V\(^{4+}\) ion, H is magnetic field and \(H_{sat}\) is the saturation field, \(68.46\) T, following the description given in Ref. . The red dotted line in Fig. [\[fig:magn\]](#fig:magn){reference-type="ref" reference="fig:magn"} corresponds to this equation with \(J_{1D}=4.107\) meV and \(g=2.006\) obtained by the susceptibility fit for the \(a\)-axis. No fitting procedure was done in the case of the magnetization modeling. The magnetic exchange interaction extracted from the susceptibility follows well the magnetization below \(20\) T after which it deviates given a much larger saturation field. The experimental saturation field is expected to be at \(70.7\) T just above the highest available magnetic field for this experiment. Calculation of the magnetization using the magnetic exchange interaction found in the analysis of the inelastic neutron scattering (as described in section 'Magnetic excitations') is also shown in the figure as a black dashed line. This value provides a much more accurate description of the magnetization in the whole magnetic field range.
## Ab-initio Calculations
To see which of the many possible exchange terms may be relevant in the case of \(\beta\)-VOSO\(_4\), we first perform ab-initio calculations, where we consider magnetic exchange couplings \(J_{ij}\) entering the spin Hamiltonian: \[\begin{aligned}
\label{eq:hamiltonian} \mathcal{H}= \sum_{\left\langle ij \right\rangle} J_{ij}S_iS_j. \end{aligned}\] with the summation running over lattice bonds \(\left\langle ij \right\rangle\). The exchange parameters \(J_{ij}\) were extracted using two complementary approaches. On one hand, we calculate hopping integrals between the V \(3d\) states using Wannier functions constructed for the uncorrelated (GGA) band structure, and introduce these hoppings into the Kugel-Khomskii model that then delivers the magnetic exchange couplings as follows, \[\begin{aligned}
\label{eq:KKexchange} J_{ij}= \frac{(4t_{ij}^{(nn)})^2}{U_{ {\rm eff}}}-\sum_m \frac{4(t_{ij}^{(nm)})^2 J_{ {\rm eff}}}{(U_{ {\rm eff}} + \Delta_m)(U_{ {\rm eff}}-J_{ {\rm eff}} + \Delta_m)} \end{aligned}\] where the first and second terms stand for the antiferromagnetic (\(J_{ij}^{AFM}\)) and ferromagnetic (\(J_{ij}^{FM}\)) contributions, respectively. Here, \(U_{\mathrm{eff}} = 4\) eV is the effective on-site Coulomb repulsion and \(J_{\mathrm{eff}} = 1\) eV is the effective Hund's coupling in the V \(3d\) shell. The hoppings \(t_{ij}^{(nn)}\) are between the half-filled states of vanadium, whereas \(t_{ij}^{(nm)}\) involve the empty states \(m\), and \(\Delta _m = E_m-E_n\) is the crystal-field splitting. Alternatively, we obtain the exchange couplings by a mapping procedure using total energies of collinear spin configurations evaluated within DFT+U, where correlation effects in the V \(3d\) shell are treated on the mean field level with the on-site Coulomb repulsion \(U_d = 4\) eV, Hund's coupling \(J_d = 1\) eV, and atomic-limit flavor of the double-counting correction. The DFT+U calculations were performed for the supercells doubled along the \(b\) and \(c\) directions without including the distortion reported in Ref.. Further tests were run using the reported crystallographic disorder but no changes were detected in any exchange interaction contradicting the interpretation of the low-temperature curie tail in the susceptibility data. In the absence of correlations, the GGA band structure of \(\beta\)-VOSO\(_4\) is metallic and features a broad V \(3d\) band that spans the energy range between \(-0.3\) and \(3.8\) eV. The lower part of this band is dominated by the half-filled \(d_{xy}\) states. Here, we define the local coordinate system in such a way that the \(z\)-axis points along the short (vanadyl) bond of the VO\(_{6}\) octahedron, whereas \(x\) and \(y\) point along the longer V--O bonds that are approximately perpendicular to it. Then, \(z\) is close to the \(a\) direction and thus nearly parallel to the structural chains in \(\beta\)-VOSO\(_4\). Only the \(d_{xy}\) states are half-filled, and all other \(3d\) states of vanadium are empty, as typical for V\(^{4+}\). Both ferro-and antiferromagnetic contributions to the exchange couplings obtained from eq.[\[eq:KKexchange\]](#eq:KKexchange){reference-type="ref" reference="eq:KKexchange"}, as well as total exchange couplings \(J_{ij}\) obtained by the mapping procedure, are listed and represented graphically in Fig. [\[fig:strucJs\]](#fig:strucJs){reference-type="ref" reference="fig:strucJs"}. Long-range interaction terms not included in this table are well below 1 K and can be safely neglected. The nearest-neighbour coupling along the structural chains (\(J_a\)) is relatively weak and ferromagnetic, while the strongest coupling, \(J_1\) \(\approx50\) K, runs between the VO\(_{6}\) octahedra that are linked via two SO\(_{4}\) tetrahedra (see Fig. [\[fig:strucJs\]](#fig:strucJs){reference-type="ref" reference="fig:strucJs"}). Such a coupling regime is common for V\(^{4+}\) phosphates, arsenates, and other compounds with tetrahedral polyanions. On the one hand, double tetrahedral bridges create a V--O\...O--V superexchange pathway that are reasonably efficient despite the large V--V separation of more than 5 \(\AA\). On the other hand, the position of the half-filled \(d_{xy}\) orbital in the plane perpendicular to the structural chain prevents the \(xy\)-\(xy\) hopping (and, thus, an antiferromagnetic interaction) for the shorter V--V separation along the chain and renders \(J_a\) weakly ferromagnetic. The remaining couplings \(J_2\) and \(J_{bc}\) are much weaker as each of them is mediated by single SO\(_{4}\) tetrahedron with a less efficient orbital overlap.
## Magnetic Excitations
The dynamical structure factors \(S(|Q|,\Delta E)\) of \(\beta\)-VOSO\(_4\) collected at \(3\) K with neutron wavelengths of \(2\) and \(3\) Å (\(E_i = 20.45,~9.09\) meV respectively) are shown in Figures [\[fig:inelastic_neutron\]](#fig:inelastic_neutron){reference-type="ref" reference="fig:inelastic_neutron"}(a) and (b). Magnetic scattering is clearly evident at low \(|Q|\), dispersing from \(|Q_0|\sim 1.05\) Å\(^{-1}\) and extending in energy up to around \(10\) meV \[Fig. [\[fig:inelastic_neutron\]](#fig:inelastic_neutron){reference-type="ref" reference="fig:inelastic_neutron"}(a)\], with a sharp peak at \(\Delta E \sim 6\) meV. Weaker magnetic intensity is also observed near the elastic line around \(|Q| \sim 0.6\) and \(2\) Å\(^{-1}\). Overall, the spectrum is reminiscent of powder data from other one-dimensional systems, like KTi(SO\(_4\))\(_2\), Sr\(_3\)CuPtO\(_6\) and KMoOP\(_2\)O\(_7\). The observed \(|Q_0|\) is close to the double-(SO\(_4\))\(^{2-}\)-bridged zig-zag chains that run along the \(b\) axis \[Fig. [\[fig:strucJs\]](#fig:strucJs){reference-type="ref" reference="fig:strucJs"}\] in agreement with the \(ab\)-initio calculations. To further confirm the hypothesis that the experimental data can be reproduced assuming one-dimensional zig-zag chains along \(b\), we compare our experimental data to the total dynamical structure factor \(S_{1d}(Q_y,\Delta E)\) for the 2-and 4-spinon continua calculated using the Bethe ansatz. To do this, the theoretical \(S_{1d}(Q,\Delta E)\) is first recalculated due to the strong deviation of the chains from linearity and the resulting folding of the Brillouin zone. From the orthorhombic symmetry of the material, the folded \(S(\mathbf{Q},\Delta E)\) is expressed as: \[\begin{aligned}
S(\mathbf{Q},&\Delta E) = \cos^2{(Q_x\delta_x)}\cos^2{(Q_z\delta_z)}S_{1d}(\mathbf{Q},\Delta E) \nonumber \\ +& \cos^2{(Q_x\delta_x)}\sin^2{(Q_z\delta_z)}S_{1d}(\mathbf{Q}+(110),\Delta E) \nonumber \\ +& \sin^2{(Q_x\delta_x)}\cos^2{(Q_z\delta_z)}S_{1d}{(\mathbf{Q}+(111)},\Delta E) \nonumber \\ +& \sin^2{(Q_x\delta_x)}\sin^2{(Q_z\delta_z)}S_{1d}(\mathbf{Q}+(011),\Delta E). \label{bz} \end{aligned}\] Here, \(S_{1d}(\mathbf{Q},\Delta E)\) represents the linear chain 2-and 4-spinon dynamical structure factor shifted by a \(\mathbf{Q}\)-dependent lattice vector from the antiferromagnetic zone center \((010)\), and \(\delta_x = 0.1615\) and \(\delta_z = 0.2353\) are the offsets of the V\(^{4+}\) position \((\delta_x,\frac{1}{4},\delta_z)\) from that which would generate a perfectly linear chain: \((0,\frac{1}{4},0)\). Following folding, the theoretical data is powder averaged, yielding \(S(|Q|,\Delta E)\), shown in Fig. [\[fig:inelastic_neutron\]](#fig:inelastic_neutron){reference-type="ref" reference="fig:inelastic_neutron"}(c). The correct zone-boundary energy, corresponding to the maximum in \(\Delta E\) at \(6\) meV, is recovered for an intra-chain coupling \(J_1 = 3.83(2)\) meV, and most details of the spectrum, including the weak intensities away from the main branch of the continuum are also reproduced (\(\chi^2 =\sum_{(Yobs-Ycalc)^2}= 5.3\)) for 6 constant-\(|Q|\) cuts through \(S(|Q|,\Delta E)\). On the other hand, the intensity at the zone-boundary energy is somewhat overestimated. This reduction in intensity is consistent with the presence of inter-chain couplings, as seen, for example, in the case of Cs\(_2\)Cu\(X_4\) (\(X\)=Cl, Br). This single \(J_1\) value was used to calculate the magnetization using the Fisher and Bonner expression as previously described. As neutron scattering measurements are not affected by the g-factor, the ESR literature g factor of 1.987 was used for this calculation. Fig. [\[fig:magn\]](#fig:magn){reference-type="ref" reference="fig:magn"} shows the calculation as a black dashed line, again no fitting procedure was used. This comparison shows that the a single magnetic intra-chain coupling of \(J_1 = 3.83(2)\) meV is a hihgly accurate value for \(\beta\)-VOSO\(_4\). The predictions from our *ab-initio* calculations suggest that the dominant exchange beyond \(J_1\) is \(J_a\), while \(J_{bc}\), and \(J_2\) are somewhat smaller. Therefore, we performed fits of the dynamical structure factor \(S(|Q|,\Delta E)\) for the \(J_1\)-\(J_a\) model. This was done following the random phase approximation-style approach in, from which \(S(\mathbf{Q},\Delta E)\) is given by the expression: \[\begin{aligned}
\frac{S_{1d}(Q_y,\Delta E)}{\left[1+J^{\prime}(\mathbf{Q})\chi^{\prime}(Q_y,\Delta E)\right]^2+\left[J^{\prime}(\mathbf{Q})\chi^{\prime\prime}(Q_y,\Delta E)\right]^2}, \label{eq:rpa} \end{aligned}\] where \(S_{1d}(Q_y,\Delta E)\) is defined as above and \[\begin{aligned}
J^{\prime}(\mathbf{Q})&\simeq J_a \cos{(Q_x/2)}\cos{(Q_z/2)}, \end{aligned}\] where the Fourier components for the further neighbour couplings correspond to that of the straight chain. It is necessary to normalize \(S_{1d}(Q_y,\Delta E)\) such that \(\int_{0}^{2\pi}{\int_{0}^{\infty} {S_{1d}(k_y,\Delta E)}}d\Delta E dk_{y} = 1\) for the exchanges extracted from ([\[eq:rpa\]](#eq:rpa){reference-type="ref" reference="eq:rpa"}) to be physically meaningful. The imaginary and real dynamical susceptibilities \(\chi^{\prime\prime}(Q_y,\Delta E)\) and \(\chi^{\prime}(Q_y,\Delta E)\) were calculated using \(S_{1d}(Q_y,\Delta E)=\chi^{\prime\prime}(k_y, \Delta E)/\pi\) and a numerical Kramers-Kronig transformation, respectively. Finally, Brillouin zone folding was accounted for before powder averaging to obtain \(S(|Q|,\Delta E)\). Due to the time-consuming nature of this procedure, the fitting was performed using six constant-\(|Q|\) cuts through the data rather than the full data set. The optimization was carried out using the particle swarm method, and a single constant was added to all cuts to account for the instrumental background. The fit to the \(J_1-J_a\) model yielded a slight reduction in \(\chi^2\) from \(5.3\) (for the \(J_1\) only model) to \(5.1\), and a small \(J_a = 0.2(1)\)meV and nearly unchanged \(J_1=3.81(2)\)meV \[Fig. [\[fig:inelastic_neutron\]](#fig:inelastic_neutron){reference-type="ref" reference="fig:inelastic_neutron"}(d-h)\]. The inclusion of terms beyond \(J_a\) did not result in any further improvements to the fit. Dzyaloshinskii--Moriya interactions could be further consider in the model as these are allowed due to the \(2\%\) crystallographic disorder and this is the scope of a future work.
# Conclusions
Our results show that the spin chains in \(\beta\)-VOSO\(_4\) are formed by doubly (SO\(_{4}\))\(^{2-}\)-bridged chains along the \(b\)-direction that do not coincide with the structural chains of the V\(^{4+}\)O\(_{6}\) octahedra. These spin chains are magnetically coupled by the magnetic exchange interaction \(J_1=3.83\) meV. This is a common occurrence in V\(^{4+}\) compounds due to the nature of their half-filled magnetic orbital. \(J_2\) is the second neighbour coupling along the spin chain and it is at least one order of magnitude weaker than \(J_1\). It may be responsible for a weak frustration therein, although it is by far insufficient to open a spin gap, which was indeed not detected in any of our measurements. The coupling \(J_a\) links the spin chains into a three-dimensional network but with low connectivity as there are two such couplings per vanadium site, as opposed to the four inter-chain couplings per site assumed in a standard quasi 1D-antiferromagnet. The remaining discrepancy between the modelled INS intensity and the experimental data at \(\sim\)`<!-- -->`{=html}6 meV likely has a different origin, the investigation of which must await the availability of larger single crystals. The weak connectivity of the spin chains suggests that the Néel temperature of \(\beta\)-VOSO\(_4\) could be quite low. Indeed, we do not detect any magnetic order down to \(0.085\) K within the used experimental settings. We do not detect any effect of the reported crystallographic disorder in the magnetic properties of \(\beta\)-VOSO\(_4\), other than possibly a finite amount of paramagnetic impurities. | {'timestamp': '2022-12-20T02:24:23', 'yymm': '2212', 'arxiv_id': '2212.08473', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08473'} |
# Introduction
It has long been recognized that Galactic disk stars could be scattered by giant molecular clouds or transient spiral structures, or by both, increasing the random motion (or radial action) of the stars. Such processes could drive secular changes in the energy and angular momentum (or guiding radius) of disk stars and result in a gradual increase in velocity dispersion with age for subgroups of disk stars. For this reason, such processes are loosely referred to as heating processes. These kinds of dynamical processes are also found to be responsible for the formation of the Galactic thick disk because they are able to bring disk stars from a small vertical extent into a large one, allowing the thin disk stars to develop into the thick disk stars. For example, the heating of a preexisting thin disk through minor mergers is a mechanism widely used to explain the origin of the Galactic thick disk. However, there are other mechanisms that allow the Galactic thick disk to form: i) accretion from disrupted satellites and ii) in situ triggered star formation with various causes, which form the disk in an \"upside-down\" fashion (i.e., old stars form in a relatively thick layer, while young stars form in a thin layer). In addition, as evidenced by previous works, major merger events are another possible mechanism contributing to the formation of the thick disk stars, which is supported by the work of. Previous research has shown that spiral structures do not always heat the Galactic disks. The pioneering work of demonstrated that stars near corotation of a spiral pattern can vary their angular momenta without heating, which is a process known as radial migration. Further investigations have shown that radial migration can indeed occur due to a variety of sources, for example, spirals, bars, and satellites, and conserves the vertical actions of stars. It should be noted that kinematic heating can also occur during episodes of radial migration when resonant overlap is possible. Radial migration has been found to be able to account for the formation of the Galactic thick disk by the pioneering work of and further supported by subsequent research, though these works are controversial because radial migration is found to have little effect on the thickening of the disk. However, stated that the outward radial migration process under action conservation can achieve larger scale heights, which makes this mechanism a still possible scenario for the origin of the Galactic thick disk. Moreover, kinematic heating caused by radial migration increases radial action, which could affect the vertical distribution of disk stars and partly contribute to the formation of the Galactic thick disk. Although there are different mechanisms related to the origin of the Galactic thick disk, it is helpful to investigate whether the thick disk stars are born thick or thin in order to partly distinguish the mechanism that dominates the formation of the Galactic thick disk. The cosmological simulations that have been run with different galaxy formation codes fail to give a consistent result for this question but nonetheless highlight the role of the physical processes in star formation, for example, the recipe for feedback. An observational work using more than a hundred star-forming galaxies to investigate the variation in ionized gas velocity dispersions with redshift indicates that the thick disk is born hot, while the work of studying high-redshift galaxies with emission lines of cold interstellar matter reveals the contrary. In the Milky Way, the lack of direct observational evidence causes this topic to remain unresolved. The main reason is that the estimation of birth sites for the disk stars is nontrivial, especially for the thick disk stars, and the radial migration process makes this situation even worse. However, measuring the thickness of a population of related stars with a variable that can characterize heating is informative to the study of the heating origin of the thick disk. For this purpose, radial action is a suitable parameter, since it gradually increases by heating processes. Previous works have provided several ways to characterize these heating processes by using actions, but there is no direct measurement for the relation of radial action with the thickness of the Galactic disk. In this work, we investigate the relationship between radial action and the vertical distribution of the Galactic disk using data from the Apache Point Observatory Galactic Evolution Experiment (APOGEE) and Gaia with the purpose of deepening our understanding of the formation and heating history of the Galactic disks. This paper is structured as follows: in Section 2, we briefly describe the data adopted in this work; the methods and results are presented in Section 3. Finally, a discussion and conclusions are given in Section 4.
# Data
In this work, we used a sample of the Galactic disk stars similar to that was obtained through cross-matching between the 14th data release of APOGEE (APOGEE DR14) and the second data release of Gaia (Gaia DR2). However, where the stellar heliocentric distances in J18 were computed by inverting the parallaxes from Gaia, in this work we updated the distances and kinematical and orbital parameters using a Bayesian approach that follows. The details of parameter inference can be found in the above literature and references therein. In this section, we summarize the key points. The posterior probability distribution of the parameters (heliocentric distance, tangential speed, and travel direction) were derived by multiplying a Gaussian distribution likelihood by a separable prior of parameters. The heliocentric distance prior is an exponentially decreasing space density distribution with a Galactic longitude-and latitude-dependent length scale taken from. The tangential speed prior is a beta distribution where \(\alpha=2\), \(\beta=3,\) and \(v_{max}=750\) \(\rm km\, s^{-1}\). The travel direction prior is a uniform distribution. We obtained the posterior probability distribution of radial velocity by multiplying a Gaussian distribution likelihood by a uniform prior in order to further estimate the kinematical and orbital parameters. The mean and standard deviation in Gaussian distribution of radial velocity were provided by the values of radial velocity and their error in the APOGEE catalog. We used EMCEE to draw random samples from the above posterior distributions for the parameters of heliocentric distance, tangential speed, travel direction, and radial velocity. The medians of the posterior samples were then used to derive the distance in Galactic cylindrical coordinates \(R\) and \(Z\) and the spatial velocities (\(U,V,W\)) in right-hand Cartesian coordinates. Coordinate transformation was adopted to obtain the cylindrical coordinate velocities \(V_R\) and \(V_{\phi}\) after correcting the solar peculiar motion of \((U, V, W)=(11.1, 12.24, 7.25)\, \rm km\, s^{-1}\). Then, the guiding radius (\(R_g\)) and orbital parameters, such as peri-center (\(R_p\)), apo-center (\(R_a\)), and eccentricity (\(e=(R_a-R_p)/(R_a+R_p)\)), were estimated with *galpy* under the potential of *McMillan17*. The actions were calculated with the Staeckel approximation under this potential. In the calculation, we adopted the Gaia parallax zero point of-0.0523 mas, a distance from the Sun to the Galactic center of 8.21 kpc, and a local standard of rest velocity of 233.1 \(\rm km\, s^{-1}\). The Galactic disk star sample was selected by using the same criteria as in J18, which excluded stars with metallicities less than \(-1.0\) dex; total velocities (\(\sqrt{U^2+V^2+W^2}\) have been corrected for the solar peculiar motion) of less than 150 \(\rm km\, s^{-1}\), to diminish the contamination from the halo; and stars at the fields of (\(l\), \(|b|\)) \(<\) (10\(^\circ\), 10\(^\circ\)), to remove the bulge stars. At this stage, the star sample comprised 149 685 giant stars. However, there were 736 stars among them whose actions failed to be calculated under the potential of *McMillan17*. Considering that these stars constituted less than one percent of our total star sample, we dropped them from the sample. The final number of stars in the sample adopted by this work is 148 949. Most of the stars in this sample are located in a wide spatial range of the Galactic disk: \(2<R<14\) kpc and \(|Z|<5\) kpc (as shown in Fig. [\[mh-mg\]](#mh-mg){reference-type="ref" reference="mh-mg"} where star distributions in the chemical plane of \[M/H\] versus \[Mg/Fe\] are also presented in the bottom panel). Following J18, a solid line was used to cut the sample into a high-\[Mg/Fe\] sample and a low-\[Mg/Fe\] sample, corresponding to a thick disk star sample and a thin disk star sample, respectively.
# Method and results
As mentioned in the introduction, heating processes increase the random motions of stars and tend to move stars into more eccentric orbits. We illustrate the relations among the radial actions \(J_R\), eccentricities \(e,\) and \(R-R_g\) of the sample stars in Fig. [\[actions\]](#actions){reference-type="ref" reference="actions"}. According to the theory of the epicycle approximation, the radial action of a star with a near-circular orbit can be expressed by its eccentricity through \(J_R=\frac{E_R}{\kappa}=\frac{1}{2} \kappa e^2 {R_g}^2\), where \(\kappa\) denotes radial frequency. In the lower-left panel of Fig. [\[actions\]](#actions){reference-type="ref" reference="actions"}, a black dashed line shows the relation between radial action and eccentricity for stars with a guiding radius of 8 kpc and radial frequency of 39 \(\rm km\, s^{-1}\, kpc^{-1}\). Fig. [\[actions\]](#actions){reference-type="ref" reference="actions"} shows that kinematically hot stars tend to have large radial actions and large \(|R-R_g|\). Before investigating the relations between radial actions and the vertical distributions for the thin and thick disks, we focused on the distributions of radial actions for the thin and thick disks in Fig. [\[jr-distribution\]](#jr-distribution){reference-type="ref" reference="jr-distribution"}. These distributions can be described by pseudo-isothermal distributions, \(J_R\propto \exp(-J_R/\hat{J_R})\), where \(\hat{J_R}\) is a free parameter standing for the temperature of the disk. We estimated this parameter by using the maximum likelihood technique. Specifically, we obtained the ratio of the star number in a given \(J_R\) bin and the bin's width and compared them with the simulated data deriving from the above pseudo-isothermal distribution to search for the best-fit \(\hat{J_R}\). The uncertainty was similarly estimated to that of and. We calculated the likelihood 500 times using the observed data and the simulations of the best fitting model, applying the individual Poisson noise. The resulting likelihood range defined the uncertainty. The best-fit \(\hat{J_R}\) for each disk is listed in the legends of Fig. [\[jr-distribution\]](#jr-distribution){reference-type="ref" reference="jr-distribution"}, which is 47.0 \(\rm kpc\, km\, s^{-1}\) and 88.1 \(\rm kpc\, km\, s^{-1}\) for the thin disk and the thick disk, respectively. The pseudo-isothermal distribution of radial action requires that the best-fit \(\hat{J_R}\) be equal to the mean radial action of the star sample, which in our case is 47.1 \(\rm kpc\, km\, s^{-1}\) for the thin disk and 81.9 \(\rm kpc\, km\, s^{-1}\) for the thick disk. Regarding the uncertainties of the best-fit \(\hat{J_R}\) listed in the legend of Fig. [\[jr-distribution\]](#jr-distribution){reference-type="ref" reference="jr-distribution"}, the best-fit value of \(\hat{J_R}\) almost matches the mean radial action of the star sample, indicating that either \(\hat{J_R}\) or the mean radial action is a suitable proxy for the temperature of the disk. Once a heating process increases the temperature of a star population from a pseudo-isothermal distribution with a low temperature to a high one, the mean radial action (or \(\hat{J_R}\)) for that star population would also increase.
## Results from the overall Galactic disk
In Fig. [\[Z-distribution\]](#Z-distribution){reference-type="ref" reference="Z-distribution"} we illustrate the vertical distributions in different radial action ranges for both the thin and thick disk star samples. In order to quantify the relation of radial action with the vertical distribution, every vertical number density distribution was fitted with an exponential function, that is, \(\rho(Z) \propto \exp(-|Z|/h)\), where \(h\) is the scale height, by using a procedure similar to what we used when fitting the distributions of radial action. We note that, at a large distance, this exponential function is approximate to the isothermal distribution, \(\rho(Z) \propto \mathrm{sech}^2(-|Z|/(2h))\). The resulting best-fit scale heights with their uncertainties are listed in the legend in each of the panels in Fig. [\[Z-distribution\]](#Z-distribution){reference-type="ref" reference="Z-distribution"}. In Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"} we show the relations between radial action \(J_R\) and scale height \(h\) for both the thin and thick disk, which are derived from Fig. [\[Z-distribution\]](#Z-distribution){reference-type="ref" reference="Z-distribution"}. We attempt to describe the scale height as a function of radial action with a form of \(J_R=a \cdot (h-b)^2\), where \(a\) and \(b\) are free parameters. An alternative form of this function is \(h=\sqrt{J_R/a}+b\). The best-fit relations obtained from the maximum likelihood estimation technique are also shown in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"}. Considering that the best-fit scale heights may be less reliable due to the large uncertainties in large radial action bins (especially in the thick disk), which would affect the estimated parameters, we also applied a Bayesian inference. Compared to the maximum likelihood estimation technique, the advantage of Bayesian inference is that it directly compares the data with the model, avoiding the estimation of the scale height. By combining the normalized vertical number density of \(\rho(Z)=\frac{1}{2h} \exp(-|Z|/h)\) and the relationship of \(h=\sqrt{J_R/a}+b\), we can easily write the logarithm of the probability of finding a star with a given radial action of \(J_R\) at distance \(Z\), \[\ln p(Z|J_R,a,b) =-\ln 2(\sqrt{J_R/a}+b)-\frac{|Z|}{\sqrt{J_R/a}+b}.\] Therefore, the probability of finding a star where the radial action is \(J_R\) and the distance is \(Z\) can be expressed through the known pseudo-isothermal distribution of radial action(\(p(J_R) = \exp(-J_R/\hat{J_R})/\hat{J_R}\)), that is: \[p(Z,J_R|a,b)=p(Z|J_R,a,b) \cdot p(J_R).\nonumber\] By taking the logarithm of the above equation, we obtained the likelihood function: \[\ln p(Z,J_R|a,b) =-\ln 2(\sqrt{J_R/a}+b)-\frac{|Z|}{\sqrt{J_R/a}+b}-\ln \hat{J_R}-\frac{J_R}{\hat{J_R}}.\] Then assuming a uniform prior that requires \(a>0\) and \(b>0\), we formulated the logarithm of posterior (up to a constant) to be: \[\ln p(a,b|\{Z,J_R\}) = \sum_{i=1}-\ln 2(\sqrt{J_{R,i}/a}+b)-\frac{|Z_i|}{\sqrt{J_{R,i}/a}+b}-\ln \hat{J_R}-\frac{J_{R,i}}{\hat{J_R}}.\] The summation is over all stars, and the value of \(\hat{J_R}\) was chosen to be the mean of radial action of the star sample, instead of the best-fit values shown in Fig. [\[jr-distribution\]](#jr-distribution){reference-type="ref" reference="jr-distribution"} for the thin and thick disks. We find that this choice does not affect the estimated best-fit parameters. We sampled the posterior via the EMCEE package with 1,000 steps and 100 walkers. The initial start points of the chain are just the best-fit parameters that were obtained from the maximum likelihood estimation. Fig. [\[coner-plot\]](#coner-plot){reference-type="ref" reference="coner-plot"} illustrates the sampling of the model parameter posterior after discarding the initial 100 steps, and we used the 16th, 50th, and 84th percentiles to quote the uncertainties of the parameters. The best-fit relations from Bayesian inference are \[\begin{aligned} h =& \sqrt{J_R/5431.89}+0.263, \rm{thin\, disk} \\ h =& \sqrt{J_R/2190.09}+0.609, \rm{thick\, disk.} \end{aligned}\] Predictions from these best-fit results are shown in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"} as are the uncertainties in predictions. Figure [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"} clearly shows that the maximum likelihood (or the data points) and Bayesian estimations give a consistent relationship for the thin disk, while they present a discrepancy for the thick disk. This discrepancy might arise from the estimated scale heights of the thick disk in large radial action ranges that are less reliable, reflected by their large uncertainties in Fig. [\[Z-distribution\]](#Z-distribution){reference-type="ref" reference="Z-distribution"}. However, when taking the uncertainties of the Bayesian prediction and the best-fit scale heights into account, the resulting relation from Bayesian inference is basically in line with the data points. Considering that the Bayesian inference directly compares the model with data, we favor the results derived from this method. We also explored a three-parameter function of \(J_R=a \cdot (h-b)^\alpha\) to describe the relationship between radial action and scale height. In the case where \(\alpha=2\), this three-parameter function is the same model that we used in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"}. The best-fit results of this three-parameter model are illustrated in Fig. [\[three-parameters\]](#three-parameters){reference-type="ref" reference="three-parameters"}. The best-fit \(\alpha\) estimated from Bayesian inference is \(2.08^{+0.14}_{-0.16}\) for the thin disk, which is highly consistent with the function that was used in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"}. For the thick disk, the best-fit \(\alpha\) is \(3.98^{+1.47}_{-1.08}\), which might indicate that the function used in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"} is inappropriate for the thick disk. The uncertainties of \(\alpha\) and of the prediction from Bayesian inference are too large for the thick disk, reducing the reliability of this three-parameter function. Thus, we prefer to still use the two-parameter function (\(J_R=a \cdot (h-b)^2\)) to describe the relationship between radial action and scale height for the thick disk in this work, but we note that this two-parameter function for the thick disk should be used carefully in other works. To better understand the relationships between the radial action and scale height we obtained, we started from the isothermal limit and epicycle approximation. In the isothermal limit and epicycle approximation, the radial action is \(J_R=E_R/\kappa = \sigma_R^2/\kappa\) and \(\sigma_z^2=8\pi G\rho_0 h^2\). To express radial action \(J_R\) in terms of scale height required \(\sigma_R\) to be a known function of \(\sigma_z\). In our case, \(J_R \propto (h-b)^2\) requires \(\sigma_R^2 \propto (\sigma_z-b\sqrt{8\pi G\rho_0})^2\). Therefore, the existence of the relationship between radial action and scale height reveals a fixed relationship between \(\sigma_R\) and \(\sigma_z\). Without the knowledge of the exact relation between \(\sigma_R\) and \(\sigma_z\) from other data or theory to justify the relationship we measured, we elected to use \(J_R=a \cdot (h-b)^2\) to describe the relationship between radial action and scale height for both of the disks.
## Results from the inner and outer disk
Previous works have found that some of the properties of the inner disk are different from the outer disk, which seems to indicate different evolutionary histories. By using the guiding radius, we separated the Galactic disks into an inner disk set and an outer disk set. We defined the stars with a guiding radius smaller than 8 kpc as being an inner disk, and those with a guiding radius larger than 8 kpc as being an outer disk. The reason for using the guiding radius instead of the current galactocentric radius was that the guiding radius is more suited to representing the position of a star's orbit. We obtained the vertical distributions of the defined inner and outer disks and their corresponding best-fit distributions as we had done in Fig. [\[Z-distribution\]](#Z-distribution){reference-type="ref" reference="Z-distribution"}, which are illustrated in Fig. [\[inner\]](#inner){reference-type="ref" reference="inner"} and [\[outer\]](#outer){reference-type="ref" reference="outer"}, respectively. We point out that we increased the size of the radial action ranges in the last two lower panels of Fig. [\[outer\]](#outer){reference-type="ref" reference="outer"}, as the number of stars in the outer disk is relatively small. The relationships between radial actions and scale heights for the inner and outer disks are shown in Fig. [\[inner-outer\]](#inner-outer){reference-type="ref" reference="inner-outer"}. As done in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"}, both maximum likelihood and Bayesian estimations were applied, but the latter is preferred, as we explained in the previous subsection. The best-fit relations derived from Bayesian inference for the overall thin and thick disk star samples in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"} are also plotted for comparison. The inner and outer thin disks show nearly the same relations as that of the overall thin disk. However, the inner and outer thick disks present different relationships from each other. The inner thick disk has a relatively flat relation, and the scale heights are larger than those of the overall thick disk, except at the large radial action end. In contrast, the outer thick disk shows a similar trend with that of the overall thick disk, but the scale height is smaller at all radial actions.
# Discussion and conclusions
In this work, we have analyzed the scale height \(h\) as a function of radial action \(J_R\) for the overall disk in Fig. [\[jr-scale height\]](#jr-scale height){reference-type="ref" reference="jr-scale height"} and for the inner and outer disk in Fig. [\[inner-outer\]](#inner-outer){reference-type="ref" reference="inner-outer"} via two different approaches: maximum likelihood and Bayesian estimations. In order to directly compare this function with data, we estimated the scale heights in radial action ranges and then described the relationship between the scale height and radial action with a function of \(h=\sqrt{J_R/a}+b\) using the maximum likelihood estimation technique. As the estimated scale heights for the thick disk at large radial action ranges are too large and may affect the best-fit relationship between radial action and scale height, we also applied a Bayesian estimation to directly compare the model with data, avoiding to estimate the scale height in different radial action ranges. It should be noted that we do not correct any biases in this work. The reasons are: 1) we only used chemical information to divide the thin and thick disks; 2) neither the APOGEE nor Gaia data are kinematically biased; and 3) most mono-age, mono-\[Fe/H\] populations are close to isothermal . Therefore, the disk stars are mostly phase-mixed in the vertical direction. The first two reasons demonstrate that the only bias arises from APOGEE spatial selection. However, the third reason assumes the bias does not strongly affect our main results, as we are only interested in the vertical distributions of the disk stars. We find that a function of \(h=\sqrt{J_R/a}+b\) is able to describe the thin disk well, but it should be used with care for the thick disk. The best-fit function given by Bayesian inference is \(h =\sqrt{J_R/5431.89}+0.263\) for the thin disk, whereas \(h = \sqrt{J_R/2190.09}+0.609\) for the thick disk. Moreover, we find that the thin disk has nearly the same relationship between the inner and outer disk, as seen from Fig. [\[inner-outer\]](#inner-outer){reference-type="ref" reference="inner-outer"}, while the thick disk presents different relationships: the inner thick disk shows a nearly flattened trend, and the outer thick disk has a gradually increasing trend. As heating agents can broaden the distribution of radial action and increase the scale height of disk stars, the measured relationships encode the heating history and the origin information of the star samples. In the isothermal limit and epicycle approximation, we show that a function of \(h=\sqrt{J_R/a}+b\) reveals a fixed relationship between the radial and vertical velocity dispersion, that is, \(\sigma_R^2 \propto (\sigma_z-b\sqrt{8\pi G\rho_0})^2\). This relationship is the overall consequence from all of the heating agents. Without the aid of a simulation, it is hard to discuss the implications of these relationships on the formation and heating history of the disks. However, it is reasonable to suspect that giant molecular clouds may be a dominant or non-negligible heating agent to account for in these relationships, as they heat both radially and vertically and hence have a chance to form a fixed relationship between \(\sigma_R^2\) and \(\sigma_z^2\). We would like to further explore this content in the future. Most importantly, this work provides an alternative way to challenge models and thus study the heating history of the Galactic disks by investigating the relationship between radial action and scale height. | {'timestamp': '2022-12-19T02:13:27', 'yymm': '2212', 'arxiv_id': '2212.08519', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08519'} |
null | null |
null | null |
# Introduction {#sec:intro}
The advent of deep, all-sky, untargeted surveys, such as the Zwicky Transient Facility (ZTF, @bellm19), the Dark Energy Survey (DES, @bernstein12 [@kessler15]) and the upcoming Legacy Survey in Space and Time (LSST, @ivezic19) has ushered in a new age for the study of transient astronomy, and it is now one of the most popular fields in astronomy and astrophysics. The thousands of new transients discovered each day offer the potential to build statistical samples of otherwise rare and poorly known phenomena, such as tidal disruption events and superluminous supernovae, as well as find new types of transient objects that currently remain unknown. Supernovae (SNe) are among the most well-known and frequently studied transient objects for several reasons. First, due to their outstanding peak brightness, they can be discovered and studied even at high redshifts, i.e., \(z > 1\). Moreover, the \(\sim\)month timescale of their evolution is neither too fast nor too slow compared to human timescales, which makes them ideal targets for astrophysical studies. Also, supernovae are among the best distance indicators in extragalactic astronomy, and thus, are useful from not only an astrophysical, but also a cosmological point of view: for example, the discovery of the accelerated expansion of the Universe was at first based on distance measurements from Type Ia SNe. Finally, the James Webb Space Telescope (JWST) and future space observatories will extend our cosmic horizon considerably, thus, enabling the discovery of transients beyond \(z > 2\), in the very early Universe. The recently discovered very high redshift galaxies at \(z > 12\)--including one at \(z = 16.7\)--by JWST improve our understanding of the cosmic star formation rate at \(z > 8\) and the epoch of reionization. Studies of these high redshift supernovae will ultimately constrain the (true) initial mass function of the first stars. The unprecedented number of new transients in the data stream makes their analysis very challenging. Since most of the surveys mentioned above are based on imaging and photometry, the possibility for a prompt revealing of the physical nature of new transients is limited. As most classification schemes require (mostly optical) spectroscopy, it is not currently possible to definitively classify the bulk of these new sources. Instead, the current strategy is concentrated on identifying potentially interesting sources based on either their position (e.g., being in a nearby galaxy or galaxy cluster) or their photometric properties (e.g., unusually bright, blue/red, fast/slow evolution, etc.). This still leaves most transients unclassified and they fade away with minimal spectroscopic attention. In such circumstances, survey programs that produce a massive number of spectroscopic observations can be useful. One of them is the Hobby-Eberly Telescope Dark Energy eXperiement (HETDEX, @gebhardt21 [@hill21]). HETDEX is a blind spectroscopic survey that, once completed, will use observations of 1.2 million Ly\(\alpha\) emitting galaxies to measure the baryonic acoustic oscillations (BAO) and constrain the fundamental parameters that describe the time evolution of the dark energy equation-of-state. HETDEX aims to obtain spectral coverage of 540 deg\(^2\) of the sky with a filling factor of 1 in 4.6; a spectrum is obtained at every position that is covered by a fiber in the field-of-view, including all objects that are brighter than the HETDEX \(5\sigma\) limiting magnitude of \(m_{\rm lim}(\mathrm{AB}) \sim 22.5\) in \(g\)-band. The HETDEX survey is unique as it contains untargeted, deep, wide-field spectroscopic data. While HETDEX is at heart a high-redshift galaxy survey, it also obtains spectra of all other objects in its field-of-view, including stars of all types, early and late type galaxies, high-redshift quasars, and, in principle, supernovae. The unbiased nature of the HETDEX spectroscopic selection function implies that there must exist spectra of anomalous, perhaps unprecedented, objects within the HETDEX continuum-source catalog. The HETDEX data can also be used, at least potentially, for the classification of all types of transient objects that appear within the survey footprint. Another advantage of HETDEX is that it uses integral field units (IFUs), thus, not only single objects but also their immediate environment can be observed spectroscopically. This could be promising, for example, for studying the hosts of extragalactic transients (see Section [4.6](#sec:hosts){reference-type="ref" reference="sec:hosts"}). In this paper we present the first attempt to use HETDEX data for spectroscopic classification and characterization of transients that were discovered within the survey's footprint. While our primary purpose was finding supernovae, it turned out that the HETDEX spectra are also useful for identifying variable active galactic nuclei (AGN), and studying the host galaxies of supernovae and other transients. This paper is organized as follows. Data extraction is outlined in Section [2](#sec:data){reference-type="ref" reference="sec:data"}, then the applied classification methods are detailed in Section [3](#sec:methods){reference-type="ref" reference="sec:methods"}. Results are discussed in Section [4](#sec:disc){reference-type="ref" reference="sec:disc"}, while the conclusions are summarized in Section [5](#sec:conc){reference-type="ref" reference="sec:conc"}.
# Data {#sec:data}
HETDEX uses the Visible Integral-field Replicable Unit Spectrograph (VIRUS, @hill21) instrument at the focal surface of the Hobby-Eberly Telescope (HET, @ramsey98 [@hill21]). VIRUS comprises up to 78 integral field units, each with 448 \(1\farcs 5\)-diameter fibers, such that 33,152 individual spectra covering 54 arcmin\(^2\) are obtained with each visit. For more details about the survey design concepts and specifications, see. For more details about the instrumentation, see. The HETDEX Data Release 2.1 (HDR2.1) continuum-source catalog contains spectra for \({\sim} 62\)k unique objects (Mentuch Cooper, in prep), and includes a vast amount of information that is too large to be visually classified by human experts. Thus, the very nature of the HETDEX survey design motivates machine learning efforts not only for anomaly detection, but also for object classification. In the HETDEX Data Release 3 (HDR3), the size of the continuum-source catalog has been extended further, reaching \({\sim} 230\)k sources. Supernovae identifications in the HETDEX catalog are challenging, not least because they are not the ultimate target of the survey. As a result, the various reduction pipelines that produce the survey's data products are not optimized to find extragalactic transients. For example, one such issue is that, within the continuum detection process, sources are rejected if they are within \(1\farcs 5\) of another galaxy; in these cases only the brightest continuum object remains in the catalog. This cut may discard supernovae and similar astronomical transients that are associated with a bright host galaxy, but are actually separate point sources near their host galaxy's core. We therefore opted not to use the default HETDEX data products for our initial search for supernovae. Instead, for this work we chose to perform forced spectral extractions at the known locations of transient sources. There are many astronomical surveys that can provide auxillary information to HETDEX. One of these is the Zwicky Transient Facility (ZTF, @bellm19 [@graham19; @masci19; @dekany20]), which covers the full northern sky in two filters every three nights. This survey typically produces between \(10^5\) and \(10^6\) alerts per night, 80% of which are associated with real astronomical sources. Thus, we aimed to extract spectra at the positions of public ZTF transients announced between 2018 and 2022 from the HETDEX spectroscopic dataset, in the hopes of finding SN-like signals (broad features) in those spectra. Details are given in the rest of this paper.
## Source extraction {#sec:extraction}
We used the Automatic Learning for the Rapid Classification of Events (ALeRCe; @Forster21) Application Programming Interface (API)[^1] to query the ZTF servers to find all detected ZTF transients that were within 12 arcminutes of the position of a HETDEX pointing. (This is roughly the size of the HET's usuable focal plane, which is 20 arcminutes in diameter.) We then determined which of the resulting 4585 ZTF sources fell within 3 of any of the \(\sim\)`<!-- -->`{=html}34,000 fibers that are part of VIRUS. This additional downsampling resulted in 636 spectra of 583 objects matched between ZTF and HETDEX. (Although HETDEX is a single epoch survey, some science verification fields have been observed multiple times.) Using `HETDEX-API/get_spec.py`[^2], we were then able to extract HETDEX spectra that are closely matched to each of these transient locations. We used a point spread function (PSF)-weighted 1D extraction at the exact position that the ZTF source was detected and required that at least 7 fibers be contained within the 3 circular aperture that is needed to fully capture the VIRUS PSF. Thus, the final spectrum is produced from a PSF-weighted sum over many fibers combined within a circular aperture with a radius of 3. It is worth noting here that the HETDEX spectra associated with these sources may not have been taken while the prospective transient was active. Indeed, in the majority of cases they were not. We can, however, still search for the nebular spectra of supernovae that have faded beyond the ZTF detection limit. While searches for pre-SN progenitor stars may also be possible, it is likely that those progenitors will be too faint for detection.
# Methods {#sec:methods}
This section summarizes the various methods that we applied to explore and characterize the extracted HETDEX spectra.
## Exploration: machine learning (ML) {#sec:ml}
A large number of discoveries in astronomy, and science more generally, have been catalyzed due to some degree of serendipity. The use of machine learning for anomaly detection has the potential to automate serendipity, and dramatically widen the potential discovery pool. One method that facilitates the visual exploration of large datasets is t-distributed stochastic neighbor embedding (t-SNE; @vanderMaaten08). The t-SNE program is fundamentally a dimensionality-reduction technique that preserves the pairwise similarity between points while producing a \(d\)-dimensional map of those points (usually \(d=2\)). We use the version of t-SNE implemented in the `scikit-learn` Python package. Here the dimensionality of the high-order space is set by the number of flux measurements (1036) in a single HETDEX spectrum. The t-SNE program works by first computing pairwise similarities \(p_{ij}\) between these high-dimensional data points \({\bf x}_i\) and \({\bf x}_j\), \[p_{j|i} = \frac{{\rm exp}(-||{\bf x}_i-{\bf x}_j||^2/2\sigma^2_i)}{\sum_{k \neq i}{\rm exp}(-||{\bf x}_i-{\bf x}_k||^2/2\sigma^2_i)},\] where \(\sigma_i\) is the variance of a Gaussian distribution centered on point \({\bf x}_i\). This parameter is related to a user-defined hyperparameter called the perplexity. We choose a perplexity value of \(\sqrt{N}\), where \(N\) is our number of data points. To mitigate problems associated with outliers, the symmetrical pairwise similarity is defined as: \[p_{ij} = \frac{p_{j|i} + p_{i|j}}{2N},\] where \(p_{j|i}\) is the probability that point \({\bf x}_{i}\) would choose point \({\bf x}_j\) as its nearest neighbor under the assumption of a Gaussian probability distribution centered at point \({\bf x}_i\). The next step is for t-SNE to attempt to learn a low-dimensional (\(d=2\)) representation of the points, \({\bf y}_1...,{\bf y}_N\), that preserves the high-dimensional similarities \(p_{ij}\) as analogous low-dimensional similarities \(q_{ij}\). In this low dimensional map, \[q_{ij} = \frac{(1 + ||{\bf y}_i-{\bf y}_j||^2)^{-1}}{\displaystyle\sum_{k \neq m}(1 + ||{\bf y}_k-{\bf y}_m||^2)^{-1}}.\] This metric is related to the Student's t-distribution, and its primary purpose is to avoid over-crowding in the low-dimensional embedded space. Also, the Student's t-distribution is faster to compute compared to a Gaussian as it does not contain any exponential terms. The key process of t-SNE is to determine the low-dimensional embeddings of the points \({\bf y}_i\) by minimizing (via gradient-descent) the Kullback-Leibler divergence between the low-and high-dimensional similarity distributions \(P\) and \(Q\), \[KL(P||Q) = \sum_{i \neq j} p_{ij} {\rm log} \frac{p_{ij}}{q_{ij}}.\] This will result in an optimized \(d=2\) dimensional map that preserves the similarities of points from the high-dimensional distribution.
## Classification: template matching via \(\chi^2\)-minimization {#sec:chi2template}
The traditional method for automated classification of astronomical objects is through comparisons with a list of template spectra using a mathematical criterion to judge the quality of matching; the classification then follows from the best-matched template. We employed this template-matching scheme using as templates the Pickles Stellar Spectral Flux Library for stars and the Kinney synthetic spectra for galaxies. The former contains 131 stellar spectra covering the Harvard classes from O5 through M4 and the Morgan-Keenan luminosity classes from I to V; the latter includes spectra for 11 galaxies, including an elliptical (E), a lenticular (S0), three spirals (Sa, Sb, Sc) and six star-bursting (SB1-SB6) systems. The star-forming templates are dominated by emission features of varying strength, while the elliptical and lenticular templates represent the spectra of most passive galaxies. We adopted \(\chi^2\) statistics for measuring the quality of the match between the observed spectrum and a template. The normalized \(\chi^2\) function was defined as \[\chi_N^2 ~=~ \frac{1}{N} \sum_{k=0}^{N} \frac{1}{\sigma_k^2} ( f_c^{\rm obs}(\lambda_k)-f_c^{\rm templ}(\lambda_k))^2, \label{eq:chi2}\] where \(N\) is the number of sampled wavelengths in an observed spectrum, the functions \(f_c^{\rm obs}\) and \(f_c^{\rm templ}\) are the continuum-normalized flux of the object and the template as a function of wavelength (\(\lambda\)), and \(\sigma_k\) is the uncertainty of the observed fluxes. Note that, as written, Equation [\[eq:chi2\]](#eq:chi2){reference-type="ref" reference="eq:chi2"} does not contain any terms for the the possible redshift mismatch between the observed and the template spectrum, and thus is only applicable for Milky Way objects with \(z=0\). Before using the equation for galaxies, the templates were therefore redshifted, and a value of \(\chi^2_N\) was computed for each \(z\) between \(0 < z < 0.3\) range, using \(dz = 0.001\) (see Section [4.2](#sec:disc-templ){reference-type="ref" reference="sec:disc-templ"} for further information). Before applying Equation [\[eq:chi2\]](#eq:chi2){reference-type="ref" reference="eq:chi2"}, simple linear interpolation was applied to the template spectra to correct for the different wavelength sampling between the template and the observed spectrum. In addition, both spectra were normalized to the continuum by iteratively fitting an 8th order polynomial to both the observed and the template spectrum. After each iteration, the standard deviation (`stddev`) of the residuals were calculated and data that exceeded the residual by `-2*stddev` and `+3*stddev` were rejected from the sample. After \(\sim 10\) such iterations the fitted polynomial was found to represent the continuum reasonably well. The uncertainties of the observed fluxes were estimated empirically by calculating the mean (\(\langle f_c \rangle\)) and standard deviation (\(\sigma\)) of the continuum-normalized data in the red part of the spectrum, between 4600 and 5400 Å. The same region was used to derive an empirical signal-to-noise (S/N) parameter, defined as `s/n`\(= \langle f_c \rangle / \sigma\). Because of the presence of spectral features in the fitted wavelength region, this method clearly overestimates the noise in the case of spectra with moderate to high S/N, while giving more realistic results for noise-dominated data. We chose this simple approach of adopting a single \(\sigma\) for the whole spectrum in Eq. [\[eq:chi2\]](#eq:chi2){reference-type="ref" reference="eq:chi2"}, i.e., \(\sigma_k = \sigma\) for all \(k\), because the majority of our sources did not reach high S/N, and the \(\chi^2_N\) function was used only for measuring the quality of fitting between the observed and the template spectra. We do not treat the value as a proper statistical quantity.
## Classification: template matching via cross-correlation {#sec:sntemplate}
The other widely used algorithm for template matching is cross-correlation. A code that is based on cross-correlation and applied very often in the supernova community is `SNID`, which uses the observed spectra of different types of SNe as templates. Such codes compute the cross-correlation function (*CCF*) between the observed (\(f(\lambda)\)) and the template spectrum (\(g(\lambda\))) via the following definition: \[CCF(\Delta x) = \int_{-\infty}^{+\infty} f(x) \cdot g(x-\Delta x) dx,\] where \(x = \ln \lambda\), \(\Delta x = \ln(\lambda / \lambda_0) = \ln(1+z)\) and \(z\) is the redshift of the object. In short, the *CCF* measures the overlap, i.e., the similarity, between the object and the template spectrum: the stronger the overlap, the greater than similarity of the two spectra. Thus, the height of the \(CCF\) peak is proportional to the overlap, while the horizontal shift (\(\Delta x\)) gives the redshift of the observed spectrum with respect to the template. Because the public template libraries of `SNID` are based on published observed spectra, Type Ia SNe are overrepresented in them. In addition, most of these template libraries do not fully cover the HETDEX wavelength range of 3500 to 5500 Å. Thus, instead of applying `SNID` for our sample, we developed our own pipeline based on the `fxcor` task in `IRAF`[^3]. Since our primary intention was finding supernovae, we adopted the Type Ia supernova templates compiled by Eric Hsiao for our analysis. These spectra span from \(-15\) days to \(+91\) days in phase with respect to the moment of \(B\)-band maximum, with a 1-day cadence. Since our single epoch data do not allow precise temporal resolution, we pre-selected 9 spectra from the library (having epochs at \(-15\), \(-5\), 0, +5, +10, +20, +30, +40, and +60 days with respect to maximum light) and used only those data in the cross-correlation. The selected SN Ia templates are shown in the left panel of Figure [\[fig:sn_templates\]](#fig:sn_templates){reference-type="ref" reference="fig:sn_templates"}. Each template spectrum has \(\Delta \lambda = 10\) Å resolution (which is much lower than the \(\sim 2\) Å resolution of the HETDEX spectra) and covers the wavelength interval between 2500 and 6000 Å. Because the spectra of hydrogen-poor (Type Ib/c) core collapse SNe are often similar to those of Type Ia SNe, at least around maximum light when the SN appears to be the brightest, we did not define a separate set of templates for those objects. Instead, the Hsiao-templates were utilized for the SN Ib/c-type objects. For hydrogen-rich (Type II) core collapse SNe, we adopted the published spectra of the archetypal Type IIP SN 1999em, downloaded from the Weizmann Interactive Supernova Data Repository (WiseRep)[^4]. The selected spectra are corrected for the redshift of SN 1999em (\(z = 0.002392\)) and shown in the right panel of Figure [\[fig:sn_templates\]](#fig:sn_templates){reference-type="ref" reference="fig:sn_templates"}. We chose only the spectra that were taken during the plateau phase (lasting \(\sim 100\) days after the explosion), because the SN is brighter by at least 1.5-2 magnitudes during this period than in the subsequent tail-and nebular phase. Thus, it is more probable to detect a Type II at this time.
# Results and Discussion {#sec:disc}
The results of our analysis are presented and discussed in this Section.
## Machine learning {#sec:disc-ml}
After having extracted spectra from the HETDEX database at the positions of the known ZTF transients (see Section [2.1](#sec:extraction){reference-type="ref" reference="sec:extraction"}), we normalized the spectra by their uncertainty (produced by the HETDEX reduction pipeline) to account for edge defects and other minor issues, and fed them directly into the t-SNE algorithm[^5]. This resulted in a two-dimensional map shown in Figure [\[fig:tsne\]](#fig:tsne){reference-type="ref" reference="fig:tsne"}, which we could then explore and visually inspect. We created a point-and-click tool to quickly and efficiently examine the spectra represented by each t-SNE data point, and by this visual inspection process, we identified the clusters that were predominantly AGN, star-forming galaxies, and low S/N sources. We also flagged several objects for further analysis. These objects of special interest were analyzed using `SNID` in order to determine their astronomical type. Amongst these was ZTF20aatpoos, which we classified as a Type II-P SN about 10 days after peak (see Section [4.4](#sec:sne){reference-type="ref" reference="sec:sne"}). This classification represents our first success in searching for SNe in the HETDEX database, and motivates efforts for further discovery.
## Template fitting via \(\chi^2\) {#sec:disc-templ}
The \(\chi^2_N\) value (Equation [\[eq:chi2\]](#eq:chi2){reference-type="ref" reference="eq:chi2"}) was computed for all 636 observed spectra using all templates in both the stellar and the galaxy library. We defined three groups of objects, named `star`, `galaxy` and `low`, as the initial classification categories for the observed spectra. Spectra that were noise-dominated and did not reach `s/n`=3 were moved into the `low` category. Sources that had lower \(\chi^2_N\) values for fits to \(z = 0\) stellar templates than fits to galaxies at \(z>0\) were classified as stars, and the remaining spectra were, at first, labeled as galaxies. In the next step, all spectra were inspected visually, and peculiar objects were moved into additional categories. We identified broad-lined AGNs (category `agn`) in the `galaxy` group using the measured width of their emission features. Since many of these sources showed only a single broad line in the relatively narrow wavelength interval observed by HETDEX (3500 to 5500 Å), their redshifts, as derived from the best-fitting galaxy template may differ substantially from the true value. These objects are analyzed further in Section [4.5](#sec:agn){reference-type="ref" reference="sec:agn"}. Finally, spectra that did not resemble any other members in the above categories were moved to the `uncertain` category.
[\[tab:stat\]]{#tab:stat label="tab:stat"}
It is an interesting question whether the results from our simple classification scheme have any correlation with the classifications performed by the ALeRCE pipeline. To test for this, we cross-compared our classification results for each transient (`star, galaxy, agn,` and `low`) to their most probable ALeRCE classification. From the various algorithms available in ALeRCE, we selected the one that had the highest probability value from the so-called Stamp Classification that uses the following categories: variable star (`VS`), active galactic nucleus (`AGN`), supernova (`SN`), asteroid (`AST`), and artifact (`BOGUS`). In addition, we introduced a 6th category, named unclassified (`UNC`), for those ZTF objects that were announced but had no ALeRCE classification available. As an initial step, all sources belonging to the `BOGUS` category were rejected from the sample. Such "transients" usually appear in the vicinity of bright, saturated objects, e.g., near their diffraction spikes, and are artifacts of the ZTF image subtraction process. A total of 251 such artifacts were removed from the sample, and only the remaining 385 spectra were analyzed further. The summary of our cross-comparison can be found in Table [1](#tab:stat){reference-type="ref" reference="tab:stat"} and is shown graphically in Figure [\[fig:classif_stats\]](#fig:classif_stats){reference-type="ref" reference="fig:classif_stats"}. It is seen that the results from the \(\chi^2\) template matching correlate very well with those from the ALeRCE Stamp Classification. For example, of the 149 sources with stellar-like HETDEX spectra, 114 were classified as variable stars by ALeRCE. Similarly, the 122 galaxy-like HETDEX spectra belong to ZTF transients that were classified by ALeRCE as either `AGN` or `SN`. It is especially encouraging that out of the 52 spectra that we classify as belonging to AGN, ALeRCE found 38 to be likely AGN. As expected, many of the low S/N spectra, i.e., those that belong to the `low` category in our scheme, have no ALeRCE classification. It is also interesting that most of the transients that ALeRCE identified as most likely SNe are associated with either galaxies or noise-dominated sources. This is also an expected result, because many of the HETDEX spectra were taken outside the visibility window of an active transient, resulting in either a low S/N spectrum or that of the likely host galaxy.
Our other SN candidate, ZTF19abdkelq, was originally classified as an asteroid (`AST`) by ALeRCE, but ZTF also recorded \(r\)-band magnitudes at 4 later epochs, with the last ZTF detection occurring \(\sim 30\) days after discovery. Thus, it is unlikely that this source is a moving object. Since it was not flagged as a transient, there is no available TNS record for the object. The HETDEX spectrum of ZTF19abdkelq is plotted in Figure [\[fig:sn2\]](#fig:sn2){reference-type="ref" reference="fig:sn2"}. In the left panel, we show the continuum-normalized spectrum (plotted with red) together with the best-matching Hsiao SN Ia template (black). Even though the observed spectrum is noisy, the pseudo-emission feature around 5200 Å is a good match to the template spectrum. The resemblance is strengthened by the analysis with `SNID` (right panel), which suggests SN 2000cx, a peculiar Type Ia SN, as the best-matching template. Both our *CCF* code and `SNID` provided a consistent redshift for ZTF19abdkelq of \(z \sim 0.12\). In Figure [\[fig:phot_sn2\]](#fig:phot_sn2){reference-type="ref" reference="fig:phot_sn2"}, the light curve of ZTF19abdkelq is plotted from ZTF \(r\)-band photometry. Green and blue curves show the expected brightness decline of a Type Ia SN, as inferred from rest-frame \(V\)-and \(B\)-band template light curves from MLCS2k2. At \(z \sim 0.12\), the rest-frame \(V\)-band template light curve represents the observer-frame \(r\)-band data well. The result of \(g\)-band synthetic photometry from the HETDEX spectrum (\(g = 22.144 \pm 0.098\) mag) is also shown as an orange square. It is seen that both the ZTF and the HETDEX photometry is entirely consistent with the proposed Type Ia SN classification of ZTF19abdkelq. The parameters for the two SNe found in HETDEX are summarized in Table [\[tab:sne_snid\]](#tab:sne_snid){reference-type="ref" reference="tab:sne_snid"}.
## Active Galactic Nuclei {#sec:agn}
Members of the AGN group were identified via the \(\chi^2\)-minimization method (see Section [3.2](#sec:chi2template){reference-type="ref" reference="sec:chi2template"}), although some were also found by cross-correlating the spectra with supernova templates (Section [3.3](#sec:sntemplate){reference-type="ref" reference="sec:sntemplate"}). We identified 52 spectra (Table [1](#tab:stat){reference-type="ref" reference="tab:stat"}) that had a galaxy spectrum as their best-matching template, but also showed broader emission features than those of the star-forming galaxy templates. After visual inspection, 3 of them turned out to be saturated stars, which were then removed from the sample. The remaining 49 were analyzed further. Redshifts were derived by fitting the "high luminosity QSO" template spectrum taken from the SDSS site[^7] to the AGN spectra. The fitting parameters for the AGN sample are collected in Table [\[tab:agn_z\]](#tab:agn_z){reference-type="ref" reference="tab:agn_z"} in the Appendix. Figure [\[fig:agn_z\]](#fig:agn_z){reference-type="ref" reference="fig:agn_z"} shows the `s/n` parameter of each spectrum (see Section [3.2](#sec:chi2template){reference-type="ref" reference="sec:chi2template"}) against redshift. It is seen that most of the ZTF transients that turned out to be AGNs are at low redshifts, i.e., at \(z \lesssim 0.5\). As a comparison, the full HETDEX AGN sample extends to at least \(z \sim 3.5\) and has a median redshift of \(z_{\rm med} \sim 2.1\); in other words, the AGNs identified here are more local than the majority of the HETDEX AGNs. This is most likely due to the sensitivity limit of the ZTF public stream, which is \(m_{AB} \lesssim 20\) mag. This limit is significantly brighter than the HETDEX detection threshold for AGNs. In Figure [\[fig:agn_sp\]](#fig:agn_sp){reference-type="ref" reference="fig:agn_sp"}, spectra of some of the ZTF AGN sources are plotted against rest-frame wavelengths compared to the SDSS QSO template. One object, ZTF20aappepk (\(z \sim 2.29\)) has a HETDEX spectrum that extends beyond the blue edge of the SDSS QSO template, but the overlapping part shows good agreement with the template, just like the other, lower redshifted AGNs in this small sample.
## Host galaxies of supernovae {#sec:hosts}
Information on the host galaxies of transients can be as important as the transients themselves. HETDEX presents a unique opportunity to gather spectra of transient host galaxies with a minimal observational cost: if a SN exploded within a HETDEX field, then it should be possible to extract the host galaxy spectrum from the HETDEX data products. While such a dedicated study of transient host galaxies is beyond the scope of the present study, it may be a fruitful endeavor for future work. Here we present galaxy types and redshifts for those HETDEX galaxies that were extracted *at the position* of ZTF transients that were classified as SNe by ALeRCE. We found 50 such objects according to Table [1](#tab:stat){reference-type="ref" reference="tab:stat"}, of which 21 had a refined ALeRCE supernova classification based on their ZTF light curves. ALeRCE uses the following categories from its `lc_classification_transient` pipeline: `SNIa`, `SNIbc`, `SNII` and `SLSN`. This subsample is listed in Table [\[tab:hostgx\]](#tab:hostgx){reference-type="ref" reference="tab:hostgx"} in the Appendix, together with the host galaxy properties found by the \(\chi^2\)-classification method (Section [3.2](#sec:chi2template){reference-type="ref" reference="sec:chi2template"}). The distribution of the host galaxy types for this subsample is shown in Figure [\[fig:hostgx_hist\]](#fig:hostgx_hist){reference-type="ref" reference="fig:hostgx_hist"}. Here we added the numbers of `SNIbc` SNe to that of the `SNII` objects, and use the `SNcc` category for those core-collapse SNe. Note that the numbers for the `SLSN` category are probably overestimated by ALeRCE. The same distribution for the whole sample (i.e., with those objects that do not have an ALeRCE light curve classification) is also plotted via the red bars. It is clear that even though the numbers are too low for any robust conclusions, most of the SNe occurred in starburst galaxies showing strong emission features. Since core collapse SNe originating from massive stars likely dominate our low-redshift sample, the distribution of host galaxy types as shown in Figure [\[fig:hostgx_hist\]](#fig:hostgx_hist){reference-type="ref" reference="fig:hostgx_hist"} is consistent with expectations. Note that ZTF19abuxhqd, which was classified as a likely Type II SN by ALeRCE, occurred in an early-type (S0) galaxy. This would be peculiar, since early-type galaxies usually host Type Ia SNe that originate from older population of stars; however, a close inspection of the noise-dominated ZTF light curve of ZTF19abuxhqd provides no convincing evidence that the presence of a core collapse SN. Specifically, the transient showed random-like brightness fluctuations for \(\sim 400\) days, instead of any sign of a decline. The distribution of the redshifts for the whole sample, as well as for those SNe that have refined ALeRCE types is plotted in Figure [\[fig:hostgx_zdist\]](#fig:hostgx_zdist){reference-type="ref" reference="fig:hostgx_zdist"}. This figure confirms that the ZTF SNe whose host galaxies could be measured by HETDEX are distributed in the \(z< 0.3\) low-redshift space.
## The expected number of supernovae in HETDEX {#sec:rates}
Even though HETDEX is not an ideal survey for finding extragalactic transients (see Section [1](#sec:intro){reference-type="ref" reference="sec:intro"}), the number of supernovae that are expected to show up within the survey footprint may be of interest. Since the classification of new transients generally needs spectroscopy, the future HETDEX data releases could be useful for classifying targets that might otherwise escape attention. The expected number of supernovae within the survey area (\(\Omega\)) covered by observations during the survey time (\(T_S\)) in redshift interval \(\Delta z\) at redshift \(z\) can be expressed as \[N_{SN}(z, \Delta z) = \Omega T_s \int_{z-\Delta z}^z \frac{R_{SN}}{1+z} \frac{dV}{dz} dz,\] where \(R_{SN}\) is the measured supernova volumetric rate (the number of SNe per Mpc\(^3\) per year) and \(dV/dz\) is the differential comoving volume. The factor \(1+z\) accounts for cosmological time dilation. The HETDEX survey field consists primarily of two parts: a Spring field of 390 deg\(^2\) and a Fall field of 150 deg\(^2\). Due to gaps between the IFUs, the effective survey field of view is \(\sim 94\) deg\(^2\). The area covered by a single shot with VIRUS is \(\sim 0.015\) deg\(^2\). To estimate the number of SNe found during a year, we assume that, on average, \(\sim 3\) such fields are taken on each night in a year for either the Spring or the Fall field. For the volumetric SN rates, we adopt \(R_{Ia} = 2.5 \times 10^{-5} \cdot (1+z)^{1.5}\) SNe Mpc\(^{-3}\) yr\(^{-1}\) for SNe Ia and \(R_{CC}(z=0.028) = 9.1 \times 10^{-5}\) SNe Mpc\(^{-3}\) yr\(^{-1}\) for core-collapse SNe. The latter is scaled with the cosmic star formation rate function by for different redshifts. Thus, using \(\Omega = 3 \cdot 0.015 = 0.045\) deg\(^2\), \(T_s = 1\) year and the volumetric SN rates given above, we get \(N_{Ia} \sim 0.24\) SN Ia in the \(0 < z < 0.3\) redshift range. Since core collapse SNe are fainter by \(\lesssim 1\) magnitude than Type Ia SNe, they are accessible by HETDEX only in the \(0 < z < 0.2\) redshift range (assuming \(\sim 22\) mag as the detection limit of continuum sources in HETDEX). Within this volume we get \(N_{CC} \sim 0.38\) core collapse SN per year. This optimistic calculation does not take into account detection efficiencies (signal-to-noise, for example), thus, it is likely an overestimate, but the final numbers are consistent with our results (one SN of both types within \(\sim5\) years of HETDEX operation). Note that between 2017 and 2022 the HETDEX field-of-view was not completely filled up by IFUs, which decreased our chances to find transients. Starting from 2022, having the field-of-view fully populated by 78 IFUs, the probability of discoveries are expected to increase. Overall, we conclude that we can expect to find a few more (\(N \lesssim 5\)) supernovae during the upcoming years of HETDEX operation.
# Conclusions {#sec:conc}
The conclusions of the present study are summarized as follows:
- We identified the counterparts of 583 transient objects in the HETDEX spectral database (but not necessarily at the same epoch) out of 4845 transient sources found by ALeRCe in the ZTF public data stream. In total, 636 spectra covering the 3500-5500 Å wavelength range were extracted from the HETDEX observing archive and analyzed.
- We applied the t-distributed stochastic neighbor embedding (t-SNE) machine-learning method to explore the dataset and identify potentially interesting sources. Our primary intention was to identify supernovae, which succeeded in 2 cases (see below).
- All ZTF sources having HETDEX spectra were classified into `star`, `galaxy`, `agn`, and `low` categories based on matching with template spectra (or, in the last case, having low signal-to-noise). For most objects, these categories agreed very well with their "Stamp-classification" made by the ALeRCe pipeline for their ZTF counterparts.
- We attempted to find supernovae among the HETDEX spectra in three different ways: visual inspection of the t-SNE 2D output, looking for outliers in the \(\chi^2\) template matching classification, and cross-correlation with SN templates. ZTF20aatpoos was flagged by all three methods as a potential SN, which was successfully confirmed as a Type IIP event using `SNID`. We also identified another SN candidate, ZTF19abdkelq, among the low S/N sources, and `SNID` confirmed it to be a Type Ia SN taken \(\sim 40\) days after maximum light.
- As a by-product, we identified 49 AGN sources, most of them at low (\(z \lesssim 0.5\)) redshifts; 38 of these were also classified as AGN by ALeRCe.
- We obtained host galaxy spectra for 50 ZTF transients classified as supernovae by ALeRCE. Most turned out to be starburst galaxies, which is consistent with the expectation that core-collapse events dominate the supernova sample in the low-redshift (\(z \lesssim 0.3\)) Universe.
We expect to find a few (\(\lesssim 5\)) more supernovae during the upcoming years of the HETDEX survey.
[^1]: https://alerce.readthedocs.io/en/latest/tutorials/ztf\_api.html
[^2]: <https://github.com/HETDEX/hetdex_api/blob/master/hetdex_tools/get_spec.py>
[^3]: IRAF was distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.
[^4]: `https://www.wiserep.org/`
[^5]: Some other studies have used a principle component analysis to reduce the dimensionality of the data before feeding it into t-SNE. We found that this was not necessary in our case due to the relatively small size of the data set.
[^6]: `https://www.wis-tns.org/`
[^7]: `https://http://classic.sdss.org/dr5/algorithms/spectemplates/` | {'timestamp': '2022-12-19T02:11:06', 'yymm': '2212', 'arxiv_id': '2212.08444', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08444'} |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction {#sec:intro}
Linguists differentiate between two types of metaphor: novel and conventional. While novel metaphors are creative expressions made in a particular situation by one particular individual, conventional metaphors are those which have been widely adopted by a language community. Consider:
1. The began at dawn. [\[ex:literal\]]{#ex:literal label="ex:literal"}
2. The government has come under . [\[ex:conventional\]]{#ex:conventional label="ex:conventional"}
3. The government the housing bill. [\[ex:novel\]]{#ex:novel label="ex:novel"}
Example [\[ex:literal\]](#ex:literal){reference-type="ref" reference="ex:literal"} is a literal usage of which refers to a military offensive. Example [\[ex:conventional\]](#ex:conventional){reference-type="ref" reference="ex:conventional"} is a conventional metaphor referring to intense verbal criticism. Example [\[ex:novel\]](#ex:novel){reference-type="ref" reference="ex:novel"} is a novel metaphor, which reuses the well-established imagery from [\[ex:conventional\]](#ex:conventional){reference-type="ref" reference="ex:conventional"}, but extends it with the word . Detecting novel metaphors and conventional metaphors are fundamentally different tasks. Novel metaphors are creative usages of words, which violate statistical patterns of language, and which extend the meaning of a word into unexpected areas of semantic space. Conventional metaphors, on the other hand, can be considered lexicalised: examples [\[ex:literal\]](#ex:literal){reference-type="ref" reference="ex:literal"} and [\[ex:conventional\]](#ex:conventional){reference-type="ref" reference="ex:conventional"} come from the gloss of two senses of in WordNet, where the novel sense of in example [\[ex:novel\]](#ex:novel){reference-type="ref" reference="ex:novel"} is not yet captured. Despite these inherent differences, in NLP the standard metaphor detection task, abbreviated as SMD here, makes no distinction between novel and conventional metaphor.
Although recent works have produced resources which distinguish between novel and conventional metaphors, both metaphor types are annotated as properties of tokens, suitable for SMD. In this work, we demonstrate the shortcomings of this formulation, and argue that conventional metaphoricity is best treated not as a property of word occurrences in a sentence, but of word senses in a lexicon. With this in mind, we investigate the problem of assigning metaphoricity ratings to word senses, a problem we name metaphorical polysemy detection (MPD). We build the first model of MPD, which identifies metaphorical senses in WordNet. No training data is available for this task. We design a novel training regime which utilises existing resources, which works by decomposing SMD into two steps: word sense disambiguation (WSD) and MPD (see [\[fig:overview\]](#fig:overview){reference-type="ref" reference="fig:overview"}). More specifically, we pair our MPD model with a state-of-the-art WSD model, and train them in conjunction on SMD data, treating word sense as a latent variable. To investigate the performance of our model, we establish an evaluation framework for MPD. To collect test data, we perform an annotation study, and label two subsets of WordNet for metaphoricity (\({\kappa=0.78}\)). Metaphor detection is typically evaluated using F1-score, which measures how well a model can judge metaphoricity in absolute terms. For MPD, we additionally introduce a new quantity for evaluation, which we call . It measures whether a model is able to correctly identify whether one sense is more metaphorical than another, even if it is unable to correctly determine where they sit around a threshold. It is calculated using ROC-AUC. On one of our test sets, consisting of words from a large resource of conventional metaphors, our MPD model attains \(.78\) ROC-AUC and \(.60\) F1, significantly higher than a strong baseline which uses a state-of-the-art SMD model, which scores \(.65\) and \(.54\) respectively. Additionally, for SMD on conventional metaphors, when our model is paired with a WSD model it attains \(.659\) F1, significantly better than the state-of-the-art SMD model (\(.626\)).
# Metaphor in the Lexicon
A fundamental feature of language is , the phenomenon of a wordform exhibiting multiple meanings which are systematically related. When meanings are related by metaphorical similarity, the phenomenon is called . An example is the word , which describes either a physical confrontation, or---metaphorically---heated criticism (e.g. "she my arguments"). Instances of metaphorical polysemy are called . In this section, we describe how conventional metaphors materialise in lexica ([2.1](#sec:metwordsense){reference-type="ref" reference="sec:metwordsense"}). With this in mind, we illustrate the shortcomings of the existing metaphor detection task ([2.2](#sec:md_problems){reference-type="ref" reference="sec:md_problems"}), and propose an alternative formulation which is more suitable for conventional metaphors ([2.3](#sec:mpd_solutions){reference-type="ref" reference="sec:mpd_solutions"}).
## Metaphoricity of Word Senses {#sec:metwordsense}
In a lexicon, metaphoricity can be formulated as a property of word senses. Consider the example definitions of the word in WordNet, which are shown in [\[tab:wordnet_met\]](#tab:wordnet_met){reference-type="ref" reference="tab:wordnet_met"} (some senses are excluded for brevity). Sense is literal, but the other senses are conventional metaphors. The conventionality of a metaphor lends itself to treatment as a continuous property. For instance, a metaphor used by an entire language community would be highly conventionalised, while a metaphor used by speakers in a particular geographic region would be moderately conventionalised. sit at the extreme other end of the spectrum: they are creative expressions with ad-hoc meanings.[^1] By definition, lexica designed to encode widespread language use will only encode the most conventional of metaphors as word senses. In order to formulate metaphoricity as a property of word senses, a lexicon needs to encode the literal and metaphorical meanings of a wordform as separate senses. The criteria that need to be met for two meanings to be considered distinct is highly contentious, with some lexicographers arguing that senses should be represented as spectra rather than discrete units. In practice, however, lexica describe senses as discrete objects. Moreover, in most lexica the granularity of these senses is high enough to separate metaphorical senses from literal ones, which thus matches our criteria. That being said, no existing lexicon systematically labels senses for metaphoricity (the decision of whether to label metaphorical senses as such is left to the lexicographer's discretion).
If word senses in a lexicon like WordNet were labeled for metaphoricity, this would be useful in a number of areas of study. When a language community need to reference a newly-arisen meaning, instead of creating a new wordform, often an existing word is extended metaphorically (e.g. a computer ). This process of metaphorical extension is a cornerstone of lexical-semantic language change, and knowing which word senses were metaphorical would open up new possibilities to study this phenomenon. The metaphors which do enter the lexicon participate in sophisticated patterns, preserving the structure of their literal domains: just as one can an opponent in a debate, so too can claims be , different rhetorical adopted, and so on. These patterns are known as , and knowing which word senses in a lexicon were metaphorical would create new opportunities to analyse their substructures.
## Issues with Standard Metaphor Detection {#sec:md_problems}
In the (SMD) task, metaphoricity is formulated as a property of a token in a sentence. The dataset which is almost universally used for this is the VUAMC. It contains tokenised English sentences, in which words are annotated with binary metaphoricity labels. It was the subject of two recent metaphor detection shared tasks. The VUAMC was annotated following an adapted version of the Metaphor Annotation Procedure. To annotate each token, MIP involves a three-stage process:
1. Establish the contextual meaning of the word, based on the other words in the sentence. [\[mip1\]]{#mip1 label="mip1"}
2. Determine whether the word has a more basic meaning which occurs in different contexts (where "basic" meanings are defined as those more concrete, related to bodily action, less vague, and/or historically older). [\[mip2\]]{#mip2 label="mip2"}
3. If the word does have a more basic meaning in different contexts, decide whether the meaning in this context can be understood in comparison with the more basic meaning. [\[mip3\]]{#mip3 label="mip3"}
If the wordform does have a more basic meaning which the contextual meaning relates to, the token is labeled as a metaphor. To perform well in SMD, a model must emulate this procedure. There are two drawbacks to this task formulation:
#### Learning WSD Implicitly
For conventional metaphors, stage [\[mip1\]](#mip1){reference-type="ref" reference="mip1"} of MIP is very similar to (WSD). In WSD, the goal is to identify the sense evoked by a wordform in a sentence. This is a hard task in its own right, and, in order to perform well, models of SMD are expected to learn to do this implicitly. We can check whether an SMD model has learnt to do this by analysing its predictions for wordforms which appear in multiple distinct contexts which all evoke the same sense. Tokens which evoke the same sense should always receive the same metaphoricity prediction (either they should all be metaphorical, or all literal). This is because, for conventional metaphors, metaphoricity is a property of a word sense ([2.1](#sec:metwordsense){reference-type="ref" reference="sec:metwordsense"}). After the sense of a token is established, stages [\[mip2\]](#mip2){reference-type="ref" reference="mip2"} and [\[mip3\]](#mip3){reference-type="ref" reference="mip3"} of MIP should always yield the same prediction; if a model makes inconsistent predictions for different invocations of the same sense, that suggests that it is incorrectly establishing the contextual meaning of the token. To investigate whether a state-of-the-art model is able to adequately perform stage [\[mip1\]](#mip1){reference-type="ref" reference="mip1"} of MIP, we perform an error analysis of MelBERT. More specifically, we compute MelBERT's metaphoricity predictions for every token in an English sense-tagged corpus, then extract all the predictions of the same word sense, which should always elicit a consistent prediction. [\[tab:melbert_preds\]](#tab:melbert_preds){reference-type="ref" reference="tab:melbert_preds"} shows MelBERT's predictions for different instances of the word which all evoke the same metaphorical sense, .[^2] MelBERT misclassifies the bottom three examples as literal. In general, for word senses which occur multiple times in the data, MelBERT gave contradictory classifications \(31.8\%\) of the time. It becomes more likely that a system will make an inconsistent classification when there are more occurrences of a word sense in the data: for word senses which occurred \(15\) or more times, the misclassification rate was \(49.3\%\). This error analysis suggests that MelBERT is not performing stage [\[mip1\]](#mip1){reference-type="ref" reference="mip1"} of MIP adequately. The errors in [\[tab:melbert_preds\]](#tab:melbert_preds){reference-type="ref" reference="tab:melbert_preds"} result not from a failure to reason about metaphoricity, but from MelBERT incorrectly establishing the contextual meaning of in the first place. We argue that this is beyond the remit of metaphor detection, since it is a hard challenge in its own right, covered by other NLP tasks.
#### Word's Senses Undefined
After stage [\[mip1\]](#mip1){reference-type="ref" reference="mip1"} of MIP has been performed, stages [\[mip2\]](#mip2){reference-type="ref" reference="mip2"} and [\[mip3\]](#mip3){reference-type="ref" reference="mip3"} require information about how the contextual meaning of a wordform relates to the other senses it may take. More specifically, models need to determine the most basic sense of the wordform, and relate that sense to the contextual meaning. However, the VUAMC does not provide information about other meanings a wordform could assume. Because of this, model architects typically assume that static word embeddings will represent the basic sense of a wordform. We argue that this assumption is problematic, because static word embeddings do not distinguish between word senses, and are sensitive to occurrence frequency. If one sense is much more frequent than others, static word embeddings will primarily encode that sense---but the basic sense of a wordform is often not the most frequent. Fortunately, we know that the most basic sense of a wordform will be one of its senses in a lexicon. If we model metaphoricity as a property of word senses, we create scope for researchers to explicitly investigate the relations between different senses, which we argue is the core challenge posed by metaphor.
## A New Task Formulation for Conventional Metaphor {#sec:mpd_solutions}
Let \(w\in\mathcal{W}\) be all the wordforms in a language. To represent metaphoricity, let \(\mathcal{M}\) be the set \(\{0, 1\}\) with members \(m\), such that \(1\) represents metaphorical and \(0\) represents literal. A token \({t\in\mathcal{T}}\) is an occurrence of a wordform in a sentence. Using \(\circ\) to denote concatenation, we define a token as a wordform \(w\) surrounded by a prefix and suffix string (\(\mathbf{p}\) and \(\mathbf{s}\)) in a sentence: \[\begin{aligned}
\label{eq:def_context} \mathcal{T} = \{\langle \mathbf{p}, w, \mathbf{s} \rangle \mid \mathbf{p} \circ w \circ \mathbf{s} \in \mathcal{W}^* \} \end{aligned}\] In SMD, the goal is to determine whether a token is metaphorical, i.e. to construct a model of \({p(m \mid t)}\). A word sense is a particular meaning of a word, which are typically represented in lexica as definitions. Let \(\mathcal{D}\) be the complete set of definitions in a lexicon (in WordNet these correspond to synsets), with elements \(d\). A lexicon maps a wordform \(w\) to a subset of the definitions \({\mathcal{D}_w\subseteq\mathcal{D}}\).[^3] Some lexica (WordNet included) associate multiple synonymous wordforms with the same definition; because of this, we define a word sense \(s\in\mathcal{S}\) as a tuple consisting of a wordform and a definition: \[\begin{aligned}
\label{eq:def_senses} \mathcal{S} = \{\langle w, d \rangle \mid w \in \mathcal{W}, d \in \mathcal{D}_w \} \end{aligned}\] Because conventional metaphors are lexicalised, we propose that the task of detecting conventional metaphors should be formulated as modelling the metaphoricity of word senses, i.e. constructing models of \({p(m\mid s)}\). We call this (MPD). This formulation of conventional metaphoricity alleviates the problems with SMD outlined in [2.2](#sec:md_problems){reference-type="ref" reference="sec:md_problems"}.
# Learning Conventional Metaphor {#sec:model}
In this section, we describe the first MPD model ([3.1](#sec:wn_met_pred){reference-type="ref" reference="sec:wn_met_pred"}). To train it, it is embedded into an SMD pipeline ([3.2](#sec:model_overview){reference-type="ref" reference="sec:model_overview"}).
## An Architecture for MPD {#sec:wn_met_pred}
To model MPD (see [2.3](#sec:mpd_solutions){reference-type="ref" reference="sec:mpd_solutions"}), we implement \({p(m\mid s)}\) as a multi-layer perceptron (MLP). In [\[eq:def_senses\]](#eq:def_senses){reference-type="ref" reference="eq:def_senses"}, we defined a word sense \(s\) as a tuple consisting of a wordform \(w\) and a definition \(d\). Our MPD model's MLP is input with a pair of embeddings, for \(w\) and \(d\). Suppose we have a \(k\)-dimensional embeddings space for words \({w\in\mathcal{W}}\), and a \(k\)-dimensional embedding space for definitions \({d\in\mathcal{D}}\). Let us define two functions to map inputs into these spaces, \({\texttt{TypeEmb}: \mathcal{W} \rightarrow {\mathbb{R}} ^{k}}\) and \({\texttt{SynsetEmb}: \mathcal{D} \rightarrow {\mathbb{R}} ^{k}}\).[^4] We concatenate these embeddings, then pass them through an MLP and a sigmoid to get a probability distribution. Letting \(\texttt{MLP}_{\boldsymbol \psi}^{\langle u,v \rangle}: \mathbb{R}^u \rightarrow \mathbb{R}^v\) be an MLP parameterised by \({\boldsymbol \psi}\) with an input size \(u\) and output size \(v\), we have \[\begin{aligned}
& \mathbf{h} = \texttt{TypeEmb}(w)\circ \texttt{SynsetEmb}(d) \nonumber \\ & p_\vtheta (m\mid s) = \sigma \big(\texttt{MLP}^{\langle 2k, 1 \rangle}_{\boldsymbol \theta} \big( \mathbf{h} \big)\big) \end{aligned}\]
## Teasing MPD out of SMD using WSD {#sec:model_overview}
Training this model would be trivial if we had \(\langle m,s \rangle\) tuples, but no training data of this form currently exists. In WSD, the task is to produce models of \({p(s\mid t)}\) for a corpus \({\langle s, t \rangle~\in~\mathcal{C}_\text{WSD}}\). In SMD, on the other hand, the task is to produce models of \({p(m\mid t)}\), using a corpus \({\langle m,t \rangle~\in~\mathcal{C}_\text{SMD}}\). No corpus contains both \(m\) and \(s\) annotations in parallel. In this section, we describe a training procedure that works by pairing an MPD model with a WSD model, and training them in conjunction on SMD, thus making it possible to use existing resources of \(\mathcal{C}_\text{WSD}\) and \(\mathcal{C}_\text{SMD}\) to train MPD. We do not believe that this procedure is an optimal way to learn MPD, as errors in the WSD component will degrade MPD performance. However, given that these are the only resources currently available, we design a method to use them for MPD, and in doing so empirically test whether they can result in a sufficiently good MPD model. By marginalising out word sense, we can decompose SMD into two parts, introducing sense \(s\) as a latent variable. The first part disambiguates senses (WSD), and the second predicts metaphoricity based on the sense (MPD): \[\begin{aligned}
\label{eq:joint} \smdcolour{\underbrace{\black{\vphantom{\sum}p_{\langle \vtheta, \vphi \rangle} (m\mid t)}}_{\substack{\text{SMD}}}} = \sum_{s\in \mathcal{S}} \mpdcolour{\underbrace{\black{\vphantom{\sum}p_\vtheta (m\mid s)}}_{\text{MPD}}} ~\wsdcolour{\underbrace{\black{ \vphantom{\sum}p_\vphi(s\mid t)}}_{\text{WSD}}} \end{aligned}\] where \({\boldsymbol \theta}\) and \({\boldsymbol \phi}\) are disjoint sets of parameters (colours correspond with [\[fig:overview\]](#fig:overview){reference-type="ref" reference="fig:overview"}). Introducing WSD into SMD like this makes explicit what SMD models typically attempt to learn implicitly (see [2.2](#sec:md_problems){reference-type="ref" reference="sec:md_problems"}). The WSD component outputs a probability distribution over the senses \({s\in\mathcal{S}}\), and the MPD component outputs the probability that each \(s\) is metaphorical.[^5] This formulation assumes that metaphoricity is conditionally independent of token given sense; we assume that the context of a word's usage tells us its sense, and that that alone is enough to predict metaphoricity. Taken as a whole, the combined model in [\[eq:joint\]](#eq:joint){reference-type="ref" reference="eq:joint"} can be trained end-to-end by minimising its cross-entropy on an SMD dataset, \[\begin{aligned}
\label{eq:met_loss} \calL_\text{SMD}({\boldsymbol \theta}, {\boldsymbol \phi}) =-\sum_{\mathclap{\substack{ \langle m, t \rangle \\ \in \mathcal{C}_\text{SMD}}}} \frac{\log p_{\langle \vtheta, \vphi \rangle} (m\mid t)}{|\mathcal{C}_\text{SMD}|} \end{aligned}\] However, this approach to training would leave the model to infer \(s\) implicitly. Instead, we can complement this objective with another, which trains the WSD model in isolation on another dataset (\(\mathcal{C}_\text{WSD}\)), also using cross-entropy, \[\begin{aligned}
\label{eq:wsd_loss} \calL_\text{WSD}(&{\boldsymbol \phi}) =-\sum_{\mathclap{\substack{ \langle s, t \rangle \\ \in \mathcal{C}_\text{WSD}}}} \frac{ \log p_\vphi (s\mid t) }{|\mathcal{C}_\text{WSD}|} \end{aligned}\] These two objectives can be combined into a multi-task objective, using \(\alpha\) as a hyperparameter which regulates the ratio between them, yielding the final global loss function \[\begin{aligned}
{3}\label{eq:joint_loss} \calL({\boldsymbol \theta}, {\boldsymbol \phi}) =& &&~~\alpha & & \cdot \calL_\text{SMD}({\boldsymbol \theta}, {\boldsymbol \phi})~+ \\ & (1 &&-\alpha) & & \cdot \calL_\text{WSD}({\boldsymbol \phi}) \nonumber \end{aligned}\] Any supervised WSD architecture can be used in conjunction with our MPD model.
# Evaluation {#sec:evaluation}
In this section, we describe the evaluation data we collected for MPD ([4.1](#sec:eval_sets){reference-type="ref" reference="sec:eval_sets"}), and introduce relative metaphoricity, an aspect of metaphor which it is only possible to evaluate with MPD ([4.2](#sec:relative_met){reference-type="ref" reference="sec:relative_met"}). We then describe our experimental setup ([4.3](#sec:setup){reference-type="ref" reference="sec:setup"}) and results ([4.4](#sec:results){reference-type="ref" reference="sec:results"}).
## Evaluation Data for MPD {#sec:eval_sets}
[\[sec:wn_pilot\]]{#sec:wn_pilot label="sec:wn_pilot"} Our approach to training ([3.2](#sec:model_overview){reference-type="ref" reference="sec:model_overview"}) alleviates the need for labeled MPD training data. However, we still need evaluation data. For this we collect annotation. In this section, we describe the main evaluation set we collected. This evaluation set consists of data from the Master Metaphor List, which is the largest list of conventional metaphors.
#### Data for Annotation
We sample \(250\) commonly cited metaphor examples from the MML, filtering out wordforms which have more than \(10\) or fewer than \(3\) senses. We annotate the remaining \(98\) wordforms for metaphoricity (a total of \(554\) word senses).
#### Annotation Procedure
We adapt MIP stages [\[mip2\]](#mip2){reference-type="ref" reference="mip2"} and [\[mip3\]](#mip3){reference-type="ref" reference="mip3"} for the annotation task (our guidelines are presented in [\[sec:guidelines\]](#sec:guidelines){reference-type="ref" reference="sec:guidelines"}). For MPD, we require a lexicon which consistently separates literal and metaphorical senses (see [2.1](#sec:metwordsense){reference-type="ref" reference="sec:metwordsense"}). We use the annotation phase to also check whether WordNet complies with this. Along with options for "metaphorical" and "literal", annotators also have the option to label senses as "mixed", meaning that the sense combined literal and metaphorical definitions. Our data is annotated by two judges, who are native speakers of English.
#### Analysis
At least one of the annotators labeled a sense as "mixed" \(5.4\%\) of the time. We conclude from this that although WordNet mixes some literal and metaphorical senses, for the most part the granularity of the sense inventory was high enough to separate them. Of those which neither thought was mixed, inter-annotator agreement was measured at \({\kappa=0.78}\) (\(N{=}539, n{=}2, k{=}2\)): where the senses were clearly distinct, annotation was consistent. Of the senses that the judges agreed upon, \(52\%\) were metaphorical and \(48\%\) were literal.
## Relative Metaphoricity {#sec:relative_met}
Metaphor detection performance is typically reported using F1-score, which factors in the precision and recall of a classifier. This evaluates a model's ability to judge whether the metaphoricity level of an input passes a threshold to be considered metaphorical, in an absolute sense. With absolute metaphoricity judgments, there are lots of edge cases, which do not cleanly fit into a metaphorical--literal binary. For example, while a physical (e.g. an outlaw with a knife) is widely accepted as being literal, and an with words (e.g. in the perpetrator's court hearing) is widely accepted as being metaphorical, the case is less clear-cut for a sporting (e.g. when they are playing football with other inmates): here we have a physical act which can result in injury, but that takes place within the confines of a game, and certainly does not involve weapons.[^6] It is difficult to decide whether an edge case like this is metaphorical or not, if those are the only options. Most people would likely agree, however, that a sporting attack is *more* metaphorical than an attack with a knife, and *less* metaphorical than an attack within a debate. We are not the first to note these flaws in absolute metaphoricity judgements; they have previously been used to motivate proposals for the treatment of metaphor on graded or continuous scales . Even though MPD is a binary classification task, it additionally gives us the opportunity to evaluate whether a model can judge which of a word's senses are more metaphorical than others, even if it gets the exact threshold wrong. We call this a measure of . Beneath the surface, models assign a probability of metaphoricity, rather than an absolute judgement. With MPD, we can compute a model's predictions for all the senses of a word, and investigate whether it puts these probabilities in the right order, even if it gets absolute predictions "wrong" in an absolute sense. A metric for this is described in [4.3](#sec:setup){reference-type="ref" reference="sec:setup"}.
## Experimental Setup {#sec:setup}
In this section, we describe the models and data we experimented with. Additional details can be found in the appendices.
#### MPD Models
We compare our MPD model (**Ours**) to several baselines. To establish a lower-bound of performance for MPD, we compute two baselines which make predictions randomly or by choosing the most common class (**Random** and **Majority**). In addition to these, we compute a strong baseline: using a state-of-the-art SMD model, we compute the metaphoricity probability of all occurrences of \(s\) in WSD data, and take the average (**MelBERT Average**). More formally, let \(t \in \mathcal{C}_\text{WSD}^{s}\) be every token in the WSD data which elicits a specific word sense \(s\). We compute the average metaphoricity probability of that word sense by taking the mean of MelBERT's metaphoricity prediction for all tokens that evoke that sense: \[p(m\mid s) = \mean_{t \in \mathcal{C}_\text{WSD}^{s}} \underbrace{p(m \mid t)}_{\substack{\text{MelBERT}}}\] If \(|\mathcal{C}_\text{WSD}^{s}|=0\) (i.e. there are no occurrences of \(s\) in the WSD data) then we default to the random baseline (and take the probability as random).
#### WSD Models
We experiment with two WSD models: a baseline (**BERT WSD Baseline**) and a state-of-the-art approach. Additional description is left to [\[sec:wsd_models\]](#sec:wsd_models){reference-type="ref" reference="sec:wsd_models"}.
#### SMD Models
We train the combined model (**WSD+MPD**) from [\[eq:joint\]](#eq:joint){reference-type="ref" reference="eq:joint"}, using our MPD model and both WSD models (we select the highest performing variant). We compare the overall SMD performance of this combined model to two other approaches, a baseline (**BERT SMD Baseline**) and a high-performing model (**MelBERT**). Additional description is left to [\[app:smd_models\]](#app:smd_models){reference-type="ref" reference="app:smd_models"}.
#### MPD Data
To evaluate MPD performance, we use the evaluation set described in [4.1](#sec:eval_sets){reference-type="ref" reference="sec:eval_sets"} (**MML**). We also collected a second set using the same procedure, designed to evaluate model generalisation ability to words which are out-of-vocabulary of the training data (**OOV**). We randomly sampled \(100\) wordform from the WordNet vocabulary which are not included in the vocabularies of \(\mathcal{C}_\text{SMD}\) or \(\mathcal{C}_\text{WSD}\), and asked one of the annotators from [\[sec:wn_pilot\]](#sec:wn_pilot){reference-type="ref" reference="sec:wn_pilot"} to follow the same procedure as before. The final evaluation sets for MML and OOV consist of \(535\) and \(492\) word senses respectively, after senses which the annotator labeled as "mixed" were removed. As an additional third set, we use the data of, who annotated verbs in WordNet for metaphoricity (**Verbs**). Their data consists of \(1{,}679\) word sense annotations, covering \(440\) verbal wordforms.
#### WSD Data
For \(\mathcal{C}_\text{WSD}\), we use SemCor. In SemCor each token is annotated with the WordNet sense it evokes. We remove all datapoints with trivial solutions (for which there is only a single sense to choose from, i.e. \({|\mathcal{D}_w|=1}\)).
#### SMD Data
For \(\mathcal{C}_\text{SMD}\), we use the VUAMC (see [2.2](#sec:md_problems){reference-type="ref" reference="sec:md_problems"}). We remove all part-of-speech types and words not in WordNet (meaning we remove all prepositions, which are almost always labeled as metaphorical in VUAMC, and so are easy to predict). Additionally, because our focus is on conventional metaphors, when investigating SMD performance, we also experiment using a subset of this data in which novel metaphors are filtered out (we refer to this subset as **Conventional**, as opposed to **All**). We achieve this using the annotation layer of.[^7]
#### Metrics
To evaluate MPD for absolute metaphoricity, we compute the F1-score. For relative metaphoricity, we compute (for each wordform individually) the area under a receiver operating characteristic curve (ROC-AUC, see [@fawcett2004roc] for discussion). An ROC curve is a plot of the true positive rate against the false positive rate, as a threshold shifts from \(0\) to \(1\). This tells us whether the metaphoricity probabilities assigned to the senses of a wordform are in correct high--low ordering, even if they are not properly calibrated around the \(.5\) threshold. ROC-AUC values range from \(1\) (perfectly correct) to \(0\) (perfectly incorrect), where \(.5\) indicates no correlation. To get a value for a whole evaluation set, we take the mean of the ROC-AUC scores of all wordforms in the set. To evaluate SMD, we compute the F1-score, and for WSD, where the goal is multi-class classification, we compute the micro-averaged F1-score.
#### Significance Testing
We use a two-tailed Monte Carlo permutation test with \(r=1{,}000\) permutations. We experiment with two significance levels, \(\alpha=0.05\) and \(0.01\), and differentiate between these by indicating the significance level using \(^*\) and \(^{**}\) respectively.
## Results {#sec:results}
#### Metaphorical Polysemy Detection
MPD results are presented in [\[tab:mpd_results\]](#tab:mpd_results){reference-type="ref" reference="tab:mpd_results"}. We compare two versions of our MPD model, trained in combination with different WSD implementations. In all experimental settings (each column), a variant of our model performs significantly better than the highest-performing baseline. Our model's results are highest for the MML set (\(.78\) ROC-AUC and \(.60\) F1), perhaps because the examples here exhibited the clearest cases of metaphoricity. In general, the EWISER WSD variant's results are higher for relative metaphoricity, whilst the baseline WSD variant's results are higher for absolute metaphoricity. The reason for this is unclear. Results are lowest for the OOV set, which shows that the model has difficulties generalising beyond training data. Generalising to never-before-seen senses is a well-known issue in WSD: themselves note that their model, EWISER, relies too much on corpus supervision. This problem will affect MPD models if they are trained with our methodology ([3.2](#sec:model_overview){reference-type="ref" reference="sec:model_overview"}); to improve on MPD, it will be necessary to collect dedicated training data. Nevertheless, our results show that it is possible to learn MPD using existing resources.
#### Standard Metaphor Detection
SMD results are presented in [1](#tab:md_results){reference-type="ref" reference="tab:md_results"}, for both variants of \(\mathcal{C}_\text{SMD}\) (one with only conventional metaphors, one with any). When trained and evaluated on only conventional metaphors, our combined MPD+WSD model significantly outperforms the state-of-the-art, scoring \(.659\) F1 (compared to MelBERT's \(.626\)). When all metaphors are included, MelBERT's results are higher that our model (\(.638\) compared to \(.631\)), but in this case the difference is insignificant (at both significance levels). That our model's performance is diminished in this setting is unsurprising, since this data will contain novel metaphors which are not encoded as senses in WordNet. We expect that with improvements to WSD generalisation, it should be possible to improve the SMD performance of a combined MPD+WSD model, for both variants. As mentioned above, generalising to unseen senses is a common issue in WSD. Only \(9.4\%\) of words in the SMD (All) test set have all their senses covered in the WSD training data, which is likely to have a substantial effect on the combined model's performance.
#### Word Sense Disambiguation
[2](#tab:wsd_results){reference-type="ref" reference="tab:wsd_results"} compares the performance of WSD models which are paired with MPD models and trained jointly on SMD, as opposed to trained on WSD alone. Although the auxiliary SMD objective leads marginal numerical differences (\(-0.02\) and \(+0.01\) for EWISER and the baseline WSD model respectively), in both cases these differences are not significantly distinguishable (at either significance level), suggesting that utilising SMD data to train WSD models is not useful. This is likely because of the relatively small overlap between the vocabularies of the datasets.
# Conclusion and Future Work {#sec:discussion}
In this paper, we argued that the standard metaphor detection (SMD) task in NLP is ill-suited to conventional metaphor, because it conflates metaphor detection with word sense disambiguation (WSD). As an alternative, we proposed metaphorical polysemy detection (MPD). We constructed the first MPD model, which identifies conventional metaphors in WordNet. To train it, we employed a novel training technique which exploits resources designed for SMD and WSD. To evaluate MPD, we collected two sets of evaluation data, and proposed a new performance measure based on relative judgements of metaphoricity. Our model performed significantly better than a state-of-the-art SMD model in all MPD evaluation settings (e.g. attaining \(.78\) ROC-AUC and \(.60\) F1 on a subset based on the master metaphor list). Additionally, we found that pairing our model with a WSD model led to state-of-the-art results for token-based conventional metaphor detection (\(.659\) compared to \(.626\)). We make our code and MPD evaluation data publicly available.[^8] Making serious improvements in MPD will require the collection of dedicated training data. Training MPD using SMD data necessitated the inclusion of a WSD model, which will negatively affect MPD performance when it makes mistakes. Additionally, our model did not utilise information about the set of definitions a wordform is associated with (\(\mathcal{D}_w\)), and instead made the naïve assumption that this information will be encoded in a static embedding of \(w\). Set-and graph-based architectures which exploited this information would be a natural thing to explore. With more data and better models, it may be possible to synthesise a full set of WordNet metaphoricity annotation, and even to extend this synthetic annotation to multilingual WordNet versions. This would be a valuable resource, which would open the door to study other questions surrounding metaphorical polysemy. | {'timestamp': '2022-12-19T02:09:51', 'yymm': '2212', 'arxiv_id': '2212.08395', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08395'} |