text
stringlengths 0
643k
⌀ | meta
stringlengths 137
151
⌀ |
|---|---|
# Introduction
It is known that flavor-changing neutral current (FCNC) processes of long-lived particles would provide a window to observe new physics (NP) beyond the standard model (SM). These particles, such as ground-state mesons of \(K,~B,~D\) and \(B_c\), and baryons of \(\Lambda_{c,b}\) and \(\Xi_{c,b}\), decay through weak interactions, resulting in longer lifetimes and narrower decay widths. These FCNC decays may benefit the detections of NP. Hadronic FCNC decays include \(c\to u\), \(s\to d\), \(b\to d\), and \(b\to s\) processes at quark level. In the SM, dilepton FCNC modes have been widely studied theoretically and experimentally . However, the neutrino (\(\nu\)) and anti-neutrino (\(\bar \nu\)) in the final states of the decays cannot be directly detected, but are treated as missing energy (\(\slashed E\)) in experiments. So far most experiments can only obtain the upper limits on the decay branching ratios associated with \(\nu\bar\nu\) . The experimental searches have given the strictest constraints on kaon FCNC decays. Recently, the upper bound on \(K_L\to\pi^0\bar\nu\nu\) from the KOTO experiment at J-PARC has been given to be \(\mathcal {B}(K_L\to\pi^0\bar\nu\nu)_{\rm KOTO}<3.0\times 10^{-9}\) at \(90\%\) confidence level (C.L.), which is slightly greater than the SM prediction of \(\mathcal {B}(K_L\to\pi^0\bar\nu\nu)_{\rm SM} = (3.4 \pm 0.6)\times10^{-11}\) . On the other hand, the decay of \(K^+\to \pi^+\bar\nu\nu\) has been measured, namely, \(\mathcal {B}(K^+\to\pi^+\bar\nu\nu)_{\rm NA62}=(11.0^{+4.0}_{-3.5}({\rm stat})\pm 0.3(\rm syst))\times 10^{-11}\) at \(68\%\) C.L. from the NA62 experiment at CERN and \(\mathcal {B}(K^+\to\pi^+\bar\nu\nu)_{\rm E949}= (17.3^{+11.5}_{-10.5})\times 10^{-11}\) from the E949 experiment at BNL . These results are consistent with the SM prediction of \(\mathcal {B}(K^+\to\pi^+\bar\nu\nu)_{\rm SM}= (8.4\pm1.0)\times 10^{-11}\) within one standard deviation. It is clear that the room for NP in \(K \to \pi\slashed E\) has become quite small. However, the searches for NP in the FCNC decay processes of charmed and bottomed mesons would still be possible. For example, the charmed meson and hyperon decays associated with \(\slashed E\) have been analyzed in Ref. . The bottomed ones have been measured in Refs. , which are listed in Table [5](#exp){reference-type="ref" reference="exp"} along with the SM predictions .
As seen from Table [5](#exp){reference-type="ref" reference="exp"}, the differences between the first and second columns indicate that there are some rooms for new invisible particles of \(\chi\) (shown as the third column) emitted in such processes. In Refs. , the effects of the invisible particles with various spins in the \(b\to s(d)\) transitions have been explored. While most of these previous studies in the literature are focused on bottomed mesons. There have been no related researches on bottomed baryons up to now. In this paper, we generalize the experimental upper bounds from \(B\) mesons to the corresponding decay modes of bottomed baryons, namely, \(\Lambda_b\) and \(\Xi_b\). On experiments, the numbers of \(\Lambda_b\) and \(\Xi_b\) produced by LHC between 2011 and 2018 are of the order of \(10^{11}\) and \(10^{9}\) , respectively. In addition, LHC is expected to produce \(10\) times more \(b\) quarks in \(run~3\) (2022-2026) . Clearly, in the near future, experiments on bottomed baryons can give more results related to the invisible particles. In this work, we consider the bottomed baryonic FCNC decays of \({\bf B}_b\to\mathcal {\bf B}_n\chi\chi\), where \({\bf B}_{n(b)}\) are (bottomed) baryons and \(\chi\) represent light invisible particles, which are assumed to be Majorana fermions. Phenomenologically, these new invisible fermions of \(\chi\) can weakly interact with the SM fermions via a mediator, which can be a scalar , pseudoscalar , vector or axial-vector particle. In our study, we will concentrate on a general model-independent approach to introduce the effective Lagrangian, which contains all possible currents involving the invisible fermions with the coupling constants extracted from the experiments. The paper is organized as follows: In Sec. II, we obtain the SM expectations of \({\bf B}_b\to\mathcal {\bf B}_n\bar\nu\nu\). In Sec. III, we first construct the effective Lagrangian, which describes the coupling between the quarks and light invisible fermions. We then present the numerical results of the upper limits for the decay branching ratios of \({\bf B}_b\to\mathcal {\bf B}_n\chi\chi\). The hadronic transition matrix elements are evaluated based on the QCD light-cone sum rules (LCSR) and modified bag model (MBM). Finally, we give the conclusion in Sec. IV.
# The SM expectations
The FCNC decay processes of bottomed baryons with missing energy are described in Fig. [\[Feyn\]](#Feyn){reference-type="ref" reference="Feyn"}, where \({\bf B}_b\) and \({\bf B}_n\) represent the initial and final baryons, \(q=b\) and \(q_{f}=s(d)\) are initial and final quarks, and \(q_{2(3)}\) are the spectator quarks, respectively. In the SM, there is no tree-level contribution to the FCNC decays of \({\bf B}_b \to {\bf B}_n\bar\nu\nu\). The first-order contributions to these processes come from the penguin and box diagrams as shown in Fig. [\[Feyn\]](#Feyn){reference-type="ref" reference="Feyn"}a, which can be described by the effective Lagrangian, given by \[\begin{aligned} \mathcal L_{\bar\nu\nu} = \frac{4G_F}{\sqrt{2}}\frac{\alpha}{2\pi\sin^2\theta_W}\sum_{\ell=e, \mu, \tau}\sum_{q=u, c, t}V_{bq}V_{sq}X^\ell(x_q)(\bar s_{_L}\gamma^\mu b_{_{L}})(\bar\nu_{_{\ell L}}\gamma_\mu\nu_{_{\ell L}}), \label{eq25} \end{aligned}\] with \[\begin{aligned} X^\ell(x_q) = \frac{x_q}{8}\left[\frac{x_q+2}{x_q-1}+\frac{3(x_q-2)}{(x_q-1)^2}\ln x_q\right], \label{eq26} \end{aligned}\] where \(G_F\) represents the Fermi coupling constant, \(\alpha\) corresponds to the fine structure constant, \(\theta_W\) stands for the Weinberg angle, \(V_{ij}\) are the CKM matrix elements, and \(x_q=m_q^2/M_W^2\) with \(m_q\) (\(M_W\)) being the mass of the quark (\(W\)-boson). Consequently, the transition amplitude is given by \[\begin{aligned} \langle {\bf B}_n\bar\nu\nu|\mathcal L_{\bar\nu\nu}| {\bf B}_b\rangle &=\frac{\sqrt{2}G_F\alpha}{4\pi\sin^2\theta_W}V_{bt}V_{st}X^\ell(x_t)\langle {\bf B}_n|\bar s\gamma^\mu(1-\gamma^5)b|{\bf B}_b\rangle\times \bar u_{\nu_\ell}\gamma_\mu(1-\gamma^5)v_{\nu_\ell}. \label{eq27} \end{aligned}\] The baryonic transition matrix elements can be parameterized by the form factors (FFs) of \(f_i^{V,A}~(i=1,2,3)\), \(f^S\) and \(f^P\), defined by \[\begin{aligned} \langle {\bf B}_n(P_f,s_f)|(\bar q_{f}\gamma_\mu q) |{\bf B}_b(P,s)\rangle &=\bar u_{ _{{\bf B}_n}} (P_f,s_f)\left[\gamma_\mu f^V_1(q^2) + i\sigma_{\mu\nu} \frac{q^\nu}{M} f^V_2(q^2)+\frac{q^\mu}{M}f^V_3(q^2)\right]u_{_{{\bf B}_b}}(P,s),\\ \langle {\bf B}_n(P_f,s_f)|(\bar q_{f} q) |{\bf B}_b(P,s)\rangle &=\bar u_{_{{\bf B}_n}} (P_f,s_f)f^S(q^2)u_{_{{\bf B}_b}}(P,s),\\ \langle {\bf B}_n(P_f,s_f)|(\bar q_{f}\gamma_\mu \gamma^5 q) |{\bf B}_b(P,s)\rangle &=\bar u_{_{{\bf B}_n}} (P_f,s_f)\left[\gamma_\mu f^A_1(q^2) + i\sigma_{\mu\nu} \frac{q^\nu}{M} f^A_2(q^2)+\frac{q^\mu}{M}f^A_3(q^2)\right]\gamma^5u_{_{{\bf B}_b}}(P,s),\\ \langle {\bf B}_n(P_f,s_f)|(\bar q_{f} \gamma^5 q) |{\bf B}_b(P,s)\rangle &=\bar u_{_{{\bf B}_n}} (P_f,s_f)f^P(q^2)\gamma^5 u_{_{{\bf B}_b}}(P,s), \label{ffb} \end{aligned}\] where \(q\) corresponds to the momentum transfer, and \(M\) is the mass of the initial baryon. We will evaluate these elements in terms of the MBM, which works well for the heavy baryonic decays . In the MBM, the baryon wave functions at rest are read as \[\label{wave_function_inrest} \Psi(x_{q_1},x_{q_2},x_{q_3}) ={\cal N}\int d^3 \vec{x} \prod_{i=1,2,3} \phi_{q_i}(\vec{x}_{q_i}-\vec{x}) e^{-iE_{q_i}t_{q_i}} \,,\] where \(q_i\) are the quark components of the baryons, \({\cal N}\) the overall normalization constant, \(x_{q_i}\) (\(E_{q_i}\)) the spacetime coordinates (energies) of \(q_i\), and \(\phi_{q_i}(x)\) the quark wave functions inside a static bag, located at the center, given by \[\label{quark_wave_function} \phi_{q}(\vec{x}) = \left( \begin{array}{c} \omega_{q+} j_0(p_qr) \chi_q\\ i\omega_{q-} j_1(p_qr) \hat{r} \cdot \vec{\sigma} \chi_q\\ \end{array} \right)\.\] Here, \(j_{0,1}\) represent the spherical Bessel functions, \(\omega_{q\pm} = \sqrt{T_{q} \pm M_{q} }\) with \(T_{q}\) the kinematic energies, and \(\chi_q\) are the two component spinors. By demanding that quark currents shall not penetrate the boundary of bags, we have the boundary condition \[\label{momentum_condition} \tan (p_qR) = \frac{p_qR}{1-M_qR-E_qR}\,,\] where \(R\) is the bag radius, resulting in that the magnitudes of 3-momenta are quantized, which can be analogous to the well-know infinite square well. By sandwiching the operators, we arrive \[\begin{aligned}
\label{master} &&\int \langle \Lambda | \overline{ s }\Gamma b(x) e^{iqx} |\Lambda_b \rangle d^4 x= {\cal Z}\int d^3\vec{x}_\Delta \Gamma _{sb}(\vec{x}_\Delta) \prod_{q_j=u,d } D_{q_j}(\vec{x}_\Delta)\,, \end{aligned}\] with \[\begin{aligned}
\label{con} &&{\cal Z} \equiv (2\pi )^4 \delta^4(p_{\Lambda_b}-p_{\Lambda}-q) {\cal N}_{\Lambda_b} {\cal N}_{\Lambda} \,,\nonumber\\ &&D_{q_j}(\vec{x}_{\Delta}) \equiv \sqrt{1-v^2 } \int d^3 \vec{x} \phi_{q_j}^\dagger \left(\vec{x} +\frac{1}{2}\vec{x}_{\Delta}\right) \phi_{q_j} \left(\vec{x}-\frac{1}{2}\vec{x}_{\Delta}\right) e^{-2iE_{q_j} \vec{v}\cdot\vec{x}}\,,\nonumber\\ &&\Gamma _{sb} (\vec{ x}_\Delta)=\int d^3\vec{x} \phi _s\left(\vec{x} + \frac{1}{2}\vec{x}_\Delta \right)\gamma^0 S_{-\vec{ v}}\Gamma S_{-\vec{ v}} \phi_b\left(\vec{x}-\frac{1}{2}\vec{x}_\Delta \right) e^{i(M_\Lambda + M_{\Lambda_b}-E_s-E_b)\vec{ v}\cdot\vec{ x} }\,, \end{aligned}\] where \(\Gamma\) are arbitrary Dirac matrices, and \(S_{\vec{v}}\) the Lorentz boost matrix of Dirac spinors. We have taken the initial (final) state as \(\Lambda_b\) (\(\Lambda\)) for an concrete example. To simplify the algebra, the Briet frame is chosen, where \(\Lambda_b\) and \(\Lambda\) have the velocity \(-\vec{ v}\) and \(\vec{ v}\), respectively. Notably, all the parameters of the model are extracted from the mass spectra, given as \[R = 4.8 ~\text{GeV}^{-1}\,,~~~M_{u,d} =0\,,~~~M_s = 0.28~\text{GeV}\,,~~~M_b = 5.093~\text{GeV}\.\] We consider the bottomed baryon decays of \(\Lambda_b\to\Lambda\bar\nu\nu\) and \(\Xi_b^{0(-)}\to\Xi^{0(-)}\bar\nu\nu\), \(\Lambda_b\to n \bar\nu\nu\), \(\Xi_b^{0(-)}\to \Sigma^{0(-)} \bar\nu\nu\) and \(\Xi_b^{0}\to \Lambda \bar\nu\nu\), due to the \(b\to s\) and \(b\to d\) transitions at quark level, respectively. The FFs can be extracted straightforwardly after the computations, which are shown in Fig. [\[ff\]](#ff){reference-type="ref" reference="ff"} along with their \(q^2\) dependencies. By integrating the three-body phase space, we obtain the decay branching ratio to be \[\mathcal {B} ({\bf B}_b\to{\bf B}_n\bar\nu\nu) = \frac{1}{512 \pi^3 M^3 \Gamma_{{\bf B}_b}}\int\frac {dq^2}{q^2} \lambda^{1/2}(M^2, q^2, M_f^2)\lambda^{1/2}(q^2, m_1^2, m_2^2)\int d\cos\theta\sum|\mathcal M|^2, \label{ps3}\] where \(\lambda(x, y, z)= x^2 + y^2 +z^2-2xy-2xz-2yz\) is the K\({\rm \ddot a}\)llen function, \(M\), \(M_f\), \(m_1\) and \(m_2\) correspond to the masses of the initial baryon, final baryon, neutrino and anti-neutrino, respectively, \(\theta\) is the phase space angle, \(\Gamma_{{\bf B}_b}\) represents the total width of the initial baryon, and \(\mathcal M\) stands for the amplitude. As the three generations of neutrinos are indistinguishable experimentally, the final results need to be multiplied by three. For the \(b\to s\) transition, the decay branching ratios associated with neutrino and anti-neutrino are as follows: \[\begin{aligned} \mathcal{B}(\Lambda_b\to \Lambda\bar\nu\nu)&=5.52\times 10^{-6}, \\ \mathcal{B}(\Xi_b^{0(-)}\to \Xi^{0(-)}\bar\nu\nu)&=7.80\times 10^{-6}. \label{SM1} \end{aligned}\] Here, due to the \(\rm SU(3)\) flavor symmetry, the branching ratios of \(\Xi_b^0\) and \(\Xi_b^-\) are considered approximately to be equal. Note that our results of (\(\Lambda_b\to \Lambda\bar\nu\nu\)) in Eq. [\[SM1\]](#SM1){reference-type="eqref" reference="SM1"} is smaller than the previous prediction in Ref. . Similarly, for the \(b\to d\) transition we have that \[\begin{aligned} \mathcal{B}(\Lambda_b\to n\bar\nu\nu)&=2.76\times 10^{-7}, \\ \mathcal{B}(\Xi_b^-\to \Sigma^-\bar\nu\nu)&=2.65\times 10^{-7}, \\ \mathcal{B}(\Xi_b^0\to \Lambda\bar\nu\nu)&=3.88\times 10^{-8}, \\ \mathcal{B}(\Xi_b^0\to \Sigma^0\bar\nu\nu)&=1.24\times 10^{-7}, \label{SM2} \end{aligned}\] which are about one to two orders of magnitude smaller than the modes in Eq. [\[SM1\]](#SM1){reference-type="eqref" reference="SM1"}, due to the ratio of the CKM matrix elements, \(|V_{td}/V_{ts}|\sim\mathcal O(\lambda)\). As a result, the SM predictions of bottomed baryonic FCNC processes with \(\slashed E\) are \(\mathcal O(10^{-8})-\mathcal O (10^{-6})\). If the experimental detections of these decays are larger than the values in Eqs. [\[SM1\]](#SM1){reference-type="eqref" reference="SM1"}-[\[SM2\]](#SM2){reference-type="eqref" reference="SM2"}, the new invisible neutral particles from NP are expected.
# Processes with invisible particles
## Effective Lagrangian
In Fig. [\[Feyn\]](#Feyn){reference-type="ref" reference="Feyn"}b, two spin-\(1/2\) invisible Majorana particles of \(\chi\chi\) are assumed to be emitted in the process, in which the four-fermion vertex may be generated at tree or loop level by introducing new physical mediators in specific models . Under the low energy scale, the model-independent effective Lagrangian is given by \[\begin{aligned} \mathcal L_{eff}=\sum_{i=1}^{6}{g_{_{mi}} Q_i}, \label{eq1} \end{aligned}\] where \(g_{fi}\) are the phenomenological coupling constants, which are taken at the new physical energy scale \(\Lambda\). There are \(6\) independent dimension-six effective operators, which have the forms: \[\begin{aligned} & Q_1=(\bar q_{_f}q)(\chi\chi),~~~~~~~~~Q_2=(\bar q_{_f} \gamma^{5}q)(\chi\chi),~~~~~~~~~ Q_3=(\bar q_{_f}q)(\chi\gamma^{5}\chi),\\ & Q_4=(\bar q_{_f} \gamma^{5}q)(\chi\gamma^{5}\chi),~~~Q_5=(\bar q_{_f} \gamma_{\mu}q)(\chi\gamma^{\mu}\gamma^{5}\chi),~~~ Q_6=(\bar q_{_f} \gamma_{\mu}\gamma^{5} q)(\chi\gamma^{\mu}\gamma^{5}\chi), \label{op} \end{aligned}\] where the invisible particles of \(\chi\) have been assumed to be the Majorana type. Since \(\chi\gamma^{\mu}\chi=0\) and \(\chi\sigma^{\mu\nu}\chi=0\), there is no contribution from the vector or tensor current. The upper limits of the coupling constants in the effective Lagrangian can be extracted from the difference between the theoretical predictions and experimental data of the \(B\) meson FCNC decays, such as \(B ^ {-} \to K ^-(K^{\ast-}) + \slashed E\) and \(B ^ {-} \to \pi ^-(\rho^-) + \slashed E\). For the \(0^-\to(0^-,1^-)\) mesons decays of \(M^-\to(M_f^-,M_f^{*-})\chi\chi\), only operators \(Q_{1,3,5}\) and \(Q_{2,4,5,6}\) in Eq. ([\[op\]](#op){reference-type="ref" reference="op"}) give the contributions, respectively. The amplitudes of the \(0^-\to (0^-, 1^-)\) decays can be simplified as \[\begin{aligned} \langle M_f^-\chi\chi|\mathcal L_{eff}|M^-\rangle&=2g_{m1} \langle M_f^-|(\bar q_{_f} q) |M^-\rangle\bar u_{\chi}v_{\chi}+2 g_{m3} \langle M_f^-|(\bar q_{_f} q) |M^-\rangle\bar u_{\chi}\gamma^5 v_{\chi}\\ &+2g_{m5}\langle M_f^-|(\bar q_{_f}\gamma_\mu q) |M^-\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}, \\ \label{Lmm0} \end{aligned}\] and \[\begin{aligned} \langle M_f^{*-}\chi\chi|\mathcal L_{eff}|M^-\rangle&=2g_{m2} \langle M_f^{*-}|(\bar q_{_f}\gamma^5 q) |M^-\rangle\bar u_{\chi}v_{\chi}+2g_{m4} \langle M_f^{*-}|(\bar q_{_f}\gamma^5 q) |M^-\rangle\bar u_{\chi}\gamma^5 v_{\chi}\\ &+2g_{m5} \langle M_f^{*-}|(\bar q_{_f}\gamma_\mu q) |M^-\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}+2g_{m6} \langle M_f^{*-}|(\bar q_{_f}\gamma_\mu \gamma^5q) |M^-\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}, \\ \end{aligned}\] respectively. The hadronic transition matrix elements can be expressed as \[\begin{aligned} \langle M_f^-|(\bar q_{_f} q) |M^-\rangle &= \frac{M^2-M_{f}^2}{m_q-m_{q_{_f}}}f_0 (q^2),\\ \langle M_f^-|(\bar q_{_f}\gamma_\mu q) |M^-\rangle &= (P+P_f)_{\mu}f_+ (q^2)+(P-P_f)_{\mu}\frac{M^2-M_{f}^2}{q^2} \big[f_0 (q^2)-f_+ (q^2)\big],\\ \langle M_f^-|(\bar q_{_f}\sigma_{\mu\nu} q) |M^-\rangle &=i\big[P_{\mu} (P-P_f)_{\nu}-P_{\nu} (P-P_f)_{\mu}\big] \frac{2}{M+M_f} f_T(q^2), \label{} \end{aligned}\] and \[\begin{aligned} \langle M_f^{*-}|(\bar q_{_f}\gamma^5 q) |M^-\rangle &=-i\big[\epsilon \cdot (P-P_f)\big]\frac{2M_f}{m_q+m_{q_{_f}}}A_0(q^2),\\ \langle M_f^{*-}|(\bar q_{_f}\gamma_\mu\gamma^5 q) |M^-\rangle &= i\bigg\{\epsilon_{\mu} (M+M_f)A_1(q^2)-(P+P_f)_{\mu}\frac{\epsilon \cdot (P-P_f)}{M+M_f}A_2(q^2)\\ &~~~-(P-P_f)_{\mu} \big[\epsilon \cdot (P-P_f)\big] \frac{2M_f}{q^2} \big[A_3(q^2)-A_0(q^2)\big]\bigg\},\\ \langle M_f^{*-}|(\bar q_f \gamma_\mu q )|M^-\rangle &=\varepsilon _{\mu \nu \rho \sigma} \epsilon ^\nu P^\rho (P-P_f)^\sigma\frac{2 }{M+M_f}V(q^2), \label{} \end{aligned}\] where \(m_{q_f}\) are the quark masses, \(f_j~(j=0,+,T)\), \(A_k~(k=0-3)\) and \(V\) are the FFs, which are evaluated from the method of the LCSR , and \(\epsilon\) is the polarization vector of the final meson with the convention of \(\varepsilon^{0123}=1\). In our calculation, we assume that only one operator contributes to the process at a time. By integrating the three-body phase space given in Eq. [\[ps3\]](#ps3){reference-type="eqref" reference="ps3"}, the upper limits of the coupling constants \(g_{mi}\) can be obtained from Table [5](#exp){reference-type="ref" reference="exp"}, given by \[\begin{aligned} \mathcal B(M\to M ^{(*)}_f\slashed E)_{\rm exp}-\mathcal B(M\to M ^{(*)}_f \bar\nu\nu)_{\rm SM} \ge \mathcal {B}(M\to M ^{(*)}_f\chi\chi)_{Q_i}=\frac{|g_{mi}|^2\widetilde\Gamma_{ii}}{\Gamma_{M_B}}, \label{eq7} \end{aligned}\] where \(i=1-6\), \(\widetilde\Gamma_{ii}\) are independent of the coupling constants, and \(\Gamma_{M_B}\) is the total width of the initial \(B\) meson. Notably, the partial decay width should be divided by two since the Majorana fermion is identical to its antiparticle. The upper limits of \(|g_{mi}|^2\) on (\(b s\chi\chi\)) and (\(bd\chi\chi\)) vertices are shown as functions of \(m_\chi\) in Fig. [\[gm-12\]](#gm-12){reference-type="ref" reference="gm-12"} with \(m_\chi\) is the mass of \(\chi\). One can see that when \(m_\chi\to 0\), the upper limits of \(|g_{mi}|^2\) are \(\mathcal O(10^{-17})\) to \(\mathcal O(10^{-16})\). Note that the limits of \(|g_{m2,4,6}|^2\) are about an order of magnitude larger than these of \(|g_{m1,3,5}|^2\), because the experimental upper bounds on the meson decay processes of \(0^-\to1^-\) are larger than those of \(0^-\to0^-\) given in Table [5](#exp){reference-type="ref" reference="exp"}. When \(m_\chi\) is larger, the bounds are getting looser as the phase space decreases.
## Results with invisible particles
For the baryonic decays of \({\bf B}_b\to {\bf B}_n\chi\chi\), all operators in Eq. [\[op\]](#op){reference-type="eqref" reference="op"} should be considered. One has that \[\begin{aligned} \langle {\bf B}_n\chi\chi|\mathcal L_{eff}| {\bf B}_b\rangle &=2g_{m1} \langle {\bf B}_n|(\bar q_{_f} q) |{\bf B}_b\rangle\bar u_{\chi}v_{\chi}+2g_{m2} \langle {\bf B}_n|(\bar q_{_f}\gamma^5 q) |{\bf B}_b\rangle\bar u_{\chi}v_{\chi}\\ &+2 g_{m3} \langle {\bf B}_n|(\bar q_{_f} q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^5 v_{\chi} +2g_{m4} \langle {\bf B}_n|(\bar q_{_f}\gamma^5 q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^5 v_{\chi}\\ &+2g_{m5} \langle {\bf B}_n|(\bar q_{_f}\gamma_\mu q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi} +2g_{m6} \langle {\bf B}_n|(\bar q_{_f}\gamma_\mu \gamma^5q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}. \\ \end{aligned}\] Here, the baryonic transition matrix elements have been given by Eq. [\[ffb\]](#ffb){reference-type="eqref" reference="ffb"}, while the numerical values of the FFs have been shown in Fig. [\[ff\]](#ff){reference-type="ref" reference="ff"}. By integrating the three-body phase space in Eq. [\[ps3\]](#ps3){reference-type="eqref" reference="ps3"}, \(\widetilde\Gamma_{ii}\) are obtained with the numerical results in Fig. [\[width-16\]](#width-16){reference-type="ref" reference="width-16"}. One can see that \(\widetilde\Gamma_{11,22,33,44,66}\) decrease to zero as \(m_\chi\) increases due to the phase space reduction, while \(\widetilde\Gamma_{55}\) increases first and then decreases to zero. The upper bound of \(m_\chi\) can be taken as \((M-M_f)/2\). When \(m_\chi=0\), we have that \(\widetilde\Gamma_{11}=\widetilde\Gamma_{33}\) and \(\widetilde\Gamma_{22}=\widetilde\Gamma_{44}\), since \(\widetilde\Gamma_{11,22}\) and \(\widetilde\Gamma_{33,44}\) are proportional to \((P_1\cdot P_2-m_\chi^2)\) and \((P_1\cdot P_2+m_\chi^2)\), respectively. It should be noted that \(\widetilde\Gamma_{ii}\) are independent of the coupling constants. By combining \(\widetilde\Gamma_{ii}\) with the bounds of the coupling coefficients given in Fig. [\[gm-12\]](#gm-12){reference-type="ref" reference="gm-12"}, we obtain the upper limits of the decay branching ratios as shown in Fig. [\[br-16\]](#br-16){reference-type="ref" reference="br-16"}. We see that in the most regions of \(m_\chi\), the upper limits of the branching ratios are \(\mathcal O(10^{-6})\) to \(\mathcal O(10^{-5})\), which are about of the same orders or an order of magnitude larger than the SM expectations of \(10^{-8}\) to \(10^{-6}\). In particular, \(Q_{2,4,6}\) make the dominant contributions. This is because the bounds on \(|g_{m2,4,6}|^2\) are looser than these on \(|g_{m1,3,5}|^2\). When \(m_\chi\to(M-M_f)/2\), the upper limits for the branching ratios from \(Q_{2,4,6}\) approach infinity, because the mass difference between the initial and final mesons is smaller than that between the initial and final baryons. For a larger value of \(m_\chi\), the baryon decays cannot be limited by the meson decay channels. In Tables [6](#br0){reference-type="ref" reference="br0"}, [7](#br1){reference-type="ref" reference="br1"} and [8](#br2){reference-type="ref" reference="br2"}, we list the upper limits of \(\mathcal B({\bf B}_b\to{\bf B}_n\chi\chi)\) for \(m_\chi=0,~1\) and \(2\) GeV, respectively. In Table [6](#br0){reference-type="ref" reference="br0"}, we also show the SM predictions of \(\mathcal B({\bf B} _b\to{\bf B}_n\bar\nu\nu)\).
We find that for the decays with \(b\to s\) transition, the contributions from the new operators are almost of the same orders as the SM ones. While for these with \(b\to d\) transition, the upper bounds of the decay modes with the invisible particles are about one to two orders of magnitude larger than \(\mathcal {B}({\bf B}_b\to{\bf B}_n\bar\nu\nu)\) due to the CKM matrix element depressions. Clearly, it is more hopeful to distinguish new neutral particles from the SM neutrinos experimentally. When \(m_\chi\) is larger, the upper limits of the contributions from \(Q_{2,4,6}\) are getting looser. We expect that in the near future, experiments on the bottomed baryon FCNC decays could give more relevant results for comparisons.
# Conclusion
We have studied the light invisible Majorana fermions in the FCNC processes of the long-lived bottomed baryons. The model-independent effective Lagrangian which contains six operators has been introduced to describe the couplings between the quarks and invisible Majorana fermions. The bounds of the coupling constants have been extracted from the differences between the experimental upper limits and SM predictions of the relevant \(B\) meson FCNC decays. Based on these bounds, we have predicted the upper limits of \(\mathcal B({\bf B}_b\to{\bf B}_n\chi\chi)\). In particular, we have found that the decay branching ratios of \(\Lambda_b\to\Lambda\chi\chi\), \(\Xi_b^{0(-)}\to\Xi^{0(-)}\chi\chi\), \(\Lambda_b\to n \chi\chi\), \(\Xi_b^{-}\to \Sigma^{-} \chi\chi\), \(\Xi_b^{0}\to \Lambda \chi\chi\) and \(\Xi_b^{0}\to \Sigma^{0} \chi\chi\) can be as large as \((3.3,~4.5,~1.8,~2.0,~0.27,~0.92)\times10^{-5}\), \((3.8,~5.3,~2.1,~2.3,~0.32,~1.1)\times10^{-5}\), and \((10,~15,~5.8,~5.8,~1.0,~2.7)\times10^{-5}\) with \(m_\chi=0\), \(1\), and \(2\) GeV, respectively. We are looking forward to the future experiments, such as those at LHCb, to get more measurements on bottomed baryons to find signs of new particles.
# Introduction
It is known that flavor-changing neutral current (FCNC) processes of long-lived particles would provide a window to observe new physics (NP) beyond the standard model (SM). These particles, such as ground-state mesons of \(K,~B,~D\) and \(B_c\), and baryons of \(\Lambda_{c,b}\) and \(\Xi_{c,b}\), decay through weak interactions, resulting in longer lifetimes and narrower decay widths. These FCNC decays may benefit the detections of NP. Hadronic FCNC decays include \(c\to u\), \(s\to d\), \(b\to d\), and \(b\to s\) processes at quark level. In the SM, dilepton FCNC modes have been widely studied theoretically and experimentally . However, the neutrino (\(\nu\)) and anti-neutrino (\(\bar \nu\)) in the final states of the decays cannot be directly detected, but are treated as missing energy (\(\slashed E\)) in experiments. So far most experiments can only obtain the upper limits on the decay branching ratios associated with \(\nu\bar\nu\) . The experimental searches have given the strictest constraints on kaon FCNC decays. Recently, the upper bound on \(K_L\to\pi^0\bar\nu\nu\) from the KOTO experiment at J-PARC has been given to be \(\mathcal {B}(K_L\to\pi^0\bar\nu\nu)_{\rm KOTO}<3.0\times 10^{-9}\) at \(90\%\) confidence level (C.L.), which is slightly greater than the SM prediction of \(\mathcal {B}(K_L\to\pi^0\bar\nu\nu)_{\rm SM} = (3.4 \pm 0.6)\times10^{-11}\) . On the other hand, the decay of \(K^+\to \pi^+\bar\nu\nu\) has been measured, namely, \(\mathcal {B}(K^+\to\pi^+\bar\nu\nu)_{\rm NA62}=(11.0^{+4.0}_{-3.5}({\rm stat})\pm 0.3(\rm syst))\times 10^{-11}\) at \(68\%\) C.L. from the NA62 experiment at CERN and \(\mathcal {B}(K^+\to\pi^+\bar\nu\nu)_{\rm E949}= (17.3^{+11.5}_{-10.5})\times 10^{-11}\) from the E949 experiment at BNL . These results are consistent with the SM prediction of \(\mathcal {B}(K^+\to\pi^+\bar\nu\nu)_{\rm SM}= (8.4\pm1.0)\times 10^{-11}\) within one standard deviation. It is clear that the room for NP in \(K \to \pi\slashed E\) has become quite small. However, the searches for NP in the FCNC decay processes of charmed and bottomed mesons would still be possible. For example, the charmed meson and hyperon decays associated with \(\slashed E\) have been analyzed in Ref. . The bottomed ones have been measured in Refs. , which are listed in Table [5](#exp){reference-type="ref" reference="exp"} along with the SM predictions .
As seen from Table [5](#exp){reference-type="ref" reference="exp"}, the differences between the first and second columns indicate that there are some rooms for new invisible particles of \(\chi\) (shown as the third column) emitted in such processes. In Refs. , the effects of the invisible particles with various spins in the \(b\to s(d)\) transitions have been explored. While most of these previous studies in the literature are focused on bottomed mesons. There have been no related researches on bottomed baryons up to now. In this paper, we generalize the experimental upper bounds from \(B\) mesons to the corresponding decay modes of bottomed baryons, namely, \(\Lambda_b\) and \(\Xi_b\). On experiments, the numbers of \(\Lambda_b\) and \(\Xi_b\) produced by LHC between 2011 and 2018 are of the order of \(10^{11}\) and \(10^{9}\) , respectively. In addition, LHC is expected to produce \(10\) times more \(b\) quarks in \(run~3\) (2022-2026) . Clearly, in the near future, experiments on bottomed baryons can give more results related to the invisible particles. In this work, we consider the bottomed baryonic FCNC decays of \({\bf B}_b\to\mathcal {\bf B}_n\chi\chi\), where \({\bf B}_{n(b)}\) are (bottomed) baryons and \(\chi\) represent light invisible particles, which are assumed to be Majorana fermions. Phenomenologically, these new invisible fermions of \(\chi\) can weakly interact with the SM fermions via a mediator, which can be a scalar , pseudoscalar , vector or axial-vector particle. In our study, we will concentrate on a general model-independent approach to introduce the effective Lagrangian, which contains all possible currents involving the invisible fermions with the coupling constants extracted from the experiments. The paper is organized as follows: In Sec. II, we obtain the SM expectations of \({\bf B}_b\to\mathcal {\bf B}_n\bar\nu\nu\). In Sec. III, we first construct the effective Lagrangian, which describes the coupling between the quarks and light invisible fermions. We then present the numerical results of the upper limits for the decay branching ratios of \({\bf B}_b\to\mathcal {\bf B}_n\chi\chi\). The hadronic transition matrix elements are evaluated based on the QCD light-cone sum rules (LCSR) and modified bag model (MBM). Finally, we give the conclusion in Sec. IV.
# The SM expectations
The FCNC decay processes of bottomed baryons with missing energy are described in Fig. [\[Feyn\]](#Feyn){reference-type="ref" reference="Feyn"}, where \({\bf B}_b\) and \({\bf B}_n\) represent the initial and final baryons, \(q=b\) and \(q_{f}=s(d)\) are initial and final quarks, and \(q_{2(3)}\) are the spectator quarks, respectively. In the SM, there is no tree-level contribution to the FCNC decays of \({\bf B}_b \to {\bf B}_n\bar\nu\nu\). The first-order contributions to these processes come from the penguin and box diagrams as shown in Fig. [\[Feyn\]](#Feyn){reference-type="ref" reference="Feyn"}a, which can be described by the effective Lagrangian, given by \[\begin{aligned} \mathcal L_{\bar\nu\nu} = \frac{4G_F}{\sqrt{2}}\frac{\alpha}{2\pi\sin^2\theta_W}\sum_{\ell=e, \mu, \tau}\sum_{q=u, c, t}V_{bq}V_{sq}X^\ell(x_q)(\bar s_{_L}\gamma^\mu b_{_{L}})(\bar\nu_{_{\ell L}}\gamma_\mu\nu_{_{\ell L}}), \label{eq25} \end{aligned}\] with \[\begin{aligned} X^\ell(x_q) = \frac{x_q}{8}\left[\frac{x_q+2}{x_q-1}+\frac{3(x_q-2)}{(x_q-1)^2}\ln x_q\right], \label{eq26} \end{aligned}\] where \(G_F\) represents the Fermi coupling constant, \(\alpha\) corresponds to the fine structure constant, \(\theta_W\) stands for the Weinberg angle, \(V_{ij}\) are the CKM matrix elements, and \(x_q=m_q^2/M_W^2\) with \(m_q\) (\(M_W\)) being the mass of the quark (\(W\)-boson). Consequently, the transition amplitude is given by \[\begin{aligned} \langle {\bf B}_n\bar\nu\nu|\mathcal L_{\bar\nu\nu}| {\bf B}_b\rangle &=\frac{\sqrt{2}G_F\alpha}{4\pi\sin^2\theta_W}V_{bt}V_{st}X^\ell(x_t)\langle {\bf B}_n|\bar s\gamma^\mu(1-\gamma^5)b|{\bf B}_b\rangle\times \bar u_{\nu_\ell}\gamma_\mu(1-\gamma^5)v_{\nu_\ell}. \label{eq27} \end{aligned}\] The baryonic transition matrix elements can be parameterized by the form factors (FFs) of \(f_i^{V,A}~(i=1,2,3)\), \(f^S\) and \(f^P\), defined by \[\begin{aligned} \langle {\bf B}_n(P_f,s_f)|(\bar q_{f}\gamma_\mu q) |{\bf B}_b(P,s)\rangle &=\bar u_{ _{{\bf B}_n}} (P_f,s_f)\left[\gamma_\mu f^V_1(q^2) + i\sigma_{\mu\nu} \frac{q^\nu}{M} f^V_2(q^2)+\frac{q^\mu}{M}f^V_3(q^2)\right]u_{_{{\bf B}_b}}(P,s),\\ \langle {\bf B}_n(P_f,s_f)|(\bar q_{f} q) |{\bf B}_b(P,s)\rangle &=\bar u_{_{{\bf B}_n}} (P_f,s_f)f^S(q^2)u_{_{{\bf B}_b}}(P,s),\\ \langle {\bf B}_n(P_f,s_f)|(\bar q_{f}\gamma_\mu \gamma^5 q) |{\bf B}_b(P,s)\rangle &=\bar u_{_{{\bf B}_n}} (P_f,s_f)\left[\gamma_\mu f^A_1(q^2) + i\sigma_{\mu\nu} \frac{q^\nu}{M} f^A_2(q^2)+\frac{q^\mu}{M}f^A_3(q^2)\right]\gamma^5u_{_{{\bf B}_b}}(P,s),\\ \langle {\bf B}_n(P_f,s_f)|(\bar q_{f} \gamma^5 q) |{\bf B}_b(P,s)\rangle &=\bar u_{_{{\bf B}_n}} (P_f,s_f)f^P(q^2)\gamma^5 u_{_{{\bf B}_b}}(P,s), \label{ffb} \end{aligned}\] where \(q\) corresponds to the momentum transfer, and \(M\) is the mass of the initial baryon. We will evaluate these elements in terms of the MBM, which works well for the heavy baryonic decays . In the MBM, the baryon wave functions at rest are read as \[\label{wave_function_inrest} \Psi(x_{q_1},x_{q_2},x_{q_3}) ={\cal N}\int d^3 \vec{x} \prod_{i=1,2,3} \phi_{q_i}(\vec{x}_{q_i}-\vec{x}) e^{-iE_{q_i}t_{q_i}} \,,\] where \(q_i\) are the quark components of the baryons, \({\cal N}\) the overall normalization constant, \(x_{q_i}\) (\(E_{q_i}\)) the spacetime coordinates (energies) of \(q_i\), and \(\phi_{q_i}(x)\) the quark wave functions inside a static bag, located at the center, given by \[\label{quark_wave_function} \phi_{q}(\vec{x}) = \left( \begin{array}{c} \omega_{q+} j_0(p_qr) \chi_q\\ i\omega_{q-} j_1(p_qr) \hat{r} \cdot \vec{\sigma} \chi_q\\ \end{array} \right)\.\] Here, \(j_{0,1}\) represent the spherical Bessel functions, \(\omega_{q\pm} = \sqrt{T_{q} \pm M_{q} }\) with \(T_{q}\) the kinematic energies, and \(\chi_q\) are the two component spinors. By demanding that quark currents shall not penetrate the boundary of bags, we have the boundary condition \[\label{momentum_condition} \tan (p_qR) = \frac{p_qR}{1-M_qR-E_qR}\,,\] where \(R\) is the bag radius, resulting in that the magnitudes of 3-momenta are quantized, which can be analogous to the well-know infinite square well. By sandwiching the operators, we arrive \[\begin{aligned}
\label{master} &&\int \langle \Lambda | \overline{ s }\Gamma b(x) e^{iqx} |\Lambda_b \rangle d^4 x= {\cal Z}\int d^3\vec{x}_\Delta \Gamma _{sb}(\vec{x}_\Delta) \prod_{q_j=u,d } D_{q_j}(\vec{x}_\Delta)\,, \end{aligned}\] with \[\begin{aligned}
\label{con} &&{\cal Z} \equiv (2\pi )^4 \delta^4(p_{\Lambda_b}-p_{\Lambda}-q) {\cal N}_{\Lambda_b} {\cal N}_{\Lambda} \,,\nonumber\\ &&D_{q_j}(\vec{x}_{\Delta}) \equiv \sqrt{1-v^2 } \int d^3 \vec{x} \phi_{q_j}^\dagger \left(\vec{x} +\frac{1}{2}\vec{x}_{\Delta}\right) \phi_{q_j} \left(\vec{x}-\frac{1}{2}\vec{x}_{\Delta}\right) e^{-2iE_{q_j} \vec{v}\cdot\vec{x}}\,,\nonumber\\ &&\Gamma _{sb} (\vec{ x}_\Delta)=\int d^3\vec{x} \phi _s\left(\vec{x} + \frac{1}{2}\vec{x}_\Delta \right)\gamma^0 S_{-\vec{ v}}\Gamma S_{-\vec{ v}} \phi_b\left(\vec{x}-\frac{1}{2}\vec{x}_\Delta \right) e^{i(M_\Lambda + M_{\Lambda_b}-E_s-E_b)\vec{ v}\cdot\vec{ x} }\,, \end{aligned}\] where \(\Gamma\) are arbitrary Dirac matrices, and \(S_{\vec{v}}\) the Lorentz boost matrix of Dirac spinors. We have taken the initial (final) state as \(\Lambda_b\) (\(\Lambda\)) for an concrete example. To simplify the algebra, the Briet frame is chosen, where \(\Lambda_b\) and \(\Lambda\) have the velocity \(-\vec{ v}\) and \(\vec{ v}\), respectively. Notably, all the parameters of the model are extracted from the mass spectra, given as \[R = 4.8 ~\text{GeV}^{-1}\,,~~~M_{u,d} =0\,,~~~M_s = 0.28~\text{GeV}\,,~~~M_b = 5.093~\text{GeV}\.\] We consider the bottomed baryon decays of \(\Lambda_b\to\Lambda\bar\nu\nu\) and \(\Xi_b^{0(-)}\to\Xi^{0(-)}\bar\nu\nu\), \(\Lambda_b\to n \bar\nu\nu\), \(\Xi_b^{0(-)}\to \Sigma^{0(-)} \bar\nu\nu\) and \(\Xi_b^{0}\to \Lambda \bar\nu\nu\), due to the \(b\to s\) and \(b\to d\) transitions at quark level, respectively. The FFs can be extracted straightforwardly after the computations, which are shown in Fig. [\[ff\]](#ff){reference-type="ref" reference="ff"} along with their \(q^2\) dependencies. By integrating the three-body phase space, we obtain the decay branching ratio to be \[\mathcal {B} ({\bf B}_b\to{\bf B}_n\bar\nu\nu) = \frac{1}{512 \pi^3 M^3 \Gamma_{{\bf B}_b}}\int\frac {dq^2}{q^2} \lambda^{1/2}(M^2, q^2, M_f^2)\lambda^{1/2}(q^2, m_1^2, m_2^2)\int d\cos\theta\sum|\mathcal M|^2, \label{ps3}\] where \(\lambda(x, y, z)= x^2 + y^2 +z^2-2xy-2xz-2yz\) is the K\({\rm \ddot a}\)llen function, \(M\), \(M_f\), \(m_1\) and \(m_2\) correspond to the masses of the initial baryon, final baryon, neutrino and anti-neutrino, respectively, \(\theta\) is the phase space angle, \(\Gamma_{{\bf B}_b}\) represents the total width of the initial baryon, and \(\mathcal M\) stands for the amplitude. As the three generations of neutrinos are indistinguishable experimentally, the final results need to be multiplied by three. For the \(b\to s\) transition, the decay branching ratios associated with neutrino and anti-neutrino are as follows: \[\begin{aligned} \mathcal{B}(\Lambda_b\to \Lambda\bar\nu\nu)&=5.52\times 10^{-6}, \\ \mathcal{B}(\Xi_b^{0(-)}\to \Xi^{0(-)}\bar\nu\nu)&=7.80\times 10^{-6}. \label{SM1} \end{aligned}\] Here, due to the \(\rm SU(3)\) flavor symmetry, the branching ratios of \(\Xi_b^0\) and \(\Xi_b^-\) are considered approximately to be equal. Note that our results of (\(\Lambda_b\to \Lambda\bar\nu\nu\)) in Eq. [\[SM1\]](#SM1){reference-type="eqref" reference="SM1"} is smaller than the previous prediction in Ref. . Similarly, for the \(b\to d\) transition we have that \[\begin{aligned} \mathcal{B}(\Lambda_b\to n\bar\nu\nu)&=2.76\times 10^{-7}, \\ \mathcal{B}(\Xi_b^-\to \Sigma^-\bar\nu\nu)&=2.65\times 10^{-7}, \\ \mathcal{B}(\Xi_b^0\to \Lambda\bar\nu\nu)&=3.88\times 10^{-8}, \\ \mathcal{B}(\Xi_b^0\to \Sigma^0\bar\nu\nu)&=1.24\times 10^{-7}, \label{SM2} \end{aligned}\] which are about one to two orders of magnitude smaller than the modes in Eq. [\[SM1\]](#SM1){reference-type="eqref" reference="SM1"}, due to the ratio of the CKM matrix elements, \(|V_{td}/V_{ts}|\sim\mathcal O(\lambda)\). As a result, the SM predictions of bottomed baryonic FCNC processes with \(\slashed E\) are \(\mathcal O(10^{-8})-\mathcal O (10^{-6})\). If the experimental detections of these decays are larger than the values in Eqs. [\[SM1\]](#SM1){reference-type="eqref" reference="SM1"}-[\[SM2\]](#SM2){reference-type="eqref" reference="SM2"}, the new invisible neutral particles from NP are expected.
# Processes with invisible particles
## Effective Lagrangian
In Fig. [\[Feyn\]](#Feyn){reference-type="ref" reference="Feyn"}b, two spin-\(1/2\) invisible Majorana particles of \(\chi\chi\) are assumed to be emitted in the process, in which the four-fermion vertex may be generated at tree or loop level by introducing new physical mediators in specific models . Under the low energy scale, the model-independent effective Lagrangian is given by \[\begin{aligned} \mathcal L_{eff}=\sum_{i=1}^{6}{g_{_{mi}} Q_i}, \label{eq1} \end{aligned}\] where \(g_{fi}\) are the phenomenological coupling constants, which are taken at the new physical energy scale \(\Lambda\). There are \(6\) independent dimension-six effective operators, which have the forms: \[\begin{aligned} & Q_1=(\bar q_{_f}q)(\chi\chi),~~~~~~~~~Q_2=(\bar q_{_f} \gamma^{5}q)(\chi\chi),~~~~~~~~~ Q_3=(\bar q_{_f}q)(\chi\gamma^{5}\chi),\\ & Q_4=(\bar q_{_f} \gamma^{5}q)(\chi\gamma^{5}\chi),~~~Q_5=(\bar q_{_f} \gamma_{\mu}q)(\chi\gamma^{\mu}\gamma^{5}\chi),~~~ Q_6=(\bar q_{_f} \gamma_{\mu}\gamma^{5} q)(\chi\gamma^{\mu}\gamma^{5}\chi), \label{op} \end{aligned}\] where the invisible particles of \(\chi\) have been assumed to be the Majorana type. Since \(\chi\gamma^{\mu}\chi=0\) and \(\chi\sigma^{\mu\nu}\chi=0\), there is no contribution from the vector or tensor current. The upper limits of the coupling constants in the effective Lagrangian can be extracted from the difference between the theoretical predictions and experimental data of the \(B\) meson FCNC decays, such as \(B ^ {-} \to K ^-(K^{\ast-}) + \slashed E\) and \(B ^ {-} \to \pi ^-(\rho^-) + \slashed E\). For the \(0^-\to(0^-,1^-)\) mesons decays of \(M^-\to(M_f^-,M_f^{*-})\chi\chi\), only operators \(Q_{1,3,5}\) and \(Q_{2,4,5,6}\) in Eq. ([\[op\]](#op){reference-type="ref" reference="op"}) give the contributions, respectively. The amplitudes of the \(0^-\to (0^-, 1^-)\) decays can be simplified as \[\begin{aligned} \langle M_f^-\chi\chi|\mathcal L_{eff}|M^-\rangle&=2g_{m1} \langle M_f^-|(\bar q_{_f} q) |M^-\rangle\bar u_{\chi}v_{\chi}+2 g_{m3} \langle M_f^-|(\bar q_{_f} q) |M^-\rangle\bar u_{\chi}\gamma^5 v_{\chi}\\ &+2g_{m5}\langle M_f^-|(\bar q_{_f}\gamma_\mu q) |M^-\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}, \\ \label{Lmm0} \end{aligned}\] and \[\begin{aligned} \langle M_f^{*-}\chi\chi|\mathcal L_{eff}|M^-\rangle&=2g_{m2} \langle M_f^{*-}|(\bar q_{_f}\gamma^5 q) |M^-\rangle\bar u_{\chi}v_{\chi}+2g_{m4} \langle M_f^{*-}|(\bar q_{_f}\gamma^5 q) |M^-\rangle\bar u_{\chi}\gamma^5 v_{\chi}\\ &+2g_{m5} \langle M_f^{*-}|(\bar q_{_f}\gamma_\mu q) |M^-\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}+2g_{m6} \langle M_f^{*-}|(\bar q_{_f}\gamma_\mu \gamma^5q) |M^-\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}, \\ \end{aligned}\] respectively. The hadronic transition matrix elements can be expressed as \[\begin{aligned} \langle M_f^-|(\bar q_{_f} q) |M^-\rangle &= \frac{M^2-M_{f}^2}{m_q-m_{q_{_f}}}f_0 (q^2),\\ \langle M_f^-|(\bar q_{_f}\gamma_\mu q) |M^-\rangle &= (P+P_f)_{\mu}f_+ (q^2)+(P-P_f)_{\mu}\frac{M^2-M_{f}^2}{q^2} \big[f_0 (q^2)-f_+ (q^2)\big],\\ \langle M_f^-|(\bar q_{_f}\sigma_{\mu\nu} q) |M^-\rangle &=i\big[P_{\mu} (P-P_f)_{\nu}-P_{\nu} (P-P_f)_{\mu}\big] \frac{2}{M+M_f} f_T(q^2), \label{} \end{aligned}\] and \[\begin{aligned} \langle M_f^{*-}|(\bar q_{_f}\gamma^5 q) |M^-\rangle &=-i\big[\epsilon \cdot (P-P_f)\big]\frac{2M_f}{m_q+m_{q_{_f}}}A_0(q^2),\\ \langle M_f^{*-}|(\bar q_{_f}\gamma_\mu\gamma^5 q) |M^-\rangle &= i\bigg\{\epsilon_{\mu} (M+M_f)A_1(q^2)-(P+P_f)_{\mu}\frac{\epsilon \cdot (P-P_f)}{M+M_f}A_2(q^2)\\ &~~~-(P-P_f)_{\mu} \big[\epsilon \cdot (P-P_f)\big] \frac{2M_f}{q^2} \big[A_3(q^2)-A_0(q^2)\big]\bigg\},\\ \langle M_f^{*-}|(\bar q_f \gamma_\mu q )|M^-\rangle &=\varepsilon _{\mu \nu \rho \sigma} \epsilon ^\nu P^\rho (P-P_f)^\sigma\frac{2 }{M+M_f}V(q^2), \label{} \end{aligned}\] where \(m_{q_f}\) are the quark masses, \(f_j~(j=0,+,T)\), \(A_k~(k=0-3)\) and \(V\) are the FFs, which are evaluated from the method of the LCSR , and \(\epsilon\) is the polarization vector of the final meson with the convention of \(\varepsilon^{0123}=1\). In our calculation, we assume that only one operator contributes to the process at a time. By integrating the three-body phase space given in Eq. [\[ps3\]](#ps3){reference-type="eqref" reference="ps3"}, the upper limits of the coupling constants \(g_{mi}\) can be obtained from Table [5](#exp){reference-type="ref" reference="exp"}, given by \[\begin{aligned} \mathcal B(M\to M ^{(*)}_f\slashed E)_{\rm exp}-\mathcal B(M\to M ^{(*)}_f \bar\nu\nu)_{\rm SM} \ge \mathcal {B}(M\to M ^{(*)}_f\chi\chi)_{Q_i}=\frac{|g_{mi}|^2\widetilde\Gamma_{ii}}{\Gamma_{M_B}}, \label{eq7} \end{aligned}\] where \(i=1-6\), \(\widetilde\Gamma_{ii}\) are independent of the coupling constants, and \(\Gamma_{M_B}\) is the total width of the initial \(B\) meson. Notably, the partial decay width should be divided by two since the Majorana fermion is identical to its antiparticle. The upper limits of \(|g_{mi}|^2\) on (\(b s\chi\chi\)) and (\(bd\chi\chi\)) vertices are shown as functions of \(m_\chi\) in Fig. [\[gm-12\]](#gm-12){reference-type="ref" reference="gm-12"} with \(m_\chi\) is the mass of \(\chi\). One can see that when \(m_\chi\to 0\), the upper limits of \(|g_{mi}|^2\) are \(\mathcal O(10^{-17})\) to \(\mathcal O(10^{-16})\). Note that the limits of \(|g_{m2,4,6}|^2\) are about an order of magnitude larger than these of \(|g_{m1,3,5}|^2\), because the experimental upper bounds on the meson decay processes of \(0^-\to1^-\) are larger than those of \(0^-\to0^-\) given in Table [5](#exp){reference-type="ref" reference="exp"}. When \(m_\chi\) is larger, the bounds are getting looser as the phase space decreases.
## Results with invisible particles
For the baryonic decays of \({\bf B}_b\to {\bf B}_n\chi\chi\), all operators in Eq. [\[op\]](#op){reference-type="eqref" reference="op"} should be considered. One has that \[\begin{aligned} \langle {\bf B}_n\chi\chi|\mathcal L_{eff}| {\bf B}_b\rangle &=2g_{m1} \langle {\bf B}_n|(\bar q_{_f} q) |{\bf B}_b\rangle\bar u_{\chi}v_{\chi}+2g_{m2} \langle {\bf B}_n|(\bar q_{_f}\gamma^5 q) |{\bf B}_b\rangle\bar u_{\chi}v_{\chi}\\ &+2 g_{m3} \langle {\bf B}_n|(\bar q_{_f} q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^5 v_{\chi} +2g_{m4} \langle {\bf B}_n|(\bar q_{_f}\gamma^5 q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^5 v_{\chi}\\ &+2g_{m5} \langle {\bf B}_n|(\bar q_{_f}\gamma_\mu q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi} +2g_{m6} \langle {\bf B}_n|(\bar q_{_f}\gamma_\mu \gamma^5q) |{\bf B}_b\rangle\bar u_{\chi}\gamma^{\mu}\gamma^5 v_{\chi}. \\ \end{aligned}\] Here, the baryonic transition matrix elements have been given by Eq. [\[ffb\]](#ffb){reference-type="eqref" reference="ffb"}, while the numerical values of the FFs have been shown in Fig. [\[ff\]](#ff){reference-type="ref" reference="ff"}. By integrating the three-body phase space in Eq. [\[ps3\]](#ps3){reference-type="eqref" reference="ps3"}, \(\widetilde\Gamma_{ii}\) are obtained with the numerical results in Fig. [\[width-16\]](#width-16){reference-type="ref" reference="width-16"}. One can see that \(\widetilde\Gamma_{11,22,33,44,66}\) decrease to zero as \(m_\chi\) increases due to the phase space reduction, while \(\widetilde\Gamma_{55}\) increases first and then decreases to zero. The upper bound of \(m_\chi\) can be taken as \((M-M_f)/2\). When \(m_\chi=0\), we have that \(\widetilde\Gamma_{11}=\widetilde\Gamma_{33}\) and \(\widetilde\Gamma_{22}=\widetilde\Gamma_{44}\), since \(\widetilde\Gamma_{11,22}\) and \(\widetilde\Gamma_{33,44}\) are proportional to \((P_1\cdot P_2-m_\chi^2)\) and \((P_1\cdot P_2+m_\chi^2)\), respectively. It should be noted that \(\widetilde\Gamma_{ii}\) are independent of the coupling constants. By combining \(\widetilde\Gamma_{ii}\) with the bounds of the coupling coefficients given in Fig. [\[gm-12\]](#gm-12){reference-type="ref" reference="gm-12"}, we obtain the upper limits of the decay branching ratios as shown in Fig. [\[br-16\]](#br-16){reference-type="ref" reference="br-16"}. We see that in the most regions of \(m_\chi\), the upper limits of the branching ratios are \(\mathcal O(10^{-6})\) to \(\mathcal O(10^{-5})\), which are about of the same orders or an order of magnitude larger than the SM expectations of \(10^{-8}\) to \(10^{-6}\). In particular, \(Q_{2,4,6}\) make the dominant contributions. This is because the bounds on \(|g_{m2,4,6}|^2\) are looser than these on \(|g_{m1,3,5}|^2\). When \(m_\chi\to(M-M_f)/2\), the upper limits for the branching ratios from \(Q_{2,4,6}\) approach infinity, because the mass difference between the initial and final mesons is smaller than that between the initial and final baryons. For a larger value of \(m_\chi\), the baryon decays cannot be limited by the meson decay channels. In Tables [6](#br0){reference-type="ref" reference="br0"}, [7](#br1){reference-type="ref" reference="br1"} and [8](#br2){reference-type="ref" reference="br2"}, we list the upper limits of \(\mathcal B({\bf B}_b\to{\bf B}_n\chi\chi)\) for \(m_\chi=0,~1\) and \(2\) GeV, respectively. In Table [6](#br0){reference-type="ref" reference="br0"}, we also show the SM predictions of \(\mathcal B({\bf B} _b\to{\bf B}_n\bar\nu\nu)\).
We find that for the decays with \(b\to s\) transition, the contributions from the new operators are almost of the same orders as the SM ones. While for these with \(b\to d\) transition, the upper bounds of the decay modes with the invisible particles are about one to two orders of magnitude larger than \(\mathcal {B}({\bf B}_b\to{\bf B}_n\bar\nu\nu)\) due to the CKM matrix element depressions. Clearly, it is more hopeful to distinguish new neutral particles from the SM neutrinos experimentally. When \(m_\chi\) is larger, the upper limits of the contributions from \(Q_{2,4,6}\) are getting looser. We expect that in the near future, experiments on the bottomed baryon FCNC decays could give more relevant results for comparisons.
# Conclusion
We have studied the light invisible Majorana fermions in the FCNC processes of the long-lived bottomed baryons. The model-independent effective Lagrangian which contains six operators has been introduced to describe the couplings between the quarks and invisible Majorana fermions. The bounds of the coupling constants have been extracted from the differences between the experimental upper limits and SM predictions of the relevant \(B\) meson FCNC decays. Based on these bounds, we have predicted the upper limits of \(\mathcal B({\bf B}_b\to{\bf B}_n\chi\chi)\). In particular, we have found that the decay branching ratios of \(\Lambda_b\to\Lambda\chi\chi\), \(\Xi_b^{0(-)}\to\Xi^{0(-)}\chi\chi\), \(\Lambda_b\to n \chi\chi\), \(\Xi_b^{-}\to \Sigma^{-} \chi\chi\), \(\Xi_b^{0}\to \Lambda \chi\chi\) and \(\Xi_b^{0}\to \Sigma^{0} \chi\chi\) can be as large as \((3.3,~4.5,~1.8,~2.0,~0.27,~0.92)\times10^{-5}\), \((3.8,~5.3,~2.1,~2.3,~0.32,~1.1)\times10^{-5}\), and \((10,~15,~5.8,~5.8,~1.0,~2.7)\times10^{-5}\) with \(m_\chi=0\), \(1\), and \(2\) GeV, respectively. We are looking forward to the future experiments, such as those at LHCb, to get more measurements on bottomed baryons to find signs of new particles. | {'timestamp': '2022-06-06T02:14:42', 'yymm': '2206', 'arxiv_id': '2206.01575', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01575'} |
# Introduction {#sec:intro}
Highly energetic partons (quarks and gluons) are expected to lose energy while traversing the extremely hot and dense quark gluon plasma (QGP) created in ultra-relativistic heavy ion collisions . This phenomenon, known as "jet quenching", was discovered at the Relativistic Heavy Ion Collider (RHIC) through the observation of suppression of high transverse momentum (\(p_T\)) hadrons and back-to-back dihadron correlations . These observations were confirmed at the Large Hadron Collider (LHC) and extended into a larger kinematic range . In addition to measurements on high \(p_T\) particles, which are the leading fragments of jets, the LHC opened new opportunities for direct observations of jet quenching and jet modifications in QGP. Significant dijet transverse momentum asymmetry and suppressed jet production were observed in lead-lead (PbPb) collisions. Jet-hadron correlations gave insights into how the energy lost by the jets is distributed in the medium produced in the collisions. Measurements of the jet shapes (transverse momentum radial profile) and the jet fragmentation functions were performed to further investigate the mechanism of the jet shower modifications in QGP. To quantify the medium modifications, measurements in proton-proton (pp) collisions are used as a reference for jet propagation in vacuum. Recent studies have focused on the jet substructure. The jet splitting function has been measured in both pp and PbPb collisions , through the usage of a jet grooming algorithm that is able to split ("decluster") a single jet into two subjets and locate the hard splitting by removing softer wide-angle radiation contributions. The hard splitting, characterized by two well-separated subjets ("two-prong" structure), provides access to the early stages in the parton shower evolution. Inspired by the investigations of the jet substructure, in this study we extend the declustering procedure continuously following the leading subjet in each splitting, and capture sequential substructure variables related to successive splittings in the parton showering process. These sequential variables, which carry information about the parton-medium interactions, can be perceived by a long short-term memory (LSTM) neural network. The LSTM neural network is a special type of recurrent neural network that is capable of learning long-term dependencies. We perform supervised machine learning in which a well-trained classifier is able to distinguish quenched jets from vacuum jets on a jet-by-jet basis. A previous study of this classification problem explored several ways of jet representation and different neural network architectures. It showed some promising first results of deep learning techniques applied for the study of quenched jets in QGP at simulation level. However, the application of the convolutional neural network often involves image preprocessing techniques, such as rotation, and normalization of the pixel intensity. These techniques are usually designed for a specific jet species, for example the boosted W/Z jets as studied in ref. . How such techniques can bring benefits to studies of the QCD jets is not well known. In addition, as a common practice in the implementation of modern Monte Carlo event generators, events associated with small cross-sections which often produce high \(p_T\) jets, can be over-sampled for statistical benefits, and event weights are assigned accordingly for compensation. How the event weights fit into the development of a machine learning approach is not well studied. Another challenge that has not yet been addressed is that in the experiment the jets evolve in the presence of a thermal background of particles created in heavy ion collisions. Therefore, before the feasibility of machine learning can be established, techniques such as event mixing and background subtraction should be applied to simulated events, after which the jet finding algorithms can be performed. In this study, we present a complete machine learning approach to the identification of quenched jets in the presence of uncorrelated underlying events. In Section [2](#sec:event_simulation){reference-type="ref" reference="sec:event_simulation"}, we describe the procedure of event simulation and background subtraction. In Section [3](#sec:machine_learning){reference-type="ref" reference="sec:machine_learning"}, we describe how the supervised machine learning is performed, including a general feature engineering method on jets, a training method with event weights taken into consideration, a method of hyper-tuning and study of robustness. In Section [4](#sec:results){reference-type="ref" reference="sec:results"}, we test the trained classifier about its discrimination power of quenched jets against non-quenched ones. Finally, we present our conclusions and outlook in Section [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}. The full code and data used in this study are available as open-source .
# Data Sample {#sec:event_simulation}
In this study, hard scattering events are simulated with Monte Carlo event generators such as [Pythia 8]{.smallcaps} and [Jewel v2.2.0]{.smallcaps} at a center-of-mass energy per nucleon pair (\(\sqrt{s_{NN}}\)) of \(\SI{5.02}{TeV}\) with a parton transverse momentum cutoff \(\hat p_t = \SI{100}{GeV}\). [Pythia]{.smallcaps} is used to simulate dijet events in pp collisions, while [Jewel]{.smallcaps} simulates dijet events in heavy ion collisions with medium-induced effects on the parton showering process taken into consideration. An important aspect of the jet-medium interactions implemented in [Jewel]{.smallcaps} is the medium response. In our studies, the medium recoils are turned on to simulate underlying events correlated with the dijets. It has to be pointed out that such implementation of the medium response is a model-dependent component, while other generators may implement their own descriptions of the medium response. More details can be found from ref. . To simulate the thermal background, which is not correlated with the jets, dijet events from both event generators are embedded into underlying events, which have general features similar to those observed experimentally. Following the implementation in ref. , the particles in the uncorrelated underlying events are generated from a Boltzmann distribution in \(p_T\), and uniform distribution in pseudorapidity \(\eta\) and azimuthal angle \(\varphi\). We prepare the uncorrelated underlying events using two different multiplicity settings, corresponding roughly to the mid-central (40-50%) and the most-central (10-20%) PbPb collisions at \(\sqrt{s_{NN}}=\SI{5.02}{TeV}\), as inferred from published experiment results . More details about the uncorrelated underlying events are listed in Table [1](#tab:tab1){reference-type="ref" reference="tab:tab1"}. With the inclusion of correlated and uncorrelated underlying event backgrounds, the influence on substructure variables introduced by particles not originating from the parton shower can be investigated.
:::
## Robustness
The trained neural network can be used as a classifier, making predictions on the probability for a jet to be quenched. Here, the distributions of the raw LSTM output from the top three best-trained networks are shown in Fig. [\[fig:calib\]](#fig:calib){reference-type="ref" reference="fig:calib"}, which exhibit a non-deterministic behavior in the metric. During the training, samples of the negative class are labeled with 0, while samples of the positive class are labeled with 1, without intermediate labels in between. Hence, the metric that quantifies the probability for a jet to be quenched, which is established during the training process, suffer from the stochastic nature of machine learning. Such non-determinism is rooted in the randomness of selecting samples to form a batch. To solve this issue, a calibration method is designed. The calibration method is designed such that samples of the negative class are uniformly distributed along the newly established metric, meanwhile the relative quantitative relationships between the samples' raw LSTM outputs are kept in order. The new metric can be obtained with the help of the receiver operating curve (ROC), which plots the true positive rate (TPR) against the false positive rate (FPR) at various threshold settings. The FPR quantifies the percentage of samples in the negative class that are falsely predicted as positive. For illustration, with a decision threshold of 0, all negative samples should have predictions larger than 0, which leads to FPR=1. The new metric is bound to the FPR, meanwhile the threshold on the raw LSTM output corresponding to a specific FPR value can be determined from the ROC curve. The distribution of the raw LSTM output from trained neural networks and their values after calibration are shown in Fig. [\[fig:calib\]](#fig:calib){reference-type="ref" reference="fig:calib"} (top and middle panels, respectively). For the three well-trained classifiers our calibration method produces almost identical distributions of the calibrated LSTM values. Given the distributions of the LSTM outputs, either raw or calibrated, it can be seen that not all samples in the positive class (jets from [Jewel]{.smallcaps}) are predicted to be quenched by the trained neural network. To determine if a jet is quenched, a threshold should be used, such that jets that have predictions larger than such a threshold would be classified as quenched, while those with predictions smaller than the threshold would be considered non-quenched. In the subsequent discussion use the threshold determined from TPR=0.4. With this threshold, the positive samples from [Jewel]{.smallcaps} are divided into top 40% subset, and bottom 60% subset. It can be seen from the red point on the ROC curve that this TPR corresponds to a FPR of roughly 10%, indicating that about 10% of the samples in the negative class have predictions larger than the threshold. Further, we consider the top 40% [Jewel]{.smallcaps} jets as quenched and compare their substructure variables with the remaining 60%, as well as those simulated with [Pythia]{.smallcaps}.
# Results {#sec:results}
Samples of the positive class ([Jewel]{.smallcaps}) are divided into two subsets, consisting of top 40% and bottom 60% samples, respectively, based on the raw LSTM output. To better understand the relation between the LSTM output and the quenching effects, we plot the distributions of the substructure variables of the groomed jets in Fig. [\[fig:stack\]](#fig:stack){reference-type="ref" reference="fig:stack"}, and the primary Lund plane density for the hard splittings after grooming in Fig. [\[fig:lund_lstm\]](#fig:lund_lstm){reference-type="ref" reference="fig:lund_lstm"}, for two subsets separately. Samples of the negative class ([Pythia]{.smallcaps}) that serve as a baseline are also shown for comparison. These two subsets of samples show quenching effects at different levels. The top 40% samples of the positive class show severe modifications of the jet substructure with enhancement of wider and softer splittings. In comparison, the bottom 60% samples exhibit a quenching pattern that is similar to that of the negative class. Two neural networks that are trained at two different multiplicity scenarios behave consistently, and show a similar ability in quantifying quenching effects. This indicates that the machine learning approach presented here successfully handles the uncorrelated underlying events independent of the background density.
# Conclusion {#sec:conclusion}
Our study has shown a promising machine learning approach to identifying jet quenching in the presence of uncorrelated underlying event background. Sequential substructure variables extracted from the jet clustering history carry information about the parton showering process with or without quenching effects, and can be learned by an LSTM neural network. The well-trained neural network can then be used to select jets at different quenching levels. We also studied the non-deterministic behavior of the supervised machine learning algorithm. With the help of the ROC curve, a calibration method is designed to avoid the non-determinism. The well-trained neural network is capable of quantifying quenching effects, making it possible to divide jets simulated from a quenching model ([Jewel]{.smallcaps}) into two subsets. The two separated subsets of jets have shown differences in the jet substructure variables, related to two different parton showering patterns: a vacuum-like pattern, similar to what is implemented in [Pythia 8]{.smallcaps}, and a medium-like pattern, which produces highly quenched jets. The procedure of extracting sequential substructure variables can be used as a general feature engineering method on jets. The techniques involved in our study, such as the background subtraction and the jet grooming, have been previously applied to experimental data and their effectiveness is well-studied. Thus, our machine learning approach is very promising in the exploration of different jet classification topics with real experimental data. We thank the Advanced Computing Center for Research and Education (ACCRE) at Vanderbilt University for providing computing resources. This work was supported in part by the US Department of Energy Grant No. DE-FG05-92ER40712, and by the Alfred P. Sloan Foundation, whose contributions are gratefully acknowledged.
[^1]: The output from LSTM layer has a dimension of \[batch size, sequence length, hidden size\] (<https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html>). The hidden size is a configurable parameter, in other words a hyper-parameter.
[^2]: The input dimension of the first FC layer is equal to the LSTM hidden size. | {'timestamp': '2022-06-06T02:16:16', 'yymm': '2206', 'arxiv_id': '2206.01628', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01628'} |
# Collinear limit of the two-parton colour correlation {#app:SC}
In this appendix we consider the \(u_{(\text{dip})}^{(\text{nab})}\) term in eq. [\[eq:U_dipnab_def\]](#eq:U_dipnab_def){reference-type="eqref" reference="eq:U_dipnab_def"}, | {'timestamp': '2022-06-06T02:14:53', 'yymm': '2206', 'arxiv_id': '2206.01584', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01584'} |
null | null |
# Introduction {#sec:introduction}
The generation of turbulent inflow conditions is essential in simulating spatially-developing turbulent boundary layers (TBLs), considering its effect on the accuracy of the simulations and computational cost. It is also a challenging topic due to the need for the time-dependent turbulent inflow data to be accurately described. The generated data should satisfy the momentum and continuity equations and consequently match the turbulent statistics and spectra of the flow. Several approaches have been proposed to generate turbulent inflow conditions with different levels of success. Adding infinitesimal perturbations on the laminar mean velocity profile at the inlet section of the computational domain and allowing the transition of the boundary layer is a straightforward approach that guarantees a realistic spatially-developing TBL. However, the need for a development section that is long enough for the flow to reach the fully-turbulent state can result in a high computational cost, making this approach not applicable for most turbulence simulations, where fully-turbulent inflow conditions are required. The use of precursor (auxiliary) parallel flow (fully developed flow) simulations with periodic boundary conditions applied to the streamwise direction is another approach that can be used by extracting flow fields from a plane normal to the streamwise direction and applying the data as inflow conditions to the main simulations. Although this method can produce accurate turbulence statistics and spectra for fully developed flows, it requires a high computational cost. Additionally, the streamwise periodicity effect, caused by the recycling of the flow within a limited domain size, can lead to physically unrealistic streamwise-repetitive features in the flow fields.
Furthermore, using parallel flow data as inflow for a simulation of a spatially-developing TBL can result in a long development distance downstream of the domain inlet to produce the correct boundary layer characteristics. To address this issue, a recycling-rescaling method was introduced by, which is a modified version of the method by. Here the velocity fields in the auxiliary simulation are rescaled before being reintroduced at the inlet section. Another well-known approach for generating turbulent inflow conditions is adding random fluctuations based on known turbulence statistics. The methods that are based on this approach are usually called synthetic turbulent inflow generation methods. Several methods, such as the synthetic random Fourier method, synthetic digital filtering method, synthetic coherent eddy method, synthetic vortex method, synthetic volume-force method and numerical counterpart of the experimental tripping methods have been proposed to feature a fast generation of turbulence with various levels of precision. However, a long-distance downstream of the domain inlet is required to allow the boundary layer to recover from the unphysical random fluctuations of the generated velocity fields and produce the right flow characteristics, resulting in a high computational cost. Another approach based on proper orthogonal decomposition (POD) and Galerkin projection has been proposed to build a reduced-order flow model and generate turbulent inflow conditions by utilising the most energetic eddies. A similar approach has been applied to experimental measurements to reconstruct turbulent inflow velocity fields from hot-wire anemometry and particle image velocimetry using POD and linear stochastic estimation. This approach showed the possibility of utilising the experimental results as turbulent inflow conditions. However, the costly experimental setup makes this approach not applicable as a general method to generate turbulence.
The rapid development of deep learning algorithms and the increase in the graphic processing unit (GPU) capability, accompanied by the enormous amounts of high-fidelity data generated from experimental and numerical simulations, encourage exploring new data-driven approaches that can efficiently tackle various fluid-flow problems. Deep learning is a subset of machine learning, where deep neural networks are used for classification, prediction and feature extraction. Recently, several models have shown great potential in solving different problems in the field of turbulence, such as turbulence modelling, turbulent flow prediction, reduced-order modelling, flow control, non-intrusive sensing and turbulent flow reconstruction.
Furthermore, recent studies on the generation of turbulent inflow conditions using deep learning models (DLMs) have shown promising results. showed that convolutional neural networks (CNNs) could be used to generate turbulent inflow conditions by proposing a model based on a convolutional autoencoder (CAE) with a multilayer perceptron (MLP) to generate turbulent inflow conditions using turbulent channel flow data. proposed a generative adversarial network (GAN) and a recurrent neural network (RNN)-based model as a representative of unsupervised deep learning to generate turbulent inflow conditions at various Reynolds numbers using data of turbulent channel flow at various friction Reynolds numbers. Recently, utilised a combination of a multiscale CAE with a subpixel convolution layer ( MSC~SP~-AE ) having a physical constrains-based loss function and a long short-term memory (LSTM) model to generate turbulent inflow conditions from turbulent channel flow data.
In all the above models, the prediction of the turbulent inflow conditions is based on parallel flows, which, as mentioned before, is more suitable as inflow for fully developed TBLs. Therefore, it is necessary to develop a model that considers the spatial development of TBLs. In this context, this paper proposes a deep learning method consisting of a transformer and a multiscale-enhanced super-resolution generative adversarial network (MS-ESRGAN) to generate turbulent inflow conditions for spatially-developing TBL simulations.
The remainder of this paper is organised as follows. In Section [2](#sec:2){reference-type="ref" reference="sec:2"}, the methodology of generating the turbulent inflow data using the proposed DLM is explained. The datasets used for training and testing the model are described in Section [3](#sec:3){reference-type="ref" reference="sec:3"}. Section [4](#sec:4){reference-type="ref" reference="sec:4"} presents the results obtained from testing the proposed model. Finally, Section [5](#sec:5){reference-type="ref" reference="sec:5"} presents the conclusions.
# Methodology {#sec:2}
The proposed DLM is a combination of two architectures. The first one is the transformer and the second one is the MS-ESRGAN. The transformer is used to predict the temporal evolution of extremely coarse velocity fields obtained by selecting distributed points at various sections along the streamwise direction of a spatially-developing TBL flow obtained through direct numerical simulation (DNS), as shown in figure [\[fig:k1\]](#fig:k1){reference-type="ref" reference="fig:k1"}(a). Meanwhile, the MS-ESRGAN is used to perform a super-resolution reconstruction of the data for all the sections predicted by the transformer, leading to a final high-resolution (HR) data, i.e. velocity fields with the same resolution as the ground truth data, as shown in figure [\[fig:k1\]](#fig:k1){reference-type="ref" reference="fig:k1"}(b). In other words, the transformer is trained for the data at each section, whereas MS-ESRGAN is trained for all the sections used in the training process. Figure [\[fig:k1\]](#fig:k1){reference-type="ref" reference="fig:k1"}(c) shows the schematic representation of the proposed DLM for generating turbulent inflow conditions. As shown in the figure, the input to the DLM is represented by coarse velocity data obtained from a plane normal to the streamwise direction with time interval \[\(t_0,... ,t_n\)\], and the output is represented by predicted high-resolution velocity data at instants, \(t_{n+1}\), where \(n\) is set to 12 in this study.
## Transformer {#sec:Transformer}
LSTM is an artificial neural network that can handle sequential data and time-series modelling. LSTM is a type of RNN. It has also played an essential role in modelling the temporal evolution of turbulence in various problems. Although LSTM is designed to overcome most of the traditional RNN limitations, such as vanishing gradients and explosion of gradients, it is usually slow in terms of training due to its architecture, which requires that the time-series data are introduced to the network sequentially. This prevents parallelisation of the training process, which is why GPU is used in deep learning calculations. Furthermore, LSTM has shown a limitation in dealing with long-range dependencies.
The transformer was introduced to deal with these limitations by applying the self-attention concept to compute the representations of its input and output data without feeding the data sequentially. In this study, a transformer is used to model the temporal evolution of the velocity fields that represent the turbulent inflow data.
Figure [\[fig:k2\]](#fig:k2){reference-type="ref" reference="fig:k2"} shows the transformer used in this study. Similar to the original transformer proposed by, it has two main components: encoder and decoder. The inputs of both components are passed through a positional encoding using sine and cosine functions, which can encode the order information of the input data into a vector and add it directly to the input vector. The encoder consists of six stacked encoder layers. Each layer contains a multi-head self-attention sublayer and a feed-forward sublayer. The input of the multi-head self-attention sublayer consists of queries (\(Q\)), keys (\(K\)), and values (\(V\)). Note that attention is a function that can map a query and set of key-value pairs to output, where the queries, keys, values and output are all vectors. The output can be calculated as the weighted sum of the values. The attention function is represented by scaled dot-product attention, which is an attention mechanism with the dot products are scaled down by \(\sqrt{d_k }\), where \(d_k\) is the dimension of \(Q\), \(K\) and it is equal to the dimension of \(V\), i.e. \(d_v\).
The scaled dot-product attention is calculated as follows: \[\label{eqn:eq1} \mathrm{Attention} (Q, K, V) = \mathrm{softmax} \left( \frac{Q K^T}{\sqrt{d_k}} \right) V,\] where softmax is a function that takes an input vector and normalizes it to a probability distribution so that the output vector has values that sum to 1. In the multi-head self-attention sublayer, \(d_v = d_{model}/h\), where \(d_{model}\) and \(h\) are the dimension of the input data to the model and number of heads, respectively.
The multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions such that \[\label{eqn:eq2} \mathrm{Multi-Head} (Q, K, V) = Concat \left( head_1 ,..., head_h \right) W^o,\] \[\label{eqn:eq3} head_i = \mathrm{Attention} \left( Q W_i^Q, K W_i^K, V W_i^V \right),\] where \(W_i^Q\), \(W_i^K\) and \(W_i^V\) are the weights corresponding to \(Q\), \(K\), \(V\) at every head, respectively; \(W^o\) represents the weights of the concatenated heads. \(W_i^Q \in \mathbb{R}^{d_{model} \times d_k }\), \(W_i^K \in \mathbb{R}^{d_{model} \times d_k }\), \(W_i^V \in \mathbb{R}^{d_{model} \times d_v }\) and \(W^o \in \mathbb{R}^{ hd_v \times d_{model} }\).
The multi-head self-attention sublayer contains six heads of scaled dot-product attention.
A residual connection is applied around the multi-head attention, followed by layer normalisation.
The second part of the encoder layer, i.e. the feed-forward sublayer, contains two dense layers with linear and rectified linear unit (ReLU) activation functions. This layer projects the vector to a larger space, where it is easier to extract the required information and then projects it back to the original space. As in the multi-head self-attention sublayer, a residual connection is employed before applying layer normalisation. Similar to the encoder, the decoder contains six decoder layers. In addition to the multi-head self-attention and feed-forward sublayers, the decoder layer has a third sublayer that performs multi-head attention over the output of the encoder stack. Furthermore, the multi-head self-attention sublayer is changed to a masked multi-head self-attention sublayer, as shown in figure [\[fig:k2\]](#fig:k2){reference-type="ref" reference="fig:k2"}, which is similar to the multi-head self-attention sublayer with the difference that the scaled dot-product attention is changed to a masked scaled dot-product attention. The masking operation ensures that the prediction can only depend on the known outputs, a fact that prevents later information leakage. In this study, the dropout technique is applied to every sublayer before the residual connection and the rate of dropout is set to 0.1.
The square of the \(L_2\) norm error is chosen as a loss function for the transformer such that \[\label{eqn:eq4} \mathcal{L}_{transformer}= \frac{1}{M} \sum_{m=1}^M \left\| {Output}_m-{Target}_m \right\|_2^2\] where \(Output\) and \(Target\) represent the output from the transformer and ground truth data, respectively, at a specific time step, \(m\). \(M\) represents the size of the training mini-batch, which is set to 64. The adaptive moment estimation (Adam) optimisation algorithm is used to update the weights of the model.
## MS-ESRGAN {#sec:MS-ESRGAN}
GANs have shown great success in image transformation and super-resolution problems. GAN-based models have also shown promising results in reconstructing HR turbulent flow fields from coarse data. In a GAN model that is used for image generation, two adversarial neural networks called the generator (\(G\)) and the discriminator (\(D\)) compete with each other. \(G\) tries to generate artificial images with the same statistical properties as those of the real ones, whereas \(D\) tries to distinguish the artificial images from the real ones. After successful training, \(G\) should be able to generate artificial images that are difficult to distinguish by \(D\). This process can be expressed as a min-max two-player game with a value function \(V(D, G)\) such that \[\label{eqn:eq5} \begin{split} \substack{min\\G} ~\substack{max\\D} ~V(D,G) = \mathbb{E}_{x_r \sim P_{data}(x_r)} [ {\rm log} D(x_r )] + \mathbb{E}_{z \sim P_z(z) } [ {\rm log} (1-D(G(z)))], \end{split}\] where \(x_r\) is the image from the ground truth data (real image) and \(P_{data}(x_r )\) is the real image distribution. \(\mathbb{E}\) represents the operation of calculating the average of all the data in the training mini-batch. In the second right term of equation [\[eqn:eq5\]](#eqn:eq5){reference-type="ref" reference="eqn:eq5"}, \(z\) is a random vector used as an input to \(G\) and \(D(x_r)\) represents the probability that the image is real and not generated by the generator. The output from \(G\), i.e. \(G(z)\), is expected to generate an image that is similar to the real one, such that the value of \(D(G(z))\) is close to 1. Meanwhile, \(D(x_r)\) returns a value close to 1, whereas \(D(G(z))\) returns a value close to 0. Thus, in the training process, \(G\) is trained in a direction that minimises \(V(D,G)\), whereas \(D\) is trained in a direction that maximises \(V(D,G)\). In this study, MS-ESRGAN is used to perform super-resolution reconstruction of the velocity fields predicted by the transformer. MS-ESRGAN is based on the enhanced super-resolution GAN (ESRGAN). Figure [\[fig:k3\]](#fig:k3){reference-type="ref" reference="fig:k3"} shows the architecture of MS-ESRGAN. As shown in figure [\[fig:k3\]](#fig:k3){reference-type="ref" reference="fig:k3"}(a), \(G\) consists of a deep CNN represented by residuals in residual dense blocks (RRDBs) and multiscale parts (MSP). Note that the input to \(G\) is low-resolution data, which are first passed through a convolutional layer and then through a series of RRDBs. The MSP, consisting of three parallel convolutional sub-models with different kernel sizes, is applied to the data features extracted by RRDBs. More details for MSP can be found in. The outputs of the three sub-models are summed and passed through a final convolutional layer to generate HR artificial data \((x_a)\). Figure [\[fig:k3\]](#fig:k3){reference-type="ref" reference="fig:k3"}(b) shows that the artificial and real data are fed to \(D\) and passed through a series of convolutional, batch normalisation and leaky ReLU layers. As a final step, the data are crossed over a convolutional layer. The non-transformed discriminator outputs using the real and artificial data, i.e. \(C(x_r )\) and \(C(x_a )\), are used to calculate the relativistic average discriminator value \(D_{Ra}\): \[\label{eqn:eq6} D_{Ra} (x_r, x_a ) = \sigma (C \left(x_r ) \right)-\mathbb{E}_{x_a} \left[C ( x_a) \right],\] \[\label{eqn:eq7} D_{Ra} (x_a, x_r ) = \sigma (C \left(x_a) \right)-\mathbb{E}_{x_r} \left[C ( x_r) \right],\] where \(\sigma\) is the sigmoid function. In equations [\[eqn:eq6\]](#eqn:eq6){reference-type="ref" reference="eqn:eq6"} and [\[eqn:eq7\]](#eqn:eq7){reference-type="ref" reference="eqn:eq7"}, \(D_{Ra}\) represents the probability that the output from \(D\) using the real image is relatively more realistic than the output using the artificial image.
Then, the discriminator loss function is defined as follows: \[\label{eqn:eq8} \ell_D^{Ra} =-\mathbb{E}_{x_r} \left[ {\rm log} (D_{Ra} (x_r, x_a)) \right]-\mathbb{E}_{x_a} \left[ {\rm log} (1-D_{Ra} (x_a, x_r )) \right].\] The adversarial loss function of the generator can be expressed in a symmetrical form as follows: \[\label{eqn:eq9} \ell_G^{Ra} =-\mathbb{E}_{x_r} \left[ {\rm log} (1-D_{Ra} (x_r, x_a )) \right]-\mathbb{E}_{x_a} \left[ {\rm log} (D_{Ra} (x_a, x_r )) \right].\] The total loss function of the generator is defined as \[\label{eqn:eq10} \mathcal{L}_G = \ell_G^{Ra} + \beta \ell_{pixel} + \ell_{perceptual},\] where \(\ell_{pixel}\) is the error calculated based on the pixel difference of the generated and ground truth data; \(\ell_{perceptual}\) represents the difference between features that are extracted from the real and the artificial data. The pre-trained CNN VGG-19 is used to extract the features using the output of three different layers. Here, \(\beta\) is a weight coefficient and its value is set to be 5000. The square of the \(L_2\) norm error is used to calculate \(\ell_{pixel}\) and \(\ell_{perceptual}\). The size of the mini-batch is set to 32. As in the transformer model, the Adam optimisation algorithm is used to update the model weights of the model.
# Data description and pre-processing {#sec:3}
The transitional boundary layer database available at the Johns Hopkins turbulence databases (JHTDB) is considered in this study for the training and testing of the DLM. The database was obtained via DNS of incompressible flow over a flat plate with an elliptical leading edge. In the simulation, the half-thickness of the plate, \(L\) and the free stream velocity, \(U_\infty\) are used as a reference length and velocity. Additionally, \(x\), \(y\) and \(z\) are defined as the streamwise, wall-normal and spanwise coordinates, respectively, with the corresponding velocity components, \(u\), \(v\), and \(w\). Note that the same definitions of the coordinates and velocity components are used in this study.
In the simulation, the length of the plate, \(L_x = 1,050L\) measured from the leading edge (\(x\) = 0), the domain height, \(L_y = 40L\) and the width of the plate, \(L_z = 240L\). The stored database in JHTDB is in the range \(x \in [30.2185, 1000.065]L\), \(y \in \left[0.0036, 26.488 \right]L\), and \(z \in \left[ 0, 240\right]L\). The corresponding number of grid points is \(N_x \times N_y \times N_z = 3320 \times 224 \times 2048 \approx 1.5 \times 10^9\). The database time step is \(\Delta t = 0.25L/U_\infty\). The stored database spans the following range in momentum-thickness-based Reynolds number, \(Re_\theta = U_\infty \theta/\nu \in \left[105.5,1502.0 \right]\), where \(\theta\) represents the momentum thickness and \(\nu\) is the kinematic viscosity. More details for the simulation and database can be found in and on the website of JHTDB.
In this study, the datasets within the range of \(Re_\theta \in \left[ 661.5, 1502.0 \right]\) are considered for training and testing the DLM. This range of \(Re_\theta\) in the database represents the fully-turbulent part of the flow. Datasets of the velocity components are collected from various (\(y-z\)) planes along the streamwise direction, with the number of snapshots = 4700 for every plane. To reduce the computational cost, the original size of each plane, \(N_y \times N_z = 224 \times 2048\) is reduced to \(112 \times 1024\). Furthermore, to increase the amount of training and testing data, every selected plane is divided into four identical sections along the spanwise direction, resulting in \(N_y \times N_z = 112 \times 256\) for each section. To obtain the coarse data, the size of the data is further reduced to \(N_y \times N_z = 14 \times 32\), which is obtained by selecting distributed points in the fields. A time series of 4000 snapshots for each section are used to train the DLM, resulting in a total number of training snapshots = \(4000\times4\times3=48000\). The fluctuations of the velocity fields are used in the training and prediction processes. The input data to the DLM are normalised using the min-max normalisation function to produce values between 0 and 1.
# Results and discussion {#sec:4}
## Results from the DLM trained at various \(Re_\theta\)
This section examines the capability of the proposed DLM to generate turbulent inflow data at three different Reynolds numbers for which the network has already been trained, \(Re_\theta\) = 661.5, 905.7, and 1362.0. Figures [\[fig:k4\]](#fig:k4){reference-type="ref" reference="fig:k4"}-[\[fig:k6\]](#fig:k6){reference-type="ref" reference="fig:k6"} show the instantaneous streamwise velocity (\(u^+\)) and vorticity (\(\omega_x^+\)) fields of the DNS and the predicted data for three different time steps, where the superscript '\(+\)' denotes normalisation by viscous inner scale; in the figures, \(\delta\) represents the boundary layer thickness. The figure shows that the instantaneous flow fields can be predicted using the model with a commendable agreement with the DNS data. Note that the model has shown a capacity to predict the instantaneous flow fields for a long period of time, more than the one required for the flow data to reach a statistically stationary state (reaching fixed first and second-order statistics over time), i.e. for a number of time steps = 10000.
Figure [\[fig:k7\]](#fig:k7){reference-type="ref" reference="fig:k7"} shows the probability density functions (p.d.f.) of the velocity components (\(u^+\), \(v^+\), and \(w^+\)) plotted against the wall-normal distance (\(y^+\)). The figure shows that the p.d.f. plots of the generated velocity components are generally consistent with the p.d.f. plots obtained from the DNS data, indicating the capability of the model in predicting the velocity fields with distributions of the velocity components that are consistent with those of the DNS data.
Figures [\[fig:k8\]](#fig:k8){reference-type="ref" reference="fig:k8"}-[\[fig:k10\]](#fig:k10){reference-type="ref" reference="fig:k10"} compare the turbulence statistics of the generated velocity fields with the turbulence statistics of the DNS data. As shown in the figures, the mean streamwise velocity profile (\(U^+\)) for all the three Reynolds numbers shows excellent agreement with the results obtained from the DNS. The comparison of root-mean-square profiles of the velocity components (\(u_{rms}^+\), \(v_{rms}^+\), and \(w_{rms}^+\)) reveals good agreement with the results obtained from the DNS. However, the profile of the Reynolds shear stress (\(\overline{u'v'}^+\)) shows a slight underprediction in the region between near the wall and the maximum Reynolds shear stress and the profile values in this region improve as the Reynolds number increases. This might be attributed to the fact that with the increase in the boundary-layer thickness, the effect of zero padding in the convolution processes is decreased in MS-ESRGAN, resulting in a better prediction of the velocity fields in this region of the boundary layer.
The capability of the proposed DLM to produce realistic spatial spectra of the velocity fields is investigated by employing the premultiplied spanwise wavenumber spectra, \(k_z \Phi_{\alpha \alpha}\), where \(\Phi_{\alpha \alpha}\) represents the spanwise wavenumber spectra, \(\alpha\) represents the velocity component and \(k_z\) is the spanwise wavenumber. Figure [\[fig:k11\]](#fig:k11){reference-type="ref" reference="fig:k11"} shows the contour plots of \(k_z^+ \Phi_{\alpha \alpha}^+\) as a function of \(y^+\) and the spanwise wavelength, \(\lambda_z^+\). The figure shows that the spectra of the velocity components are generally consistent with those obtained from the DNS data with a slight deviation at the high wavenumbers. This indicates that the two-point correlations of the generated velocity components are consistent with those obtained from the DNS data, further supporting the excellent performance of the proposed DLM to properly represent the spatial distribution of the velocity fields. It is worth noting that the ability of the model to reproduce accurate spectra is essential in generating the turbulent inflow conditions to guarantee that the turbulence will be sustained after introducing the synthetic-inflow; otherwise, the generated inflow would require very long distances to reach \"well-behaved\" turbulent conditions, and these fluctuations could also dissipate.
To evaluate the performance of the proposed DLM to generate the velocity fields with accurate dynamics, the frequency spectra, \(\phi_{\alpha \alpha}^+\), as a function of \(y^+\) and the frequency, \(f^+\) are represented in figure [\[fig:k12\]](#fig:k12){reference-type="ref" reference="fig:k12"}. Note that the spectra obtained from the generated velocity fields show a commendable agreement with those of the DNS data, indicating that the proposed DLM can produce turbulent inflow conditions with a temporal evolution of the velocity fields that is consistent with that of the DNS.
## Interpolation and extrapolation capability of the DLM
This section investigates the performance of the proposed DLM to generate turbulent inflow conditions at Reynolds numbers that are not used in the training process. The velocity fields at \(Re_\theta\) = 763.8 and 1155.1 are used as examples of the velocity fields that fall between the Reynolds numbers used in the training process, i.e. the interpolation ability of the model is investigated using the flow fields at these Reynolds numbers.
Figure [\[fig:k13\]](#fig:k13){reference-type="ref" reference="fig:k13"} shows the instantaneous streamwise velocity and vorticity fields for the flow at \(Re_\theta\) = 763.8. It is worth noting that the transformer is trained for the flow at the nearest \(Re_\theta\), i.e. \(Re_\theta\) = 661.5 is used to predict the temporal evolution of the velocity fields. The figure shows that the main features of the flow fields can be obtained with relatively good precision; however, the details of the predicted velocity fluctuations are not clearly shown. Similar results can be observed in figure [\[fig:k14\]](#fig:k14){reference-type="ref" reference="fig:k14"} for the predicted velocity fields at \(Re_\theta\) = 1155.1. Here, the current transformers trained for the flow at \(Re_\theta\) = 905.7 and 1362.0 are used to predict the temporal evolution of the velocity fields.
The turbulence statistics of the flow at \(Re_\theta\) = 763.8 and 1155.1 are shown in figures [\[fig:k15\]](#fig:k15){reference-type="ref" reference="fig:k15"} and [\[fig:k16\]](#fig:k16){reference-type="ref" reference="fig:k16"}, respectively. Although the mean streamwise velocity and the root-mean-square profiles of the spanwise and wall-normal velocity components show an ability of the DLM to predict reasonably well, the root-mean-square profile of the streamwise velocity component and the mean Reynolds shear stress show an under-prediction due to the lack of detailed information on the velocity fluctuations. The extrapolation ability of the DLM is evaluated using the flow fields at \(Re_\theta\) = 1502.0, which is higher than the maximum \(Re_\theta\) used to train the transformer and MS-ESRGAN, i.e. \(Re_\theta\) = 1362.0. The transformer trained for the flow at \(Re_\theta\) = 1362.0 is used to predict the dynamics of the velocity fields. Figure [\[fig:k17\]](#fig:k17){reference-type="ref" reference="fig:k17"} shows that the generated instantaneous streamwise velocity and vorticity fields generally have similar accuracy to the interpolated flow fields. Meanwhile, the turbulence statistics show a deviation from the DNS statistics, as shown in figure [\[fig:k18\]](#fig:k18){reference-type="ref" reference="fig:k18"}. This can be attributed to the lack of details of the velocity fluctuations and the extrapolation process that relies on the flow information at one Reynolds number compared with the interpolation process where the flow falls within the range of the Reynolds numbers that the MS-ESRGAN is trained for.
Finally, the accuracy of the spectral content of the interpolated and extrapolated velocity components is examined in figure [\[fig:k19\]](#fig:k19){reference-type="ref" reference="fig:k19"} by employing the premultiplied spanwise wavenumber spectra. These results indicate that the spectra are produced with relatively good accuracy for the low-moderate wavenumbers.
## Error analysis, transfer learning and computational cost
The performance of the proposed DLM is further statistically investigated using the \(L_2\) norm error of the predicted data of all the Reynolds numbers used in this study, \[\label{eqn:eq11} \varepsilon = \frac{1}{J} \sum_{j=1}^J \frac{\left\| {\alpha}_j^{DNS}-{\alpha}_j^{pred}\right\|_2 }{\left\| {\alpha}_j^{DNS}\right\|_2}\] where \(\alpha_j^{DNS}\) and \(\alpha_j^{pred}\) represent the ground truth (DNS) and the predicted velocity components, respectively, and \(J\) represents the number of the test snapshots. Figure [\[fig:k20\]](#fig:k20){reference-type="ref" reference="fig:k20"} shows that as the Reynolds number increases, no significant differences can be seen in the error values of the predicted velocity fields. However, as expected, the error shows higher values for the interpolated and extrapolated velocity fields compared with the error of the predicted velocity fields at the Reynolds numbers that the DLM is trained for. Furthermore, in contrast with the aforementioned statistical results, the error values are relatively high for the wall-normal and spanwise velocity fields. This indicates that the DLM has learned to model the structure of the flow with generally accurate turbulence statistics and spatio-temporal correlations, rather than reproducing the time sequence of the flow data. This observation is consistent with the results obtained by.
It is worth mentioning that the transfer learning (TL) technique is used in this study. The weights of the transformer are sequentially transferred for every training (\(y-z\)) plane in the flow. First, the transformer is trained for the flow at the lowest Reynolds number, i.e. \(Re_\theta\) = 661.5. After that, the weights of the model are transferred for the training using the next \(Re_\theta\) data and so on. The results from using TL in this study show that with the use of only 25% of the training data for the transformer model, the computational cost (represented by the training time) can be reduced by 52% without affecting the prediction accuracy. These results are consistent with the results obtained by.
The total number of trainable parameters of the DLM is 356.5 million (305.5 million for the transformer and 51 million for the MS-ESRGAN). The training of the transformer model for all the three Reynolds numbers used in this study using a single NVIDIA TITAN RTX GPU with the aid of TL requires approximately 23 hours. Meanwhile, the training of the MS-ESRGAN requires approximately 32 hours. Thus, the total training time of the DLM model is 55 hours, indicating that the computational cost of the model is relatively lower than the cost of DNS required to generate the velocity fields. Furthermore, this computational cost is required only once, i.e. for the training of the model. Since the prediction process is computationally inexpensive and does not need any data for the prediction (except the initial instantaneous fields), the DLM can also be considered efficient in terms of storing the inflow data.
# Conclusions {#sec:5}
This study proposed a deep-learning-based method to generate turbulent inflow conditions for spatially-developing TBL simulations. A combination of a transformer and MS-ESRGAN was used to build the inflow generator. The transformer was trained to model the temporal evolution of the velocity fields represented by various (\(y-z\)) planes of spatially limited data. Meanwhile, MS-ESRGAN was trained to perform super-resolution reconstruction of the predicted velocity fields.
The generated instantaneous velocity fields showed an excellent agreement with the DNS results for the velocity fields at Reynolds numbers that the DLM was trained for. The model also successfully reproduced the turbulence statistics with commendable accuracy. Furthermore, the model reproduced the spectra of the velocity components with accurate precision, indicating accurate spatial and temporal correlations of the generated velocity components, which further supports the ability of the model to maintain the realistic behaviour of the velocity fields.
The performance of the proposed model was further examined using velocity fields at Reynolds numbers that were not used in the training process. The instantaneous and statistical results showed a reasonable accuracy for the interpolated and extrapolated velocity fields. The spectra of the velocity components revealed a relatively good agreement with the results from the actual velocity data, with a deviation that can be observed at high wavenumbers. These results suggest that the model can generate the turbulent inflow conditions for the flow at Reynolds numbers that are not necessarily used in the training of the model.
The results obtained from the error analysis showed that the increase in the Reynolds number has no significant effect on the error values of the predicted velocity fields, indicating that the model is robust to the increase of the Reynolds number. The use of TL in the training of the transformer revealed a noticeable reduction in the computational cost of the DLM without affecting the precision of the prediction.
This study showed for the first time that a transformer-based model could be effectively used for modelling the dynamics of turbulent flows with the ability to perform parallel computing during the training process, which is not possible in the LSTM-based models. It also paved the way for utilising synthetic-inflow generators for large-scale turbulence simulations using deep learning, with significant promise in terms of computational savings. | {'timestamp': '2022-06-06T02:15:59', 'yymm': '2206', 'arxiv_id': '2206.01618', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01618'} |
# Introduction [\[sec:intro\]]{#sec:intro label="sec:intro"}
Infinite power series solutions to ordinary differential equations (ODEs) are useful if they are convergent while also satisfying the given constraints over the domain on which they are defined. It is within this context that we develop a convergent series solution to a classical nonlinear ODE in fluid mechanics--the Sakiadis boundary layer flow along a moving wall. To date, this problem, as well as the related well-known Blasius problem along a stationary wall, do not have proven exact analytical solutions in the literature, although approximate analytical solutions have been put forward and are discussed later in this section. The Sakiadis boundary layer is an important flow field in configurations where thin liquid films are coated onto moving substrates , and is an essential component of hydrodynamic assist in high speed curtain coating . The boundary layer surrounding a flat plate moving through a viscous incompressible fluid was first examined in the literature by Sakiadis who applied Blasius's similarity transform to Prandtl's boundary layer equations (with appropriate boundary conditions, B.C.s) to arrive at a third-order nonlinear ODE in \(f(\eta)\) refered to here as the 'Sakiadis Problem',
By contrast, the Blasius problem describing a stationary plate in a moving fluid is governed by the same operator ([\[sDE\]](#sDE){reference-type="ref" reference="sDE"}) but has conditions \[\text{Blasius B.C.s: }~f(0)=0, ~ f'(0)=0, ~ f'(\infty) = 1. \label{bBC}\] While both the Sakiadis and Blasius problems can be handled in similar ways numerically (e.g., shooting, transformation) , the difference in boundary conditions leads to different (approximate) analytical approaches; and, the nonlinear nature of the equations yields distinctly different solutions. A common measure of the accuracy of any solution technique applied to either problem is the quantity \(\kappa\), defined as \[\kappa\equiv f''(0), \label{eq:kappa}\] which is directly related to the wall shear stress in the boundary-layer flow, and is typically referred to as the "wall shear" parameter . For a given \(\kappa\)---which is known numerically and is determined in this paper algorithmically---an infinite power series solution for both problems can be obtained through standard means, given as \[f=\sum_{n=0}^\infty a_n \eta^n,~~|\eta|<S, \label{eq:series}\] where \(S\) is a finite radius of convergence. Singularities that lie outside of the physical domain at a distance \(S\) away from \(\eta=0\) are the cause of this radius. The Blasius problem has 3 well-known singularities lying the same distance \(S\) from the origin (see and historical review therein) and are reported to lie at values \(\eta=S\text{ exp }[i(2j+1)\pi/3]\) (\(j\)=0, 1, 2) where \(S\approx 5.6900380545\). For the Sakiadis problem, \(S\approx4.07217\), arising from singularities lying off the real-line in the left half-plane . Approximate resummations are available, that bypass the original series' convergence barrier caused by these singularities for both the Sakiadis and Blasius problems. One approach for avoiding (non-physical) singularities that restrict series convergence is to re-expand the series by mapping the independent variable such that the (non-physical) closest singularities no longer affect the physical domain. The divergent Blasius series has been successfully re-summed in this way by Boyd through re-expanding ([\[eq:series\]](#eq:series){reference-type="ref" reference="eq:series"}) as
is the *gauge function* by which ([\[eq:Boydseries\]](#eq:Boydseries){reference-type="ref" reference="eq:Boydseries"}) may be cast as a formal Taylor series, specifying how terms are asymptotically ordered . In ([\[eq:Boyd\]](#eq:Boyd){reference-type="ref" reference="eq:Boyd"}), the three-fold symmetry of the closest singularities (with modulus \(S\)) in the complex \(\eta\)-plane are mapped to infinity, due to the effect of the above transformation on their orientation. This leads to a new radius of convergence \(\tilde{S}\) that lies outside the original physical domain, thus creating a convergent series solution for the Blasius problem on the positive real line. The coefficients \(\tilde{b}_n\) are obtained by equating ([\[eq:series\]](#eq:series){reference-type="ref" reference="eq:series"}) with the expansion of ([\[eq:Boyd\]](#eq:Boyd){reference-type="ref" reference="eq:Boyd"}) about \(\eta=0\); these may be obtained recursively (only depending on prior coefficients) due to the gauge function ([\[eq:gauge\]](#eq:gauge){reference-type="ref" reference="eq:gauge"}) and its derivatives equaling \(0\) at \(\eta=0\). The substitution made in ([\[eq:Boyd\]](#eq:Boyd){reference-type="ref" reference="eq:Boyd"}) may be considered a modified Euler transformation[^1]. While the series ([\[eq:Boyd\]](#eq:Boyd){reference-type="ref" reference="eq:Boyd"}) explicitly incorporates singularities in order to bypass the radius of convergence of the original series ([\[eq:series\]](#eq:series){reference-type="ref" reference="eq:series"}), another approach is to consider the other side of the domain (\(\eta\to\infty\)) as a possible expansion point. Using the method of dominant balance, the \(\eta\to\infty\) behaviors for the Blasius and Sakiadis solutions are given in as \[\text{Blasius: }f \sim \eta+B+4Q \frac{\exp[-\eta^2/4-B\eta/2]}{(\eta+B)^2}\left[1+O\left(\frac{1}{(\eta+B)^2}\right)\right],~\textrm{as} \ \eta \rightarrow \infty, \label{eq:blasiusA}\] \[\text{Sakiadis: }f \sim C+ G e^{-C\eta/2} + \frac{G^2}{4C} e^{-C\eta} + O\left(e^{-3C\eta/2}\right),\ \textrm{as} \ \eta \rightarrow \infty, \label{eq:sakiadisA}\] where \(B\), \(Q\), \(C\), and \(G\) are constants arising from integration. We repeat the above expressions here to highlight a key difference and to motivate the technique used to examine the Sakiadis problem in this paper. A necessary condition for an expansion to converge is that it does not introduce new singularities into the problem that lie in the physical domain. In the Blasius problem, the constant \(B\) takes on a negative value and thus ([\[eq:blasiusA\]](#eq:blasiusA){reference-type="ref" reference="eq:blasiusA"}) has a singularity at the positive real value of \(-B\). The expansion ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}) for the Sakiadis problem does not have this issue, and it is thus possible that the expansion converges; this will be discussed in more detail shortly and revisited throughout the paper. Barlow et al. used the asymptotic form ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}) to propose an approximant to sum the divergent series ([\[eq:series\]](#eq:series){reference-type="ref" reference="eq:series"}) for the Sakiadis problem as \[\text{Sakiadis: }f_{A,N}=C+\sum_{n=1}^N \mathcal{A}_n \left(e^{-C\eta/2}\right)^n. \label{eq:AA}\] In ([\[eq:AA\]](#eq:AA){reference-type="ref" reference="eq:AA"}), the \(\mathcal{A}_n\) coefficients are computed by solving an \(N\times N\) system enforcing that the \(N\)-term Taylor expansion of ([\[eq:AA\]](#eq:AA){reference-type="ref" reference="eq:AA"}) about \(\eta=0\) is exactly the \(N^\mathrm{th}\) truncation of ([\[eq:series\]](#eq:series){reference-type="ref" reference="eq:series"}). Unlike the recursive computation of \(\tilde{b}_n\) in ([\[eq:Boyd\]](#eq:Boyd){reference-type="ref" reference="eq:Boyd"}), all \(\mathcal{A}_n\) coefficients change their value as \(N\) changes, and thus ([\[eq:AA\]](#eq:AA){reference-type="ref" reference="eq:AA"}) cannot be considered a formal power series solution to ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}). Nevertheless, \(f_{A,N}\) converges to the numerical solution of \(f\) over the physical domain as \(N\) is increased. Note that, although not remarked on in , each \(\mathcal{A}_n\) coefficient converges to a specific value as \(N\) is increased. Additionally, ([\[eq:AA\]](#eq:AA){reference-type="ref" reference="eq:AA"}) was used in to compute \(\kappa\), as well as the asymptotic constants in ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}) (\(C\) and \(G\)) to within 12 digits of accuracy (beyond previously reported numerical results) before hitting a round-off barrier. Note that, in using approximant ([\[eq:AA\]](#eq:AA){reference-type="ref" reference="eq:AA"}), did not utilize knowledge of the singularities that caused the original series ([\[eq:series\]](#eq:series){reference-type="ref" reference="eq:series"}) to diverge. However, the convergence of ([\[eq:AA\]](#eq:AA){reference-type="ref" reference="eq:AA"}) suggests that changing the gauge function from \(\eta\) to \(e^{-C\eta/2}\) fortuitously creates a mappping that removes the influence of the limiting singularities present in ([\[eq:series\]](#eq:series){reference-type="ref" reference="eq:series"}); this is similar in behavior to the modified Euler transformation used in ([\[eq:Boyd\]](#eq:Boyd){reference-type="ref" reference="eq:Boyd"}) for the Blasius problem. In the current work, we develop a formal Taylor series solution to the Sakiadis problem using the exponential gauge transformation mention above, with the goal of providing further explanation for the improved accuracy of ([\[eq:AA\]](#eq:AA){reference-type="ref" reference="eq:AA"}). Interestingly, another classical problem in fluid mechanics that has a similar mathematical structure is a static air--liquid meniscus (with surface tension) formed when a flat wall is placed in an infinite horizontal pool. Both the meniscus and the Sakiadis problem have (a) the same asymptotic decaying exponential behavior at infinite distance from the wall and (b) a singularity structure that limits convergence of the standard power series solution, yet benefits from an exponential transformation. For the meniscus problem, the exact solution---given as distance from the wall as function of meniscus height---has long been known . Thus, the emphasis of this 2\(^\mathrm{nd}\) analysis is not in providing a new solution form, but rather in exploiting the structural similarities with the Sakiadis problem to better understand the nature of exponential gauge transformations and their effect on convergence properties. The paper is organized as follows. In section [\[sec:Sakiadis\]](#sec:Sakiadis){reference-type="ref" reference="sec:Sakiadis"}, we review the series and approsimant solutions to the Sakiadis boundary layer problem, expanding upon the background given above. In section [\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}, an asymptotically motivated gauge function is used to map the Sakiadis ODE and B.C.s to a domain in which its series solution converges over the whole un-mapped physical domain. In section [\[sec:predictions\]](#sec:predictions){reference-type="ref" reference="sec:predictions"}, we provide a procedure for using the analytic solution (from section [\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}) to compute the asymptotic constants \(C\) and \(G\), the wall shear parameter \(\kappa\), and locations of the closest singularities (from \(\eta\)=0) of the Sakiadis function to arbitrary precision. In section [\[sec:Meniscus\]](#sec:Meniscus){reference-type="ref" reference="sec:Meniscus"}, the classical problem of a meniscus at a flat wall is reviewed and its solution is obtained in section [\[sec:divergent\]](#sec:divergent){reference-type="ref" reference="sec:divergent"} as a divergent power series expansion about the wall location. An asymptotic expansion away from the wall and a convergent series solution motivated by this expansion are provided in section [\[sec:MeniscusMapped\]](#sec:MeniscusMapped){reference-type="ref" reference="sec:MeniscusMapped"}, following a similar technique to the Sakiadis problem in section [\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}. Remarks on the overall methodology applied to both problems are provided in section [\[sec:conclusions\]](#sec:conclusions){reference-type="ref" reference="sec:conclusions"}.
# The Sakiadis boundary layer
## A divergent power series and a convergent approximant [\[sec:Sakiadis\]]{#sec:Sakiadis label="sec:Sakiadis"}
Although the main results of this subsection are found in (and references therein), here we elucidate elements of that work that were previously not discussed and are relevant to the work herein. The Sakiadis boundary layer along a moving flat wall is obtained as a solution to the nonlinear ODE ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}) where \(f\) and \(\eta\) are the similarity variables, related to the physical variables in the governing boundary layer equation through a similarity transform. The general power series solution to the ODE ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}a) was first developed by , and is expressed as
For the Sakiadis problem, \(a_0=a_1=0\) as prescribed by ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}b), and \(a_2=\kappa/2\) as defined in ([\[eq:kappa\]](#eq:kappa){reference-type="ref" reference="eq:kappa"}). The "wall shear" parameter \(\kappa\) is computed either numerically or analytically in a self-consistent way such that the condition \(f'(\infty)=0\) in ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}b) is met. The value of \(\kappa\) is given in as \[\kappa=-0.443748313369\dots, \label{eq:kappaEst1}\] the precision of which is improved upon in section [\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}. The series ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) diverges within the physical domain, as indicated in figure [\[fig:SakiadisUnMapped\]](#fig:SakiadisUnMapped){reference-type="ref" reference="fig:SakiadisUnMapped"}a (dashed lines), where it is compared to the numerical solution (\(\bullet\)'s, RK4 with \(\Delta \eta=10^{-5}\)). Using the modified Padé approximants given in , an estimate of the locations of the convergence-limiting singularities in the complex plane are found[^2] to be at \(\eta_{s\pm} \approx-1.2114 \pm 3.8878i\), indicating a radius of convergence of \(S=|\eta_{s\pm}| \approx 4.0722\), shown in figure [\[fig:SakiadisUnMapped\]](#fig:SakiadisUnMapped){reference-type="ref" reference="fig:SakiadisUnMapped"}a as a vertical line. The series ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) is analytically continued beyond this radius of convergence in in an approximate way by constructing an approximant based on the \(\eta\to\infty\) expansion given by ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}). As mentioned in section [\[sec:intro\]](#sec:intro){reference-type="ref" reference="sec:intro"}, a particular sequence of approximants having the form of the asymptotic behavior ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}) are examined in as \[f_{A, N}(\eta) = C+\sum_{k=1}^N \mathcal{A}_k \left(e^{-C\eta /2}\right)^k \label{SakApproxForm}\] where the \(\mathcal{A}_k \dots \mathcal{A}_N\) coefficients are computed such that the expansion of ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) about \(\eta=0\) exactly matches the Taylor series solution ([\[SakTaylor0\]](#SakTaylor0){reference-type="ref" reference="SakTaylor0"}) to order \(N\). Taylor expanding each term of ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) to order \(N\), switching the order of summation, and setting the result equal to ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) leads to a \(N\times N\) Vandermonde system, whose explicit inversion formula is known . For example, at \(N=2\), we have \[\left[ \begin{array}{cccc} 1 &&& 1 \\ 1 &&& 2 \end{array} \right]\left[ \begin{array}{c} \mathcal{A}_1\\ \mathcal{A}_2 \end{array} \right]=\left[ \begin{array}{c} 0!~a_0-C\\ 1!~(-2/C)~a_1 \end{array} \right], \label{matrix2}\] and for \(N\)=3 we have \[\left[ \begin{array}{ccccccc} 1 &&& 1 &&& 1 \\ 1 &&& 2 &&& 3 \\ 1 &&& 4 &&& 9 \end{array} \right]\left[ \begin{array}{c} \mathcal{A}_1\\ \mathcal{A}_2 \\ \mathcal{A}_3 \end{array} \right]=\left[ \begin{array}{c} 0!~a_0-C\\ 1!~(-2/C)~a_1 \\ 2!~ (-2/C)^2 ~a_2 \end{array} \right], \label{matrix3}\] and so on. Note that, after inverting the matrices above, \(\mathcal{A}_1\) and \(\mathcal{A}_2\) attain different values in ([\[matrix2\]](#matrix2){reference-type="ref" reference="matrix2"}) than in ([\[matrix3\]](#matrix3){reference-type="ref" reference="matrix3"}). The relevance of this issue is discussed in what follows. Also, although the Vandermonde matrix has an explicit inversion formula, high precision arithmetic (i.e., beyond double) must be used for large \(N\) to avoid round-off error. In principle, if \(\kappa\) and \(C\) are known, ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) provides a complete approximate solution. A slight modification can also be used to estimate \(\kappa\) and \(C\) as well, wherein the \(\mathcal{A}_k\) coefficients are chosen to match the Taylor coefficients to order \(N\), but \(\kappa\) and \(C\) are chosen to make \(\mathcal{A}_N=\mathcal{A}_{N-1}=0\) (effectively switching \(\mathcal{A}_N\) and \(\mathcal{A}_{N-1}\) with \(\kappa\) and \(C\) as unknowns of the \(N\times N\) system). This is precisely what is done in to compute the value of \(\kappa\) given by ([\[eq:kappaEst1\]](#eq:kappaEst1){reference-type="ref" reference="eq:kappaEst1"}) and to provide a \(C\) value of \[C=1.61612544681\dots, \label{CEst1}\] the precision of which is improved upon in section [\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}. After using the \(\kappa\) and \(C\) obtained via the above process, approximant ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) converges uniformly over the domain as \(N\) is increased, as shown in figure [\[fig:SakiadisUnMapped\]](#fig:SakiadisUnMapped){reference-type="ref" reference="fig:SakiadisUnMapped"}a (solid curves). Curves for \(f'\) and \(f''\), found analytically from ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}), are also shown to uniformly converge in . To revisit the discussion just above, it is found that, as the approximant ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) itself converges and the values of \(A_k\) change with increasing \(N\), \(A_1\) and \(A_2\) converge respectively to values \[A_1=G=-2.1313459241\dots \label{Gest1}\] and \(A_2=G^2/(4C)\), matching the \(\eta\to\infty\) expansion ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}) obtained from dominant balance. The other \(A_k\) coefficients also converge, albeit slower as \(k\) increases, as the portion of the approximant that is used to match the behavior at \(\eta=0\) is pushed to higher-order terms. Although the approximant ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) is a *convergent* series, its coefficients depend on truncation \(N\) and so it cannot be considered a formal *Taylor* series. Thus, we cannot apply Taylor's theorem to ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) to relate convergence and singularity location in the context of switching between a gauge function of \(\eta\) in ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) to a gauge function of \(e^{-C\eta/2}\) in ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}). In the next section, we remove this coefficient dependence on truncation while retaining the asymptotically motivated gauge function of \(e^{-C\eta/2}\); in doing so, we provide an explanation for why the approximant ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) converges without explicitly incorporating singular behavior.
## Variable transform and convergent power series solution [\[sec:SakiadisMapped\]]{#sec:SakiadisMapped label="sec:SakiadisMapped"}
Motivated by the asymptotic expansion ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}) and encouraged by the convergence of ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}), we define the variable transformations \[\begin{aligned}
\omega(\eta) &= \exp\left(-\frac{C}{2} \eta\right) \label{SakTransDefs} \end{aligned}\] \[\begin{aligned}
f(\eta) &= g(\omega(\eta)) \label{composition} \end{aligned}\] with the goal of finding the Taylor coefficients of the function \(g\) such that \(g \circ \omega\) satisfies the ODE ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}). Substituting ([\[SakTransDefs\]](#SakTransDefs){reference-type="ref" reference="SakTransDefs"}) and ([\[composition\]](#composition){reference-type="ref" reference="composition"}) into ([\[sDE\]](#sDE){reference-type="ref" reference="sDE"}), applying the chain rule, and rearranging terms, leads to the transformed ODE \[\frac{g}{C}\left(\omega\ddot{g} + \dot{g}\right) = \omega^2 \dddot{g} + 3\omega \ddot{g} + \dot{g} \label{SakTransformedODE}\] where \(\dot{g}\) denotes a derivative with respect to \(\omega\). The boundary conditions ([\[sBC\]](#sBC){reference-type="ref" reference="sBC"}) become \[g(1)=0,~~\dot{g}(1)=\frac{-2}{C},~~\ddot{g}(1)=\frac{2}{C}+\frac{4\kappa}{C^2} \label{eq:transcond}\] where the third condition in ([\[sBC\]](#sBC){reference-type="ref" reference="sBC"}) here leads to \(0=0\) and so the transformation of condition ([\[eq:kappa\]](#eq:kappa){reference-type="ref" reference="eq:kappa"}) is written instead for \(\ddot{g}(1)\). In the usual way, we assume a solution to ([\[SakTransformedODE\]](#SakTransformedODE){reference-type="ref" reference="SakTransformedODE"}) of the form \[g(\omega) = \sum_{n=0}^\infty \tilde{a}_n \omega^n, \label{eq:omega}\] which is readily differentiated term-by-term to compute \(\dot{g}\), \(\ddot{g}\), and \(\dddot{g}\). After employing Cauchy's product rule to handle the nonlinear term on the left-hand side of ([\[SakTransformedODE\]](#SakTransformedODE){reference-type="ref" reference="SakTransformedODE"}), the ODE becomes \[\begin{gathered}
\frac{\omega}{C}\sum_{n=0}^\infty \left[\sum_{k=0}^n (k+1)(k+2)\tilde{a}_{k+2}\tilde{a}_{n-k}\right]\omega^n + \frac{1}{C}\sum_{n=0}^\infty \left[\sum_{k=0}^n (k+1)\tilde{a}_{k+1}\tilde{a}_{n-k}\right]\omega^n \\ = \omega^2 \sum_{n=0}^\infty (n+1)(n+2)(n+3)\tilde{a}_{n+3}\omega^n + 3\omega \sum_{n=0}^\infty (n+1)(n+2)\tilde{a}_{n+2}\omega^n \\ + \sum_{n=0}^\infty (n+1)\tilde{a}_{n+1}\omega^n. \label{SakTransformedExpandedForm} \end{gathered}\] After carrying through the \(\omega\) and \(\omega^2\) in the above and shifting the series accordingly, we may equate like-terms to obtain the coefficients. In particular, for \(n\geq 2\), equating both sides of ([\[SakTransformedExpandedForm\]](#SakTransformedExpandedForm){reference-type="ref" reference="SakTransformedExpandedForm"}) leads to the recurrence relation
As we can see, ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) is a Taylor series solution to ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) in terms of the gauge function \(e^{-C\eta /2}\). Although the implementation of ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) requires knowledge of \(C\) and \(G\), we will show in section [\[sec:predictions\]](#sec:predictions){reference-type="ref" reference="sec:predictions"} that these may be computed (to any desired precision) directly from ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) via the conditions ([\[eq:transcond\]](#eq:transcond){reference-type="ref" reference="eq:transcond"}). Interestingly, ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) appears to be a recursive way to compute the asymptotic series ([\[eq:sakiadisA\]](#eq:sakiadisA){reference-type="ref" reference="eq:sakiadisA"}), originally obtained in via the method of dominant balance. Using values of \(C\) and \(G\) determined in section [\[sec:predictions\]](#sec:predictions){reference-type="ref" reference="sec:predictions"}, convergence of ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) in shown as solid lines in figure [\[fig:SakiadisUnMapped\]](#fig:SakiadisUnMapped){reference-type="ref" reference="fig:SakiadisUnMapped"}b. An error plot is shown in figure [\[fig:SakiadisErrorMapped\]](#fig:SakiadisErrorMapped){reference-type="ref" reference="fig:SakiadisErrorMapped"}b, indicating convergence to within machine precision over the full domain (shown versus \(\omega\in[0,1]\) for clearer representation of the limit \(\eta\to\infty\)). For comparison, error from the divergent Sakiadis series ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) and approximant ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) are shown in figure [\[fig:SakiadisErrorMapped\]](#fig:SakiadisErrorMapped){reference-type="ref" reference="fig:SakiadisErrorMapped"}a over the same domain. Although not shown here, error for \(f'\) and \(f''\), which may be analytically obtained from ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}), follow the same error trends as in figure [\[fig:SakiadisErrorMapped\]](#fig:SakiadisErrorMapped){reference-type="ref" reference="fig:SakiadisErrorMapped"}b. We now provide an explanation of why ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) (and by extension ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"})) converge to the numerical solution of ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}). As alluded to in section [\[sec:intro\]](#sec:intro){reference-type="ref" reference="sec:intro"}, the singularities that cause the original series ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) to diverge are mapped via the gauge function ([\[SakTransDefs\]](#SakTransDefs){reference-type="ref" reference="SakTransDefs"}) such that their influence lies beyond the physical domain; this is shown in figure [\[fig:SakiadisMapped\]](#fig:SakiadisMapped){reference-type="ref" reference="fig:SakiadisMapped"} where the complex \(\eta\) and \(\omega\) planes are compared. Here, we track the movement of the two singularities \(\eta_{s,\pm}\) (shown by \(\ast\)'s in the figure) closest to \(\eta=0\), whose locations are predicted by the Padé analysis of (see section [\[sec:Sakiadis\]](#sec:Sakiadis){reference-type="ref" reference="sec:Sakiadis"}) and later confirmed in section [\[sec:predictions\]](#sec:predictions){reference-type="ref" reference="sec:predictions"}. In the \(\eta\) plane of figure [\[fig:SakiadisMapped\]](#fig:SakiadisMapped){reference-type="ref" reference="fig:SakiadisMapped"}, the circle of convergence for ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) is drawn based on these two singularities. Note that this circle intersects the positive real line at the same location as the radius of convergence of series ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) drawn in figure [\[fig:SakiadisUnMapped\]](#fig:SakiadisUnMapped){reference-type="ref" reference="fig:SakiadisUnMapped"}a, as is expected from Taylor's theorem, which only guarantees convergence of Taylor series within such circles . In the \(\omega\) plane of figure [\[fig:SakiadisMapped\]](#fig:SakiadisMapped){reference-type="ref" reference="fig:SakiadisMapped"}, the circle of convergence for ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) is drawn, centered at \(\omega=0\) and with mapped radius \(|e^{-\left(C\eta_{s,\pm}\right) /2}|\); note that this circle now extends beyond the physical domain itself[^3], which explains why ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) converges over the entire physical domain. On the mapped \(\omega\)-plane shown in figure [\[fig:SakiadisMapped\]](#fig:SakiadisMapped){reference-type="ref" reference="fig:SakiadisMapped"}, note that, in addition to drawing the circle of convergence corresponding to expansion ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}), a larger circle corresponding to the expansion about \(\omega=1\) is also drawn, which is larger because the mapped singularities are farther away. This expansion point corresponds to \(\eta=0\) in the original domain. Using the same procedures employed above to obtain the series about \(\omega=0\), the series about \(\omega=1\) is defined as \[g(\omega) = \sum_{n=0}^\infty \hat{a}_n (\omega-1)^n \label{eq:g1}\] where the coefficients are given as
which may be considered a series solution to ([\[eq:Sakiadis\]](#eq:Sakiadis){reference-type="ref" reference="eq:Sakiadis"}) written in terms of the gauge function \(\left(e^{-C\eta/2}-1\right)\). In figure [\[fig:SakiadisErrorMapped\]](#fig:SakiadisErrorMapped){reference-type="ref" reference="fig:SakiadisErrorMapped"}b, the error associated with ([\[SakTransRec2\]](#SakTransRec2){reference-type="ref" reference="SakTransRec2"}) (dashed line) is shown to reduce to machine precision as additional terms are used in the series. Although it is difficult to identify an advantage in using ([\[SakTransRec2\]](#SakTransRec2){reference-type="ref" reference="SakTransRec2"}) over ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}), as each naturally perform better near their expansion point as they converge, a "best choice" can be chosen by examining the rapidity of convergence. This can be determined by considering the infinity norm of the error (over \(\omega\in[0,1]\)) versus truncation \(N\), shown in figure [\[fig:SakiadisNorm\]](#fig:SakiadisNorm){reference-type="ref" reference="fig:SakiadisNorm"}. In the figure, one can see that, for a fixed number of terms \(N\), the expansion about \(\omega=1\) (given by ([\[SakTransRec2\]](#SakTransRec2){reference-type="ref" reference="SakTransRec2"})) is more accurate than that about \(\omega=0\) (given by ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"})). For comparison, the approximant ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) is also shown in figure [\[fig:SakiadisNorm\]](#fig:SakiadisNorm){reference-type="ref" reference="fig:SakiadisNorm"} and outperforms both ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) and ([\[SakTransRec2\]](#SakTransRec2){reference-type="ref" reference="SakTransRec2"}). The cost in using ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) is computational, since either inverting an increasingly ill-conditioned Vandermonde matrix or using a Vandermonde inversion algorithm requires precision well-beyond double to avoid round-off error.
## Computation of \(C\), \(G\), \(\kappa\), and singularities of the Sakiadis function [\[sec:predictions\]]{#sec:predictions label="sec:predictions"}
Previous estimates for \(C\), \(G\), and \(\kappa\) have been found by quadrature, shooting, and approximant ([\[SakApproxForm\]](#SakApproxForm){reference-type="ref" reference="SakApproxForm"}) . Here, we improve upon those estimates and provide formulae for obtaining them to any desired precision, by considering the application of conditions ([\[eq:transcond\]](#eq:transcond){reference-type="ref" reference="eq:transcond"}) to the convergent solution ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}), i.e., \[\begin{aligned}
&g(1)=\sum_{n=0}^\infty \tilde{a}_n=0 \label{eq1}\\ &\dot{g}(1)=\sum_{n=1}^\infty n\tilde{a}_n=\frac{-2}{C} \label{eq2}\\ &\ddot{g}(1)=\sum_{n=2}^\infty n(n-1)\tilde{a}_n=\frac{2}{C}+\frac{4\kappa}{C^2}. \label{eq3} \end{aligned}\] Noting that ([\[eq1\]](#eq1){reference-type="ref" reference="eq1"}) and ([\[eq2\]](#eq2){reference-type="ref" reference="eq2"}) are functions of \(C\) and \(G\) only (see ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"})), we have 2 equations and 2 unknowns and one may employ any desired solver (here we use Newton's method). After finding \(C\) and \(G\), one may then use ([\[eq3\]](#eq3){reference-type="ref" reference="eq3"}) to explicitly compute \(\kappa\). In preparation for Newton's method, we rewrite ([\[eq1\]](#eq1){reference-type="ref" reference="eq1"}) and ([\[eq2\]](#eq2){reference-type="ref" reference="eq2"}) as \[\begin{aligned}
\phi_0(C, G) &= \sum_{n=0}^\infty \tilde{a}_n=0 \\ \phi_1(C, G) &= \frac{2}{C} + \sum_{n=1}^\infty n\tilde{a}_n=0, \end{aligned}\] such that the \(\nu^\mathrm{th}\) Newton iterate becomes
Using the method outlined above, the following improved estimates are obtained beyond double precision using only a few Newton iterations (if previous estimates are used as the initial guess) in ([\[eq:Newton\]](#eq:Newton){reference-type="ref" reference="eq:Newton"}) and a truncation of 50 terms in the series of ([\[eq:newtonseries\]](#eq:newtonseries){reference-type="ref" reference="eq:newtonseries"}): \[C=1.616125446804603717\dots\] \[G=-2.1313459240475714821 \dots\] Using the above values of \(C\) and \(G\), ([\[eq3\]](#eq3){reference-type="ref" reference="eq3"}) is used to compute \(\kappa\) as \[\kappa=\frac{C^2}{4}\left[G+\sum_{n=2}^\infty n^2\tilde{a}_n\right], \label{getkappa}\] where, taking 50 terms in ([\[getkappa\]](#getkappa){reference-type="ref" reference="getkappa"}), we obtain \[\kappa=-0.443748313368861 \dots\] Using the values of \(C\) and \(G\) given above, we may now explore the convergence of ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) with the aim of deducing the precise singularity locations sketched in figure [\[fig:SakiadisMapped\]](#fig:SakiadisMapped){reference-type="ref" reference="fig:SakiadisMapped"}. As shown in figure [\[fig:SakiadisROC\]](#fig:SakiadisROC){reference-type="ref" reference="fig:SakiadisROC"}, the radius of convergence of ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) can be deduced from the ratio test, and is approximately\[\lim_{N\to\infty}\left|\frac{\tilde{a}_N}{\tilde{a}_{N+1}}\right| = 2.66149513\dots\] Relating the above to the original variables of the Sakiadis problem via ([\[SakTransDefs\]](#SakTransDefs){reference-type="ref" reference="SakTransDefs"}), we have \[\begin{aligned}
\nonumber \lim_{N\to\infty}\left|\frac{\tilde{a}_N}{\tilde{a}_{N+1}}\right|&=\left|e^{-C\left(\eta_{s,r}\pm i \eta_{s,i}\right)/2}\right|\\ &=e^{-C\eta_{s,r}/2} \label{eq:ratio} \end{aligned}\] where \(\eta_{s,r}\) and \(\eta_{s,i}\) are the respective real and imaginary parts of the conjugate singularity pair closest to \(\eta=0\) in the original Sakiadis function. While ([\[eq:ratio\]](#eq:ratio){reference-type="ref" reference="eq:ratio"}) may be used to solve directly for \(\eta_{s,r}\), the imaginary part may be deduced by recognizing that ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) appears (for all terms investigated) to be an alternating series and thus the closest \(\omega\)-singularity is on the negative real line[^6] (indicated in figure [\[fig:SakiadisMapped\]](#fig:SakiadisMapped){reference-type="ref" reference="fig:SakiadisMapped"}b) such that \[\arg\left[e^{-C\left(\eta_{s,r}\pm i \eta_{s,i}\right)/2}\right]=\mp C\eta_{s,i}/2=\mp(2n-1)\pi,~~n=1,2,3\dots. \label{eq:arg}\] where, choosing \(n=1\) in ([\[eq:arg\]](#eq:arg){reference-type="ref" reference="eq:arg"}) leads to \(\eta_{s,i}\) values consistent with both the Padé analysis in and the radius of convergence of ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}) indicated in figure [\[fig:SakiadisUnMapped\]](#fig:SakiadisUnMapped){reference-type="ref" reference="fig:SakiadisUnMapped"} (when combined with \(\eta_{s,r}\) above). Hence we conjecture that \(n\) equals 1 in ([\[eq:arg\]](#eq:arg){reference-type="ref" reference="eq:arg"}) and, consequently, that the closest singularities to \(\eta=0\) in the Sakiadis problem are given, through use of ([\[eq:ratio\]](#eq:ratio){reference-type="ref" reference="eq:ratio"}) and ([\[eq:arg\]](#eq:arg){reference-type="ref" reference="eq:arg"}), by \[\begin{aligned}
\nonumber \eta_{s,\pm}&=\frac{-2}{C}\lim_{N\to\infty}\ln\left|\frac{\tilde{a}_N}{\tilde{a}_{N+1}}\right|\pm i\frac{2\pi}{C}\\&\approx-1.211401\pm3.887808i. \label{eq:SakSing} \end{aligned}\] It is worth mentioning that, although we are able to use the convergent series solution ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) to deduce the locations of the convergence-limiting singularities of the divergent series ([\[eq:SakSeries\]](#eq:SakSeries){reference-type="ref" reference="eq:SakSeries"}), knowledge of the singularity ([\[eq:SakSing\]](#eq:SakSing){reference-type="ref" reference="eq:SakSing"}) does not enter into the formulation of our solution at any point, in contrast with the modified Euler summation ([\[eq:Boyd\]](#eq:Boyd){reference-type="ref" reference="eq:Boyd"}) used for the Blasius problem. Consequently (and fortunately), the convergence of ([\[SakTransRec\]](#SakTransRec){reference-type="ref" reference="SakTransRec"}) does not rely on retaining "enough" digits of ([\[eq:SakSing\]](#eq:SakSing){reference-type="ref" reference="eq:SakSing"}). In section [\[sec:Meniscus\]](#sec:Meniscus){reference-type="ref" reference="sec:Meniscus"}, an exponential transformation similar to ([\[SakTransDefs\]](#SakTransDefs){reference-type="ref" reference="SakTransDefs"}) is used to obtain a convergent series from a divergent one. However, in contrast with the Sakiadis problem, the meniscus problem of section [\[sec:Meniscus\]](#sec:Meniscus){reference-type="ref" reference="sec:Meniscus"} has an exact solution that allows us to anticipate the location convergence-limiting singularity *a priori*. Still, we shall show that, as in the problem above, precise knowledge of this singularity is not required to obtain a convergent expansion.
# Meniscus at a flat wall [\[sec:Meniscus\]]{#sec:Meniscus label="sec:Meniscus"}
## Background and formulation [\[sec:MeniscusBackground\]]{#sec:MeniscusBackground label="sec:MeniscusBackground"}
To examine the nature of the exponential gauge function transformation used in section [\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}, we examine a second classical problem which has similar structural features. In particular, we consider the well-studied problem of the shape of a static meniscus rising above an infinite horizontal pool . The pool of liquid has density, \(\rho\), is subjected to gravity, \(g\), and is in contact with air of negligible density. A flat wall is placed vertically in the pool, and the liquid intersects the wall with a contact angle \(\theta\) (measured through the liquid) as shown in figure [\[fig:ErrorPlot\]](#fig:ErrorPlot){reference-type="ref" reference="fig:ErrorPlot"}b; for purposes of this study, we assume that \(\theta\) lies between \(0\) and \(\pi/2\). The location of the air--liquid interface, with surface tension \(\sigma\), is parameterized as \(y=h(x)\) where \(x\) is the horizontal distance from the wall, and \(y=0\) is the undisturbed location of the interface as \(x\to\infty\). For this configuration, the Young--Laplace equation couples with the hydrostatic field to yield the following dimensionless equations and boundary conditions :
In ([\[eq:secondorder\]](#eq:secondorder){reference-type="ref" reference="eq:secondorder"}), derivatives in \(\bar{x}\) are denoted with primes, and the over-bars denote dimensionless variables defined as \[\bar{h}=\frac{h}{L},~\bar{x}=\frac{x}{L},~ L=\sqrt{\frac{\sigma}{\rho g}}. \label{eq:dlvars}\] In ([\[eq:dlvars\]](#eq:dlvars){reference-type="ref" reference="eq:dlvars"}), the characteristic length scale \(L\) is the well-known capillary length. Multiplying both sides of ([\[eq:secondorderODE\]](#eq:secondorderODE){reference-type="ref" reference="eq:secondorderODE"}) by \(\bar{h}'\), integrating, and applying the boundary conditions ([\[eq:slopeboundarycondition\]](#eq:slopeboundarycondition){reference-type="ref" reference="eq:slopeboundarycondition"}) and ([\[eq:flatcondition\]](#eq:flatcondition){reference-type="ref" reference="eq:flatcondition"}), we obtain
where ([\[eq:heightatthewall\]](#eq:heightatthewall){reference-type="ref" reference="eq:heightatthewall"}) represents the height of the interface at the wall as function of the contact angle, \(\theta\in[0,\pi/2]\). Note that, in ([\[eq:firstorderOperator\]](#eq:firstorderOperator){reference-type="ref" reference="eq:firstorderOperator"}), \(\bar{h}'\to-\infty\) as \(\bar{h}\to\sqrt{2}\). The exact inverse solution \(\bar{x}(\bar{h})\) of ([\[eq:firstorderODE\]](#eq:firstorderODE){reference-type="ref" reference="eq:firstorderODE"}) is obtainable via variable separation and integration and is given as \[\bar{x}= \cosh^{-1}\frac{2}{\bar{h}}-\cosh^{-1}\sqrt{\frac{2}{1-\sin\theta}} + \sqrt{2+2\sin\theta}-\sqrt{4-{\bar{h} ^{2}}}. \label{eq:Batcher}\] In what follows, we obtain an analytic solution for \(\bar{h}(\bar{x})\) directly via series expansion. Our intention in doing so is not to replace equation ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}) in usage, but it is to elucidate structural similarities and provide insights to the Sakiadis series solution in section ([\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}). Equation ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}) is used in what follows to assess the accuracy of the series solutions provided in sections [\[sec:divergent\]](#sec:divergent){reference-type="ref" reference="sec:divergent"} and [\[sec:MeniscusMapped\]](#sec:MeniscusMapped){reference-type="ref" reference="sec:MeniscusMapped"}.
## Divergent power series solution [\[sec:divergent\]]{#sec:divergent label="sec:divergent"}
The standard power series solution of ([\[eq:firstorderODE\]](#eq:firstorderODE){reference-type="ref" reference="eq:firstorderODE"}) is found by assuming
In ([\[eq:meniscusform\]](#eq:meniscusform){reference-type="ref" reference="eq:meniscusform"}), the series is stated to converge within the region \(|\bar{x}|<|\bar{x}_s(\theta)|\) with \(\bar{x}_s(\theta)\) being the as-of-yet undetermined closest singularity to \(\bar{x}=0\). Without formal proof, we conjecture that \(\bar{x}_s(\theta)\) is the value that satisfies \(\bar{h}(\bar{x}_s)=\sqrt{2}\) for a given \(\theta\) in the exact solution ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}), causing \(\bar{h}'\to-\infty\) in ([\[eq:secondorderODE\]](#eq:secondorderODE){reference-type="ref" reference="eq:secondorderODE"}). The value of the conjectured limiting singularity \(\bar{x}_s(\theta)\) is known exactly in closed-form from the substitution of \(\bar{h}=\sqrt{2}\) into ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}) and is given as \[\bar{x}_s(\theta)= \cosh^{-1}\sqrt{2}-\cosh^{-1}\sqrt{\frac{2}{1-\sin\theta}} + \sqrt{2+2\sin\theta}-\sqrt{2}, \label{eq:BatcherSing}\] where we note that \(\bar{x}_s\leq0\) for \(\theta\in[0,\pi/2]\). We support this conjecture in what follows. Figure ([\[fig:Interface_45deg\]](#fig:Interface_45deg){reference-type="ref" reference="fig:Interface_45deg"}) shows \(N\)-term truncations of series ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}) (dashed curves), compared with the exact solution ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}) (\(\bullet\)'s) for \(\theta=\pi/4\). As expected the series ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}) agrees for small \(\bar{x}\) but ultimately diverges at a finite radius of convergence (indicated as a solid vertical line in the figure) and given by \(|\bar{x}_s(\pi/4)|\approx 0.3\) from ([\[eq:BatcherSing\]](#eq:BatcherSing){reference-type="ref" reference="eq:BatcherSing"}). Although we have not formally established that \(\bar{x}_s\) is the closest singularity to \(\bar{x}=0\) in ([\[eq:firstorderODE\]](#eq:firstorderODE){reference-type="ref" reference="eq:firstorderODE"}) and thus is responsible for divergence, evidence to support this conjecture is given by the Domb-Sykes plot in figure [\[fig:Domb-SkyesPlot_OriginalODE\]](#fig:Domb-SkyesPlot_OriginalODE){reference-type="ref" reference="fig:Domb-SkyesPlot_OriginalODE"}, where the magnitude of the ratios of coefficients of ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}) is shown to approach \(|\bar{x}_s(\pi/4)|\) as \(n\to\infty\) (i.e., \(1/n\to0\) in the figure). Although only shown in this section for \(\theta=\pi/4\), all permissible \(\theta\) values lead to similarly divergent series, limited by a radius of convergence of \(|\bar{x}_s(\theta)|\). In section [\[sec:MeniscusMapped\]](#sec:MeniscusMapped){reference-type="ref" reference="sec:MeniscusMapped"}, we apply the same type of transformation used in section [\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"} to overcome this convergence barrier for all contact angles.
## Asymptotic expansion, variable transform, and convergent power series solution [\[sec:MeniscusMapped\]]{#sec:MeniscusMapped label="sec:MeniscusMapped"}
Now that we have a solution to ([\[eq:firstorderOperator\]](#eq:firstorderOperator){reference-type="ref" reference="eq:firstorderOperator"}), given by ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}) that---although ultimately divergent beyond some positive \(\bar{x}\)-value---provides the correct interface shape close to the wall, the asymptotic behavior away from the wall is examined. To this end, the method of dominant balance is employed. To meet the constraint ([\[eq:flatcondition\]](#eq:flatcondition){reference-type="ref" reference="eq:flatcondition"}), it is assumed that
The asymptotic solution to ([\[eq:asympformODE\]](#eq:asympformODE){reference-type="ref" reference="eq:asympformODE"}) is
The same process is repeated to generate higher order correction to ([\[eq:asympform\]](#eq:asympform){reference-type="ref" reference="eq:asympform"}). We obtain: \[h\sim C_0e^{-x} + \frac{3}{16}C_0^{3}e^{-3\overline{x}} +\frac{3}{256}C_0^{5}e^{-5\overline{x}} + O(e^{-7x}),~~\bar{x}\to\infty. \label{eq:asympformFinal}\] The pattern of exponentials is thus evident in equation ([\[eq:asympformFinal\]](#eq:asympformFinal){reference-type="ref" reference="eq:asympformFinal"}). We note here that a similar exponential behavior is also observed in the Sakiadis problem previously and provides a linkage between the problems in approach and interpretation to follow. The asymptotic solution ([\[eq:asympformFinal\]](#eq:asympformFinal){reference-type="ref" reference="eq:asympformFinal"}) motivates us to transform ([\[eq:firstorderODE\]](#eq:firstorderODE){reference-type="ref" reference="eq:firstorderODE"}) to reflect the exponential pattern of ([\[eq:asympformFinal\]](#eq:asympformFinal){reference-type="ref" reference="eq:asympformFinal"}). This is achieved by transformations in both the independent and dependent variables, given respectively as
Substituting ([\[eq:transformations\]](#eq:transformations){reference-type="ref" reference="eq:transformations"}) into ([\[eq:firstorderODE\]](#eq:firstorderODE){reference-type="ref" reference="eq:firstorderODE"}) leads to the transformed ODE: \[\left[H+2U\dot{H}\right]\left[UH^{2}-2\right] =-H\sqrt{4-UH^{2}},~~ \label{eq:transformedODE}\] with transformed condition \[H(1)=\sqrt{2(1-\sin\theta)}, \label{eq:onecondition}\] where \(\dot{H}\) denotes the derivative with respect to \(U\). Although an exact explicit solution to ([\[eq:transformedODE\]](#eq:transformedODE){reference-type="ref" reference="eq:transformedODE"}) cannot be found as \(H(U)\), an exact implicit solution in \(H\) and \(U\) can be found by separating variables, integrating ([\[eq:transformedODE\]](#eq:transformedODE){reference-type="ref" reference="eq:transformedODE"}), and applying ([\[eq:onecondition\]](#eq:onecondition){reference-type="ref" reference="eq:onecondition"}) to arrive at \[\sqrt{2+2\sin\theta}-\sqrt{4-UH^2}=\ln\frac{H\left[1+\sqrt{1+\sin\theta}\right]}{\sqrt{1-\sin\theta}\left[2+\sqrt{4-UH^2}\right]}. \label{eq:impsoln}\] Equation ([\[eq:impsoln\]](#eq:impsoln){reference-type="ref" reference="eq:impsoln"}) is used to extract the condition \[H(0)=\frac{4\sqrt{1-\sin\theta }~ e^{-2+\sqrt{2+2\sin\theta}}}{\sqrt{1+\sin\theta}+\sqrt{2}}, \label{eq:zerocondition}\] which is needed for the series solution that follows. Using the same procedure as in sections ([\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}) and ([\[sec:divergent\]](#sec:divergent){reference-type="ref" reference="sec:divergent"}), the power series solution of ([\[eq:transformedODE\]](#eq:transformedODE){reference-type="ref" reference="eq:transformedODE"}) (with condition ([\[eq:zerocondition\]](#eq:zerocondition){reference-type="ref" reference="eq:zerocondition"})) is given by \[H=\sum_{n=0}^\infty A_{n} U^{n}, \label{eq:InitialTry}\] where,
which, by construction, is consistent with the asymptotic ordering ([\[eq:asympformFinal\]](#eq:asympformFinal){reference-type="ref" reference="eq:asympformFinal"}) as \(\bar{x}\to\infty\), and shows explicitly the exponential gauge function \(e^{-2\bar{x}}\). Figure [\[fig:Interface_45deg\]](#fig:Interface_45deg){reference-type="ref" reference="fig:Interface_45deg"} shows \(N\)-term truncations of ([\[eq:TransformedODESolu\]](#eq:TransformedODESolu){reference-type="ref" reference="eq:TransformedODESolu"}) (solid curves) compared with both the exact solution ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}) and divergent series ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}). The difference between ([\[eq:TransformedODESolu\]](#eq:TransformedODESolu){reference-type="ref" reference="eq:TransformedODESolu"}) and the exact solution is not discernible on the scale of the figure for any \(N\) shown, and it is noteworthy to mention that the radius of convergence of the original series ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}) has been exceeded. The absolute error between ([\[eq:TransformedODESolu\]](#eq:TransformedODESolu){reference-type="ref" reference="eq:TransformedODESolu"}) and the exact solution ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}) is shown in figure [\[fig:ErrorPlot\]](#fig:ErrorPlot){reference-type="ref" reference="fig:ErrorPlot"}a, which indicates convergence (as \(N\) increases) for a continuum of angles \(\theta\), as prescribed in figure [\[fig:ErrorPlot\]](#fig:ErrorPlot){reference-type="ref" reference="fig:ErrorPlot"}b. Figure [\[fig:ErrorPlot\]](#fig:ErrorPlot){reference-type="ref" reference="fig:ErrorPlot"} is generated for the smallest possible contact angle (\(\theta=0\)) and thus, by virtue of ([\[eq:firstorderODE\]](#eq:firstorderODE){reference-type="ref" reference="eq:firstorderODE"}) being an autonomous ODE, contains interface shapes for all contact angles as shifted semi-infinite domains; this is indicated in the figure. The maximum error occurs at the wall and is shown versus \(N\) for \(\theta=0\) in figure [\[fig:ErrorvsN\]](#fig:ErrorvsN){reference-type="ref" reference="fig:ErrorvsN"}. As done in ([\[sec:SakiadisMapped\]](#sec:SakiadisMapped){reference-type="ref" reference="sec:SakiadisMapped"}) for the Sakiadis problem, we now provide an explanation of why the series ([\[eq:TransformedODESolu\]](#eq:TransformedODESolu){reference-type="ref" reference="eq:TransformedODESolu"}) converges to the exact solution ([\[eq:Batcher\]](#eq:Batcher){reference-type="ref" reference="eq:Batcher"}) of the ODE ([\[eq:firstorderODE\]](#eq:firstorderODE){reference-type="ref" reference="eq:firstorderODE"}) describing a meniscus at a flat wall. The answer again lies in the mapping provided by the gauge function (here, ([\[eq:indtrans\]](#eq:indtrans){reference-type="ref" reference="eq:indtrans"})), as shown in figure [\[fig:MeniscusMapped\]](#fig:MeniscusMapped){reference-type="ref" reference="fig:MeniscusMapped"} where the complex \(\bar{x}\) and \(U\) planes are compared. In the \(\bar{x}\) plane of figure [\[fig:MeniscusMapped\]](#fig:MeniscusMapped){reference-type="ref" reference="fig:MeniscusMapped"}, circles of convergence for ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}) centered around \(\bar{x}\)=0 are drawn for various \(\theta\) values, based on the conjectured closest singularity (to \(\bar{x}\)=0), \(\bar{x}_s(\theta)\), using ([\[eq:BatcherSing\]](#eq:BatcherSing){reference-type="ref" reference="eq:BatcherSing"}). For \(\theta=0\), no circle is drawn, since \(\bar{x}_s(\theta)\) (which sets the radius) is 0. Note, that these circles of convergence each intersect the positive real line. In particular, for \(\theta=\pi/4\), this intersection occurs at the same location as the radius of convergence of series ([\[eq:divpowerseries\]](#eq:divpowerseries){reference-type="ref" reference="eq:divpowerseries"}) drawn in figure [\[fig:Interface_45deg\]](#fig:Interface_45deg){reference-type="ref" reference="fig:Interface_45deg"}; this correspondence holds for all other values of \(\theta\), as expected from Taylor's theorem. In the \(U\) plane of figure [\[fig:MeniscusMapped\]](#fig:MeniscusMapped){reference-type="ref" reference="fig:MeniscusMapped"}, the circle of convergence for ([\[eq:TransformedODESolu\]](#eq:TransformedODESolu){reference-type="ref" reference="eq:TransformedODESolu"}) is drawn, centered at \(U=0\) and with mapped radius of \(|e^{-2\bar{x}_s}|\); note that this circle now extends to the boundary of the physical domain for \(\theta=0\) and extends beyond it for \(\theta>0\), which explains why ([\[eq:TransformedODESolu\]](#eq:TransformedODESolu){reference-type="ref" reference="eq:TransformedODESolu"}) converges over the entire physical domain for all values of \(\theta\). The circles of convergence in figure [\[fig:MeniscusMapped\]](#fig:MeniscusMapped){reference-type="ref" reference="fig:MeniscusMapped"} are, of course, still *conjectured* because, in constructing the mapping in figure [\[fig:MeniscusMapped\]](#fig:MeniscusMapped){reference-type="ref" reference="fig:MeniscusMapped"}, we are assuming that the singularity \(\bar{x}_s(\theta)\) (given by ([\[eq:BatcherSing\]](#eq:BatcherSing){reference-type="ref" reference="eq:BatcherSing"})) is the closest singularity to \(\bar{x}=0\) and that no other singularities map to \(U\) singularities closer to \(U=0\). In addition to evidence given in section [\[sec:divergent\]](#sec:divergent){reference-type="ref" reference="sec:divergent"}, further evidence that supports this conjecture is provided in the Domb-Sykes plots in figure [\[fig:Domb-SkyesPlot_TransformedODE\]](#fig:Domb-SkyesPlot_TransformedODE){reference-type="ref" reference="fig:Domb-SkyesPlot_TransformedODE"}, where the radius of convergence deduced from this numerical ratio-test is consistent with the locations of the mapped singularities in figure [\[fig:MeniscusMapped\]](#fig:MeniscusMapped){reference-type="ref" reference="fig:MeniscusMapped"}. Finally, it is worth noting that---like in the Sakiadis problem---knowledge of these singularities is not required to implement the exponential gauge function transformation ([\[eq:transformations\]](#eq:transformations){reference-type="ref" reference="eq:transformations"}) that leads to a convergent series solution.
# Conclusions [\[sec:conclusions\]]{#sec:conclusions label="sec:conclusions"}
In this work, we provide convergent power series solutions to the Sakiadis boundary layer problem and the problem of a meniscus at a flat wall, by means of transforming the original ODEs in terms of variable substitutions that are motivated by the asymptotic expansions about \(\infty\). In both cases, the transformations map the dominant convergence-limiting singularities out of the physical domain; also in both cases, convergence-limiting singularities do not need to be known *a priori* but their locations are deduced nonetheless. For the Sakiadis problem, this provides---in the absence of a formal proof of exactness---a *conjectured* exact Taylor series representation of the solution over the full physical domain. That said, for both the Sakiadis and meniscus problems, the exponential gauge functions used handle singularities similarly to achieve these demonstrably convergent series solutions. Although the nature of nonlinear ODEs precludes general conclusions, our results indicate that asymptotic behaviors can be useful to motivate gauge functions to overcome power series divergence. Additionally, the approach used here supports a growing body of literature underscoring the use of power series solutions as a viable method for analytically solving nonlinear ODEs.
[^1]: There are several definitions of Euler transformations in the literature. The definitions and interpretations that align most with this work are found in
[^2]: In , only the singularity at \(-1.2114 + 3.8878i\) is reported. Revisiting the analysis, we find two things: (1) The expected conjugate \(-1.2114-3.8878i\) is also found in the roots of the denominator of the modified Padé and (2) The two additional digits of \(\eta_s\) reported in (but not here) are found to converge to different values as more terms are taken; a more precise value is given here in section [\[sec:predictions\]](#sec:predictions){reference-type="ref" reference="sec:predictions"}.
[^3]: To paraphrase Buzz Lightyear: "To \(\eta=\infty\) and beyond"
[^4]: Although here we claim this form of \(C\) and \(G\) dependence on \(\tilde{a}_n\) by inspection, this can be verified by making the substitutions \(\omega=G\tilde{\omega}/C\) and \(g=C\tilde{g}\) into ([\[SakTransformedODE\]](#SakTransformedODE){reference-type="ref" reference="SakTransformedODE"}) and noting the \(C\) and \(G\) dependence vanishes.
[^5]: In cases where the parameter dependence is more complicated, this can always be done recursively by directly partially differentiating the recurrence relation.
[^6]: This follows from a variable substitution \(\tilde{\omega}=-\omega\) in ([\[eq:omega\]](#eq:omega){reference-type="ref" reference="eq:omega"}), application of Pringsheim's theorem to the resulting series, and then mapping back to \(\omega\). | {'timestamp': '2022-06-06T02:12:56', 'yymm': '2206', 'arxiv_id': '2206.01536', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01536'} |
# Introduction
communications systems are conceptualized, designed, and optimized under the assumption that the propagation channels are random and uncontrollable. However, the emerging idea of a reconfigurable intelligent surface (RIS), also known as an intelligent reflecting surface, has challenged this fundamental assumption. An RIS is a novel concept in wireless communications where existing artificial structures such as walls and ceilings of buildings will be equipped with many tightly packed subwavelength-sized reflecting meta-surfaces. The overall RIS is planar, while each of the metasurfaces in the RIS is software-controlled and are designed to perform a desired transformation on the incoming signal, thereby providing some control over the propagation environment. This control over the wireless propagation environment has been proposed as a means to aid wireless communication systems. It has been shown that with perfect channel state information, the presence of RISs can provide significant gains in energy and bandwidth efficiency when the transmit beamformer and the reflection coefficients at the RIS are jointly optimized. Although the initial applications of RIS to wireless systems were limited to RIS-aided communication system designs, it has recently gained attention as a means to improve localization accuracy. The basic idea is to treat RISs as virtual anchors if their positions are known, which is a reasonable assumption for stationary RISs. This assumption is conceptually similar to assuming perfect knowledge of the locations of actual anchors, such as the macro base stations (BSs). Hence, there is the potential of measuring the times of arrival and other geometric channel parameters valuable for positioning with respect to different RISs. However, in order to fully harness the power of RIS-aided localization, we must first understand its fundamental limits. These fundamental limits are unknown for the general case in the far-field, which is the focus of this paper. Thus, in the paper, we derive the Bayesian CRLB and examine its structural properties which leads to key insights into RIS-aided localization in the far-field case including the impact of nuisance parameters and the effects of RIS location uncertainty on localization performance.
## Related Works
The following three research directions are of interest to this paper: i) localization using RISs, ii) localization with large antenna arrays, and iii) Bayesian limits of localization networks and effects of anchor uncertainty. The relevant prior work from these three directions is discussed next.
### Localization using RISs
It has been shown that due to a large number of geometric channel parameters in RIS-aided wireless systems, the position error bound (PEB) and the orientation error bound (OEB) of a user equipment (UE) can be significantly reduced. A challenge with exploiting multiple RISs in localization is identifying which received signals are associated with specific RISs. In, Hadamard matrices have been proposed as a possible solution to this problem as it allows the receiver to extract channel parameters associated with each RIS. In, the focus is on the setup where an RIS is attached to each intended receiver and a joint phase design scheme is used to separate the channel parameters associated with each user resulting in submeter UE localization. In, the focus is on using an RIS to enhance the LOS link through coherent combining, thus improving both 3D localization and synchronization. In, an RIS is employed to multi-input-multi-output (MIMO) radar to provide the location of multiple targets. RIS-aided localization without a controlling BS has also been considered. In, two-step positioning is achieved without a controlling BS by employing a single receive RF chain at each RIS. In, backscatter modulation is used to enable localization in the absence of a controlling BS. A key limitation of these works is their application-specific CRLB analyzes in which the effect of each scenario is abstracted such that it can be approximated through suitable modifications of the existing CRLB results. As a result, the general structure of the fisher information matrix (FIM) resulting from deploying multiple RISs has not been rigorously investigated. Moreover, in those works, the impact of RIS perturbation/placement uncertainty and the effect of nuisance parameters have not been analyzed. Hence, the primary objective of this paper is to provide a rigorous Bayesian analysis of the FIM. For completeness, we also discuss the relevant literature on RIS-aided localization in the near-field. When the RISs are large enough such that the receivers are considered to be in their near-field, additional information provided by spherical wavefront can be used for localization as shown in. More specifically, a single RIS is attached to a receiver and used to localize a transmitter in, while multiple RISs are used to provide continuous positioning capability by improving the near-field NLOS accuracy in. In, backscatter modulation is used to empower each RIS element in a single RIS to act as virtual anchors for the time of arrival-based localization. In, the RIS-enhanced bounds for \(3\)D localization in the near-field are provided for the uplink of a system operating both synchronously and asynchronously
### Localization with Large Antenna Arrays
The reflection coefficients of the RISs can be designed such that the channel parameters associated with distinct paths can be separated and used as information for localization (see ), which is conceptually similar to localization enabled by the spatial and temporal resolution offered by large antenna systems operating with large bandwidths. Hence, as in massive MIMO-aided localization where the resulting FIM can be diagonalized due to the presence of a large number of antennas (e.g., see ), the FIM for RIS-localization can also be diagonalized, *albeit* for a different reason (unitary correlation matrix, as will be discussed in detail shortly). As a consequence, the existing results on localization with large arrays have the potential of providing useful insights into RIS-aided localization. With this in mind, we now summarize the most relevant literature from this research direction. In, source localization is considered by collecting and processing time of arrival (ToA) and angle of arrival (AoA) estimates at various distributed BSs. While the ToA estimates are used to restrict the set of possible source positions to a convex set, the AoA estimates are exploited to provide an estimate of the source position. In, \(2\)D position and orientation bounds are derived along with expectation-maximization-based estimation algorithms that achieve these bounds. In, these bounds are used to show that under certain conditions, NLOS components provide more information about position and orientation. These bounds are generalized to the \(3\)D scenario, and various scaling laws are provided for both the uplink and downlink in. More limits are provided for localization using measurements on the uplink of a massive MIMO system in. These additional limits indicate that the CRLB is unique in the limit of the number of receive antennas because each possible transmit position leads to distinct observations at the BS. The case of downlink UE positioning with a single antenna receiver is considered in. In, a single anchor is used to estimate UE trajectory, and the effect of I/Q imbalance on single anchor positioning is considered in.
### Bayesian Limits of Localization Networks and Effects of Anchor Uncertainty
In our work we are concerned with the fundamental limits of RIS-enabled localization. A major factor in this analysis is the uncertainty in the exact position/orientation of the RIS. Bayesian approaches to localization can be used to include this uncertainty. Thus, we briefly review relevant work in this area. The Bayesian philosophy to estimation presented in has been applied to the parameter estimation problem in localization, (see for a small subset). In, the received waveforms from various anchors are processed, combined with prior information about UE position, and subsequently used to provide localization bounds. This setting is extended to the case of multiple UEs that communicate with each other in, resulting in cooperative Bayesian localization bounds. In, Bayesian limits are presented for network localization which utilizes both position-related parameters and inertial measurement units. In, the possibility of improving localization and tracking systems by exploiting prior UE information is investigated. Authors in develop the concept of soft information (SI) for localization. Instead of providing hard decisions on NLOS/LOS and the position of a UE, SI quantifies these decisions through probability distributions. While the Bayesian philosophy has been used to generate a posterior distribution that provides localization estimates in the presence of anchor uncertainty, a deterministic approach to parameter estimation in the presence of anchor uncertainty has also been considered.
### Bayesian FIM for Channel Parameters along each RIS-aided Path
We derive the FIM for the RIS-related channel parameters and show that the FIM can be decomposed into: i) information provided by the receiver, ii) information provided by the transmitter, and iii) information provided by the RIS components. We show that the information provided by the RIS can be further decomposed into a correlation matrix and an information matrix representing the gains due to the RIS. This decomposition shows that the structure of the FIM can be significantly controlled through the RIS correlation matrix and this control allows us to investigate the impact of both nuisance parameters and anchor uncertainty on the localization performance.
### Bayesian Equivalent Fisher Information Matrix (EFIM) for Geometric Channel Parameters Provided by the RIS
We quantify the information loss associated with the geometric channel parameters along each RIS path due to the unknown nuisance parameters (channel path gains). This quantification is achieved by deriving the Bayesian EFIM for the geometric channel parameters. We show that the resulting information loss has a specific structure. With this computed information loss, we show that when no prior information about the nuisance parameters is available, the corresponding submatrix in the EFIM related to the RIS geometric parameters is a zero matrix. As a result of the EFIM being a zero matrix, any parametric channel estimator must provide the complex path gains before estimating the RIS-related geometric channel parameters. Because localization is dependent on the accurate estimation of the RIS-related geometric channel parameters, we note that localization is not possible without prior information on the nuisance parameters. This result is in contrast to localization with large antenna arrays. All the information is not lost in localization with large antenna arrays due to a lack of knowledge about the nuisance parameters, and localization is still feasible. We further show that for a RIS deployed as a passive uniform rectangular array (URA), when the paths are separable, the Fisher information provided by the angle of incidence (both azimuth and elevation) is redundant in the EFIM for the geometric channel parameters and thus its estimation is not necessary.
### Bayesian FIM/EFIM for UE Position and Orientation
Through a bijective transformation, the FIM of UE position and orientation is obtained from the EFIM of the geometric channel parameters. When paths are separable, the FIM of the UE position and orientation is a sum of the FIM from each of the RIS paths. While any prior information about the UE appears as an additive term in the EFIM, the prior information about the RIS appears in a less simplistic manner. Finally through Monte-Carlo simulations, we study the effect of number of RISs and number of RIS elements on the localization performance as well as the consequence of having a set of RISs with perturbed positions/orientations.
# System Model
We consider the downlink of an RIS-assisted single-cell MIMO system consisting of a single BS with \(N_T\) antennas, a UE of interest with \(N_R\) antennas, and \({M}_1\) distinct RISs. The \({m}^{th}\) RIS is assumed to contain \(N_L^{[{m}]}\) reflecting elements where \({m} \in \mathcal{M}_1 = \{1, 2, \ldots, {M}_1\}\). We further assume orthogonal frequency division multiplexing (OFDM) for this transmission. The BS has an arbitrary but known array geometry with its centroid located at \(\mathbf{p}_{BS} \in \mathbb{R}^3\). The UE is defined by its position \(\mathbf{p}_{} \in \mathbb{R}^3\), orientation \((\theta_0, \phi_0)\), and an arbitrary but known array geometry. We use notations \(\theta\) and \(\phi\), respectively, with appropriate subscripts and superscripts for all the elevation and azimuth angles. The set of RISs is also defined by their positions \(\mathbf{p}^{[{m}]} \in \mathbb{R}^3\) and orientation angles \((\theta_0^{[{m}]}, \phi_0^{[{m}]})\), for \({m} \in \mathcal{M}_1\). The geometry of each RIS is known but arbitrary. There are \({M} \geq 3\) paths between the BS and the UE, where the first path \({m} = 0\) is the LOS path, a set of paths \({m} \in \mathcal{M}_1\) are the virtual LOS paths created by \({M}_1\) distinct RISs, and there is a set of non-RIS NLOS paths \(\mathcal{M}_2 = \{{M}_1+1, {M}_1+2, \cdots, {M}_1 + {M}_2 \}\), where \({M} = {M}_1 + {M}_2+1\). The non-RIS NLOS paths (created either by scatterers or reflectors) are usually much weaker compared to the non-RIS LOS paths as well as the virtual LOS paths created by the RISs. Therefore, the non-RIS NLOS paths will not be included in the analysis for notational simplicity. Further, we partition the set of RISs \(\mathcal{M}_1\) into a set with perfectly known position and orientation \(\mathcal{M}_1^{a}\) and a set with perturbed position and orientation \(\mathcal{M}_1^{b}\). These sets are mutually exclusive, i.e. \(|\mathcal{M}_1| = |\mathcal{M}_1^{a}| + |\mathcal{M}_1^{b}|\). The number of parameters that needs to be estimated depends on these mutually exclusive sets. In general, the number of unknown parameters related to positioning is \(6 + 6|\mathcal{M}_1^{b}|\). Note that the constant term is due to the unknown \(3\)D orientation and the unknown \(3\)D position of the UE.
## Far-Field Channel Model
All paths are described in part by their angle of departure (AoD), angle of arrival (AoA), and time of arrival (ToA) as specified by \((\theta_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{u}}}^{[{m}]})\), \((\theta_{\mathrm{r},\mathrm{u}}^{[{m}]}, \phi_{\mathrm{r},\mathrm{u}}^{[{m}]})\), and \(\tau^{[{m}]}\), respectively. The array vector at the transmitter and receiver is specified by \[\label{equ:channel_model_array_response} \begin{aligned} \mathbf{a}_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}\left(\theta_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}\right) \triangleq e^{-j \Delta_{\mathrm{\mathrm{t}, {\mathrm{u}}}}^{\mathrm{T}} \mathbf{k}\left(\theta_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}\right)} \in \mathbb{C}^{N_{T}}, \; \; \mathbf{a}_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\left(\theta_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\right) \triangleq e^{-j \Delta_{\mathrm{r},\mathrm{u}}^{\mathrm{T}} \mathbf{k}\left(\theta_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\right)} \in \mathbb{C}^{N_{R}}, \end{aligned}\] respectively, where \(\mathbf{k}(\theta, \phi)=\frac{2 \pi}{\lambda}[\cos \phi \sin \theta, \sin \phi \sin \theta, \cos \theta]^{\mathrm{T}}\) is the wavenumber vector, \(\lambda\) is the wavelength, \(\mathbf{\Delta}_{\mathrm{r},\mathrm{u}} \triangleq\left[\mathbf{u}_{\mathrm{r},\mathrm{u}, 1}, \mathbf{u}_{\mathrm{r},\mathrm{u}, 2}, \ldots, \mathbf{u}_{\mathrm{r},\mathrm{u}, N_{R}}\right], \mathbf{u}_{\mathrm{r},\mathrm{u}, n} \triangleq\left[x_{\mathrm{r},\mathrm{u}, n}, y_{\mathrm{r},\mathrm{u}, n}, z_{\mathrm{r},\mathrm{u}, n}\right]^{\mathrm{T}}\) is a vector of Cartesian coordinates of the \(n^{\text {th }}\) receiver element, and \(N_{R}\) is the number of receiving antennas. Similarly, parameters \(N_{T}, \mathbf{\Delta}_{\mathrm{r},\mathrm{u}}\) and \(\mathbf{u}_{\mathrm{\mathrm{t}, {\mathrm{u}}}, n}\) are defined for the transmit vector. The array response due to the AoR (angle of reflection) and AoI (angle of incidence) at the \({m}^\text{th}\) RIS can be written as \[\label{equ:channel_model_array_response_at_ris} \begin{aligned} \begin{aligned} &\mathbf{a}_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}\left(\theta_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}\right) \triangleq e^{-j \Delta_{\mathrm{l,m}}^{\mathrm{T}} \mathbf{k}\left(\theta_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}\right)} \in \mathbb{C}^{{{N_{L}^{[{m}]}}}}, \; \; \mathbf{a}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\left(\theta_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}, \phi_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\right) \triangleq e^{-j \Delta_{\mathrm{l,m}}^{\mathrm{T}} \mathbf{k}\left(\theta_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}, \phi_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\right)} \in \mathbb{C}^{{{N_{L}^{[{m}]}}}}, \end{aligned} \end{aligned}\] with \(\mathbf{\Delta}_{\mathrm{l,m}} \triangleq\left[\mathbf{u}_{\mathrm{l,m}, 1}, \mathbf{u}_{\mathrm{l,m}, 2}, \ldots, \mathbf{u}_{\mathrm{l, m}, N_{\mathrm{L}}^{[{m}]}}\right], \text{ where } \mathbf{u}_{\mathrm{l,m}, n} \triangleq\left[x_{\mathrm{l,m}, n}, y_{\mathrm{l,m}, n}, z_{\mathrm{l,m}, n}\right]^{\mathrm{T}}\) is a vector of Cartesian coordinates of the \(n^{\text {th }}\) RIS element. We also define \(\mathbf{X}_{\mathrm{l,m}} = \operatorname{diag}([x_{\mathrm{l,m,1}}, \cdots, x_{\mathrm{l,m,N_{\mathrm{L}}^{[{m}]}}}]^{\mathrm{T}})\). The definitions of \(\mathbf{Y}_{\mathrm{l,m}}\) and \(\mathbf{Z}_{\mathrm{l,m}}\) are similar. The channel at the \(n^{th}\) subcarrier during the \(t^{th}\) OFDM symbol is written as \[\label{equ:channel_model_compositie_channel} \begin{aligned} \mathbf{H}_{t}[n]=\sum_{m=0}^{M} \frac{\beta^{[{m}]}}{\sqrt{\rho^{[{m}]}}} \mathbf{H}^{[{m}]}_{t}[n] e^{\frac{-j 2 \pi n \tau^{[{m}]}}{N T_{S}}} \in \mathbb{C}^{N_{R} \times N_{T}}, \end{aligned}\] where \(\beta^{[{m}]}\) is the complex channel gain, \(\sqrt{\rho^{[{m}]}}\) is the pathloss of the \({m}^{th}\) path and \[\label{equ:channel_model_cases_channel} \begin{aligned} &\mathbf{H}^{[{m}]}_{t}[n] = \begin{cases} \mathbf{{a}}_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\left(\theta_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\right) {\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}\left(\theta_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}\right), \; \; \; {m} \in \{ \{0\} \cup \mathcal{M}_2\},\\ \mathbf{a}_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\left(\theta_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\right) {\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}\left(\theta_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}\right) \mathbf{\Omega}_{t}^{[{m}]}[n] \mathbf{a}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\left(\theta_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}, \phi_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\right){\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}\left(\theta_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}, \phi_{\mathrm{t}, {\mathrm{u}}}^{[{m}]}\right), \; \; {m} \in\mathcal{M}_1.\\ \end{cases} \\ \end{aligned} \end{equation \vspace{-10mm} \subsection{Transmit Processing} We consider the transmission of \(T\) OFDM symbols each containing \(N\) OFDM subcarriers. The BS precodes a vector of communication symbols \(\mathbf{x}^{}[n]=\left[x_{1}[n], \ldots, x_{N_{\mathrm{B}}}[n]\right]^{\mathrm{T}} \in \mathbb{C}^{N_{\mathrm{B}}}\) at the subcarrier level with a directional precoding matrix \(\mathbf{F} \in \mathbb{C}^{N_{T} \times N_{\mathrm{B}}} \). After precoding, the symbols are modulated with an \(N-\)point inverse fast Fourier transform (IFFT). A cyclic prefix of sufficient length \(N_{cp}\) is added to the transformed symbol. In the time domain, this cyclic prefix has length \(T_{cp} = N_{cp}T_s\) where \(T_s = 1/B\) represents the sampling period. The directional beamforming matrix is defined as \(\mathbf{F} \triangleq\left[\mathbf{f}_{1}, \mathbf{f}_{2}, \ldots \mathbf{f}_{N_{\mathrm{B}}}\right]\) where \(\mathbf{f}_{\ell}=\frac{1}{\sqrt{N_{\mathrm{B}}}} \mathbf{a}_{\mathrm{t}, b}\left(\theta_{\mathrm{t}, b}^{[l]}, \phi_{\mathrm{t}, b}^{[l]}\right), \quad 1 \leq \ell \leq N_{\mathrm{B}},\) is the beam pointing in the direction \(\left(\theta_{\ell}, \phi_{\ell}\right)\) and has the same representation as (\ref{equ:channel_model_array_response}). In order to ensure a power constraint, we set \(\operatorname{Tr}\left(\mathbf{F}^{\mathrm{H}} \mathbf{F}\right)=\) 1, and \(\mathbb{E}\left\{\mathbf{x}[n] \mathbf{x}^{\mathrm{H}}[n]\right\}=\mathbf{I}_{N_{\mathrm{B}}}\), where \(\operatorname{Tr}(\cdot)\) denotes the matrix trace and \(\mathbf{I}_{N_{\mathrm{B}}}\) is the \(N_{\mathrm{B}}\)-dimensional identity matrix. \subsection{Far-Field RIS Reflection Control} The reflection coefficients of the \({m}^{th}\) RIS during the \(t^{th}\) OFDM symbol can be decomposed into \begin{equation} \label{equ:far_field_rf_control} \begin{aligned} \mathbf{\Omega}_{t}^{[{m}]} = \gamma_{t}^{[{m}]} \mathbf{\Gamma}^{[{m}]}, \end{aligned}\] where \(\gamma_{t}^{[{m}]}\) and \(\mathbf{\Gamma}^{[{m}]}\) are termed fast and slow varying RIS coefficients respectively because \(\gamma_{t}^{[{m}]}\) varies across the \(T\) OFDM symbols, \(\mathbf{\Gamma}^{[{m}]}\) is constant across those OFDM symbols. For continuous RIS coefficients, this decomposition is trivially obtained while maintaining \(T = M_1\). However, for discrete RIS coefficients, the number of OFDM symbols must be strictly greater than \(M_1\). After the removal of the cyclic prefix and the application of an \(N\)-point fast Fourier transform (FFT), the received signal at the \(n^{th}\) subcarrier during the \(t^{th}\) OFDM symbol is \[\label{equ:receive_processing} \begin{aligned} \mathbf{r}_{t}[n] = \mathbf{H}_{t}[n] \mathbf{F} \mathbf{x}[n] + \mathbf{n}_{t}[n], \end{aligned}\] where \(\mathbf{n}_{t}[n] \sim \mathcal{C}\mathcal{N}(0,N_0)\) is the Fourier transform of the thermal noise local to the UE's antenna array at the \(n^{th}\) subcarrier during the \(t^{th}\) OFDM symbol and \(\mathbf{x}[n]\) are pilots transmitted.
## Far-Field Receive Processing
As stated above already, we assume that the only non-RIS path is a single LOS path, hence the signals received at the \(t^{th}\) OFDM symbol can be written as \[\label{equ:far_field_receive_processing} \begin{aligned} \mathbf{r}_{t}[n] &= \frac{\beta^{[\mathrm{0}]}}{\sqrt{\rho^{[\mathrm{0}]}}} \mathbf{H}^{[\mathrm{0}]}_{t}[n] e^{\frac{-j 2 \pi n \tau^{[\mathrm{0}]}}{N T_{S}}} \mathbf{F} \mathbf{x}[n] + \sum_{m = 1}^{{M}_1} \frac{\beta^{[{m}]}}{\sqrt{\rho^{[{m}]}}} \mathbf{H}^{[{m}]}_{t}[n] e^{\frac{-j 2 \pi n \tau^{[{m}]}}{N T_{S}}} \mathbf{F} \mathbf{x}[n] + \mathbf{n}_{t}[n]. \end{aligned}\] The first term corresponds to the LOS path and the second term corresponds to the RIS paths. To facilitate any subsequent derivations, we also write the received signal as \[\label{equ:far_field_receive_processing_2} \begin{aligned} \mathbf{r}_{t}[n] = \bm{\mu}^{}_{t}[n] + \mathbf{n}_{t}[n], \quad t=1,2, \ldots, T, \quad n=1,2, \ldots, N, \end{aligned}\] where \(\bm{\mu}^{}_{t}[n]\) is the noise-free part of \(\mathbf{r}_{t}[n]\). Based on the received signal in ([\[equ:far_field_receive_processing_2\]](#equ:far_field_receive_processing_2){reference-type="ref" reference="equ:far_field_receive_processing_2"}), the vectors of the unknown channel parameters related to the RIS paths are defined as \[\begin{aligned} \begin{array}{llll} \bm{\theta}_{\mathrm{r},\mathrm{u}} \triangleq\left[\theta_{\mathrm{r}, {\mathrm{u}}}^{[1]}, \ldots, \theta_{\mathrm{r}, {\mathrm{u}}}^{[{M_1}]}\right]^{\mathrm{T}}, & \bm{\phi}_{\mathrm{r},\mathrm{u}} \triangleq\left[\phi_{\mathrm{r}, {\mathrm{u}}}^{[1]}, \ldots, \phi_{\mathrm{r}, {\mathrm{u}}}^{[{M_1}]}\right]^{\mathrm{T}}, \bm{\theta}_{\mathrm{t},\mathrm{l}} \triangleq\left[\theta_{\mathrm{t}, {\mathrm{l}}}^{[1]}, \ldots, \theta_{\mathrm{t}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{T}}, & \bm{\phi}_{\mathrm{t},\mathrm{l}} \triangleq\left[\phi_{\mathrm{t}, {\mathrm{l}}}^{[1]}, \ldots, \phi_{\mathrm{t}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{T}}, \\ \bm{\theta}_{\mathrm{r},\mathrm{l}} \triangleq\left[\theta_{\mathrm{r}, {\mathrm{l}}}^{[1]}, \ldots, \theta_{\mathrm{r}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{r}}, & \bm{\phi}_{\mathrm{r},\mathrm{l}} \triangleq\left[\phi_{\mathrm{r}, {\mathrm{l}}}^{[1]}, \ldots, \phi_{\mathrm{r}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{r}}, \bm{\theta}_{\mathrm{t},u} \triangleq\left[\theta_{\mathrm{t}, {\mathrm{u}}}^{[1]},\ldots, \theta_{\mathrm{t}, {\mathrm{u}}}^{[{M_1}]}\right]^{\mathrm{T}}, & \bm{\phi}_{\mathrm{t},u} \triangleq\left[\phi_{\mathrm{t}, {\mathrm{u}}}^{[1]}, \ldots, \phi_{\mathrm{t}, {\mathrm{u}}}^{[{M_1}]}\right]^{\mathrm{T}}, \\ \bm{\beta} \triangleq\left[\beta^{[1]}, \ldots, \beta^{[{M_1}]}\right]^{\mathrm{T}}, & \bm{\tau}_{} \triangleq\left[\tau^{[1]}_{}, \ldots, \tau^{[{M_1}]}_{}\right]^{\mathrm{T}}. \end{array} \end{aligned}\] Hence, the unknown channel parameters can be represented by the vector \[\label{equ:FIM_parameter_vector} \begin{aligned} \bm{\eta} \triangleq\left[\bm{\theta}_{\mathrm{r},\mathrm{u}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{r},\mathrm{u}} ^{\mathrm{T}}, \bm{\theta}_{\mathrm{t},\mathrm{l}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{t},\mathrm{l}}^{\mathrm{T}}, \bm{\theta}_{\mathrm{r},\mathrm{l}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{r},\mathrm{l}}^{\mathrm{T}}, \bm{\theta}_{\mathrm{t},\mathrm{u}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{t},\mathrm{u}}^{\mathrm{T}}, \bm{\tau}^{\mathrm{T}}_{}, \bm{\beta}_{\mathrm{R}}^{\mathrm{T}}, \bm{\beta}_{\mathrm{I}}^{\mathrm{T}}\right]^{\mathrm{T}}, \end{aligned}\] where \(\bm{\beta}_{\mathrm{R}} \triangleq \Re\{\bm{\beta}\}\), and \(\bm{\beta}_{\mathrm{I}} \triangleq \Im\{\bm{\beta}\}\) are the real and imaginary parts of \(\bm{\beta}\), respectively. Also, the unknown channel parameters related to the LOS path are written as \(\begin{array}{llll} \bm{\psi}_{} \triangleq\left[\theta_{\mathrm{r}, {\mathrm{u}}}^{[0]}, \phi_{\mathrm{r}, {\mathrm{u}}}^{[0]}, \theta_{\mathrm{t}, {\mathrm{u}}}^{[0]}, \phi_{\mathrm{t}, {\mathrm{u}}}^{[0]}, \tau^{[\mathrm{0}]}_{}, {\beta}_{\mathrm{R}}^{\mathrm{[0]}}, {\beta}_{\mathrm{I}}^{\mathrm{[0]}} \right]^{\mathrm{T}}. \end{array}\) Finally, the unknown channel parameters related to the LOS plus RIS paths can be written as \(\bm{\zeta}_{} \triangleq\left[ \bm{\psi}_{}^{\mathrm{T}}, \bm{\eta}_{}^{\mathrm{T}} \right]^{\mathrm{T}}.\)
# Fisher Information for RIS Paths {#section:Fisher_Information_for_RIS_Paths}
We define the geometric channel parameters \(\bm{\eta}_{1} \triangleq\left[\bm{\theta}_{\mathrm{r},\mathrm{u}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{r},\mathrm{u}} ^{\mathrm{T}}, \bm{\theta}_{\mathrm{t},\mathrm{l}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{t},\mathrm{l}}^{\mathrm{T}}, \bm{\theta}_{\mathrm{r},\mathrm{l}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{r},\mathrm{l}}^{\mathrm{T}}, \bm{\theta}_{\mathrm{t},\mathrm{u}}^{\mathrm{T}}, \bm{\phi}_{\mathrm{t},\mathrm{u}}^{\mathrm{T}}, \bm{\tau}^{\mathrm{T}}_{}\right]^{\mathrm{T}}\) and the nuisance parameter as \(\bm{\eta}_{2} \triangleq\left[\bm{\beta}_{\mathrm{R}}^{\mathrm{T}}, \bm{\beta}_{\mathrm{I}}^{\mathrm{T}} \right]^{\mathrm{T}}\). To derive the FIM of \(\bm{\eta}_{}\), we define the PDF as \[\label{equ:pdf_joint_channel} \begin{aligned} \chi(\mathbf{r}_{t}[n]; \bm{\eta}_{} ) = \chi(\mathbf{r}_{t}[n] | \bm{\eta}_{} ) \chi(\bm{\eta}_{} ), \end{aligned}\] where \(\chi(\bm{\eta}_{} ) = \chi(\bm{\eta}_{1}) \chi( \bm{\eta}_{2})\). The FIM of the channel parameters due to the observation \(\mathbf{r}\) is an \(11 {M}_1 \times 11 {M}_1\) matrix which can be viewed as a collection of \({M}_1 \times {M}_1\) submatrices \[\label{equ:FIM_parameter_matrix} \begin{aligned} \mathbf{J}_{\eta}^{\mathrm{D}} \triangleq\left[\begin{array}{cccc} \mathbf{J}_{\bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{r},\mathrm{u}}} & \mathbf{J}_{\bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\phi}_{\mathrm{r},\mathrm{u}}} & \cdots & \mathbf{J}_{\bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{I}}} \\ \mathbf{J}_{\bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\phi}_{\mathrm{r},\mathrm{u}}}^{\mathrm{T}} & \ddots & \cdots & \vdots \\ \vdots & \cdots & \ddots & \vdots \\ \mathbf{J}_{\bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{I}}}^{\mathrm{T}} & \cdots & \cdots & \mathbf{J}_{\bm{\beta}_{\mathrm{I}} \bm{\beta}_{\mathrm{I}}} \end{array}\right] \end{aligned}\] in which \(\mathbf{J}_{\bm{\eta}_{{\mathbf{v}_1}} \bm{\eta}_{{\mathbf{v}_2}}} \triangleq \frac{2}{\sigma^2} \sum_{n = 1}^{N} \sum_{t=1}^{{T}} \Re\left\{\frac{\partial \bm{{\mu}^{}_{t}[n]}^{\mathrm{H}}}{\partial {\bm{\eta}_{}}_{\mathbf{v}_1 }} \frac{\partial \bm{{\mu}^{}_{t}[n]}^{{}}}{\partial {\bm{\eta}_{}}_{\mathbf{v}_2 }}\right\}\), where \(\bm{\eta}_{\mathbf{v}_1}\), \(\bm{\eta}_{\mathbf{v}_2}\) are both dummy variables that stand for the parameters of interest, and \(\sigma^2\) is the signal-to-noise-ratio which incorporates the pathloss and composite noise power. To incorporate any prior knowledge about the unknown channel parameters, the Bayesian FIM is also analyzed. The Bayesian FIM of the channel parameters \(\bm{\eta}_{}\) is also an \(11{M}_1 \times 11 {M}_1\) matrix which can also be viewed as several \({M}_1 \times {M}_1\) submatrices such that its entries are written as \(\Tilde{\mathbf{J}}_{\mathbf{v}_1 \mathbf{v}_2 } = {\mathbf{J}}_{\mathbf{v}_1 \mathbf{v}_2 } + {\mathbf{J}}_{\mathbf{v}_1 \mathbf{v}_2 }^{\mathrm{P}},\) where \(\mathbf{v}_{1}, \mathbf{v}_{2} \in \bm{\eta}\). To continue and allow for a compact representation of the FIM, we define the following terms relating to the AoA at the UE
The corresponding terms for the AoD at the BS can be obtained by replacing \(r\) with \(t\) in ([\[equ:FIM_parameter_terms_los_1\]](#equ:FIM_parameter_terms_los_1){reference-type="ref" reference="equ:FIM_parameter_terms_los_1"}) and ([\[equ:FIM_parameter_terms_los_2\]](#equ:FIM_parameter_terms_los_2){reference-type="ref" reference="equ:FIM_parameter_terms_los_2"}). In addition to replacing \(\Delta_{\mathrm{r},\mathrm{u}}\) with \(\Delta_{\mathrm{l,m}}\), the term related to the elevation AoI defined as \({\mathbf{K}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\) can be obtained by replacing \(\theta_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\) with \(\theta_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\). A similar term for the azimuth AoI defined as \({\mathbf{P}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\) can be obtained by swapping \(\phi_{\mathrm{r}, {\mathrm{u}}}^{[{m}]}\) with \(\phi_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}\). The corresponding terms related to the AoR can be obtained from the terms related to the AoI by replacing \(r\) with \(t\). Other terms relating to AoI and AoR at the RISs are \[\label{equ:FIM_parameter_terms_ris_2} \begin{aligned} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]} \triangleq {\mathbf{\Gamma}}^{[{m}]} {\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}, \; \Tilde{\mathbf{a}_k}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]} \triangleq {\mathbf{K}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]} \; \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}, \; {\mathbf{a}}_{k_{\mathrm{t}, {\mathrm{l}}}}^{[{m}]} \triangleq {\mathbf{K}}_{\mathrm{t}, {\mathrm{l}}}^{[{m}]} {\mathbf{a}}_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}, \; \Tilde{\mathbf{a}_p}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]} & \triangleq {\mathbf{P}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[{m}]}, \; {\mathbf{a}}_{p_{\mathrm{t}, {\mathrm{l}}}}^{[{m}]} \triangleq {\mathbf{P}}_{\mathrm{t}, {\mathrm{l}}}^{[{m}]} {\mathbf{a}}_{\mathrm{t}, {\mathrm{l}}}^{[{m}]}, \end{aligned}\] \[\label{equ:FIM_parameter_terms_ris_2b} \begin{aligned} \mathbf{k}_{\mathrm{l}} & \triangleq \mathbf{p}_{\mathrm{l}} \triangleq \left[{\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[1]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[1]},{\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[2]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[2]}, \ldots, {\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[{M_1}]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{H}}, \\ \mathbf{k}_{\mathrm{t},\mathrm{l}} & \triangleq\left[{{\mathbf{{a}}^{\mathrm{H}}}}_{k_{\mathrm{t}, {\mathrm{l}}}}^{[1]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[1]},{{\mathbf{{a}}^{\mathrm{H}}}}_{k_{\mathrm{t}, {\mathrm{l}}}}^{[2]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[2]}, \ldots, {{\mathbf{{a}}^{\mathrm{H}}}}_{k_{\mathrm{t}, {\mathrm{l}}}}^{[{M_1}]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{H}}, \; \; \mathbf{p}_{\mathrm{t},\mathrm{l}}&\triangleq\left[{{\mathbf{{a}}^{\mathrm{H}}}}_{p_{\mathrm{t}, {\mathrm{l}}}}^{[1]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[1]},{{\mathbf{{a}}^{\mathrm{H}}}}_{p_{\mathrm{t}, {\mathrm{l}}}}^{[2]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[2]}, \ldots, {{\mathbf{{a}}^{\mathrm{H}}}}_{p_{\mathrm{t}, {\mathrm{l}}}}^{[{M_1}]} \Tilde{\mathbf{a}}_{\mathrm{r}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{H}}, \\ \mathbf{k}_{\mathrm{\mathrm{r},\mathrm{l}}} & \triangleq\left[{\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[1]} \Tilde{\mathbf{a}_k}_{\mathrm{r}, {\mathrm{l}}}^{[1]},{\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[2]} \Tilde{\mathbf{a}_k}_{\mathrm{r}, {\mathrm{l}}}^{[2]}, \ldots, {\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[{M_1}]} \Tilde{\mathbf{a}_k}_{\mathrm{r}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{H}}, \mathbf{p}_{\mathrm{\mathrm{r},\mathrm{l}}} &\triangleq\left[{\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[1]} \Tilde{\mathbf{a}_p}_{\mathrm{r}, {\mathrm{l}}}^{[1]},{\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[2]} \Tilde{\mathbf{a}_p}_{\mathrm{r}, {\mathrm{l}}}^{[2]}, \ldots, {\mathbf{{a}}^{\mathrm{H}}}_{\mathrm{t}, {\mathrm{l}}}^{[{M_1}]} \Tilde{\mathbf{a}_p}_{\mathrm{r}, {\mathrm{l}}}^{[{M_1}]}\right]^{\mathrm{H}}. \end{aligned}\] The scalar part of the RIS reflecting coefficient used for multipath separation is arranged as \[\label{equ:FIM_parameter_terms_ris_3} \begin{aligned} \bm{\gamma}^{[{m}]} = [{\gamma}^{[{m}]}_{1}, {\gamma}^{[{m}]}_{2}, \ldots, {\gamma}^{[{m}]}_{T}]^{\mathrm{T}}, \end{aligned}\] and can be arranged in a matrix as \[\label{equ:FIM_parameter_terms_ris_4} \begin{aligned} \mathbf{D}_{\gamma} = [\bm{\gamma}^{[1]}, \bm{\gamma}^{[2]}, \ldots, \bm{\gamma}^{[{M_1}]}]. \end{aligned}\] This matrix, henceforth referred to as a *sequence matrix* provides control in both spatial and temporal domains through the fast-varying part of reflecting coefficients of the RISs. The complex channel gains are arranged in a diagonal matrix \(\mathbf{B} \triangleq \operatorname{diag}(\bm{\beta})\) and the signal factor representing the effect of the transmitted beams is specified by \[\label{equ:FIM_parameter_terms_ris_5} \begin{aligned} \left[\mathbf{R}_{k}\right]_{u v} \triangleq \sum_{n = 1}^{N}\left(2 \pi n /\left(N T_{s}\right)\right)^{k} \mathbf{x}[n] \mathbf{x}^{\mathrm{H}}[n] e^{-j 2 \pi n \frac{\tau^{[v]}-\tau^{[u]}}{N T_{s}}}, \end{aligned}\] where \(k \in \{0,1,2\}\).
## Entries of the FIM
After taking the first derivative, and using the definitions in the previous sections, the submatrices of the FIM in ([\[equ:FIM_parameter_matrix\]](#equ:FIM_parameter_matrix){reference-type="ref" reference="equ:FIM_parameter_matrix"}) that are related to AoA at the UE can be written as \[\label{equ:FIM_exact_submat_theta_r_u_theta_r_u} \begin{aligned} \mathbf{J}_{\bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{r},\mathrm{u}}} &= \frac{2}{\sigma^2}\Re\left\{ \left( \mathbf{B}^{\mathrm{H}} \mathbf{K}_{\mathrm{r},\mathrm{u}}^{\mathrm{H}} \mathbf{K}_{\mathrm{r},\mathrm{u}} \mathbf{B}\right) \odot\left(\mathbf{k}_{\mathrm{l}} \mathbf{k}_{\mathrm{l}}^{\mathrm{H}}\right) \odot\left(\mathbf{D}_{\mathrm{\gamma}}^{\mathrm{H}} \mathbf{D}_{\mathrm{\gamma}}\right) \odot \left(\mathbf{A}_{\mathrm{t},\mathrm{u}}^{\mathrm{H}} \mathbf{F} \mathbf{F}^{\mathrm{H}} \mathbf{A}_{\mathrm{t},\mathrm{u}}^{}\right)^{\mathrm{T}} \odot \mathbf{R}_{0}\right\}. \end{aligned}\] where \(\odot\) represents element-wise matrix multiplication. From this equation, we notice that the FIM decreases with signal-to-noise (SNR). In general, we notice that all submatrices of the FIM in ([\[equ:FIM_parameter_matrix\]](#equ:FIM_parameter_matrix){reference-type="ref" reference="equ:FIM_parameter_matrix"}) can be written similarly. These equations are presented in Appendix [\[appendix:FIM_entries\]](#appendix:FIM_entries){reference-type="ref" reference="appendix:FIM_entries"} and they have a general form of \[\mathbf{J}_{\mathbf{v}_1 \mathbf{v}_{2}}=\Re\{(\text {Rx factor}) \odot(\text {RIS gain}) \odot(\text {RIS correlation}) \odot(\text {Tx factor}) \odot(\text {signal factor})\}.\] We notice that the submatrices can be decomposed into: i) information provided by the receiver specified by some combination of the terms \(\{ \mathbf{P}_{\mathrm{r}, \mathrm{u}}, \mathbf{K}_{\mathrm{r}, \mathrm{u}}, \mathbf{B}, \mathbf{A}_{r, u} \}\), ii) information provided by the transmitter specified by some combination of the terms \(\{ \mathbf{P}_{\mathrm{t}, \mathrm{u}}, \mathbf{K}_{\mathrm{t}, \mathrm{u}}, \mathbf{F}, \mathbf{A}_{\mathrm{t}, \mathrm{u}} \}\), iii) information provided by the RIS gain specified by some combination of the terms \(\{ \mathbf{k}_{\mathrm{l}}, \mathbf{k}_{\mathrm{t},\mathrm{l}}, \mathbf{k}_{\mathrm{r},\mathrm{l}}, \mathbf{p}_{\mathrm{t},\mathrm{l}}, \mathbf{p}_{\mathrm{r},\mathrm{l}} \}\), iv) the correlation across the RIS specified by the product \(\mathbf{D}_{\gamma}^{\mathrm{H}} \mathbf{D}_{\gamma}\), and v) the transmit signal factor.
## Prior Information Concerning the Channel Coefficients
In this subsection, we develop the Bayesian FIM by introducing prior information about the channel parameters. We assume that prior information is unavailable for the geometric channel parameters \(\bm{\eta}_{1}\)[^1]. However, a prior distribution of the nuisance parameters is known. More specifically, the distribution of the nuisance parameters (complex path gains) is as follows \[\label{equ:prior_nuisance} \begin{aligned} \begin{array}{llll} \chi(\bm{\beta}_{\mathrm{R}} ) &\sim \mathcal{N}(0, \frac{\sigma^2_{ \mathbf{\beta}_{\mathrm{}}}}{2} \mathbf{I}_{_{M_{\mathrm{1}}}}), & \chi(\bm{\beta}_{\mathrm{I}} ) &\sim \mathcal{N}(0, \frac{\sigma^2_{ \bm{\beta}_{\mathrm{}}}}{2} \mathbf{I}_{_{M_{\mathrm{1}}}}), \end{array} \end{aligned}\] where \(\sigma^2_{ \mathbf{\beta}_{\mathrm{}}}\) represents the uncertainty in the knowledge of the nuisance parameters. The FIM for the prior information \(\mathbf{J}_{ \bm{\eta}_{2}}^{\mathrm{P} }\) can be written as \[\label{equ:prior_information_channel_coeff} \begin{aligned} \mathbf{J}_{ \bm{\eta}_{2}}^{\mathrm{P} } &= -\mathbb{E}_{\mathbf{r} \mid \bm{\eta}_{2}}\left[\frac{\partial^{2} \ln \chi( \bm{\eta}_{2} )}{\partial \bm{\eta}_{2} \partial \bm{\eta}_{2}^{\mathrm{T}}}\right], \\ \end{aligned}\] which is an all zero matrix except \(\mathbf{J}_{\bm{\beta}_{\mathrm{R}}\bm{\beta}_{\mathrm{R}}}^{\mathrm{P} } = \mathbf{J}_{\bm{\beta}_{\mathrm{I}}\bm{\beta}_{\mathrm{I}}}^{\mathrm{P} } = \frac{2}{\sigma^2_{ \mathbf{\beta}_{\mathrm{}}}}\mathbf{I}_{_{M_{\mathrm{1}}}}.\) Hence, the entries in the Bayesian FIM can be written as \(\Tilde{\mathbf{J}}_{\mathbf{v}_1 \mathbf{v}_2 } = {\mathbf{J}}_{\mathbf{v}_1 \mathbf{v}_2 } + {\mathbf{J}}_{\mathbf{v}_1 \mathbf{v}_2 }^{\mathrm{P}}, \; \; \Tilde{\mathbf{J}}_{\mathbf{v}_3 \mathbf{v}_4 } = {\mathbf{J}}_{\mathbf{v}_3 \mathbf{v}_4 }\) where \(\mathbf{v}_{1}, \mathbf{v}_{2} \in \bm{\eta}_2\) and \(\mathbf{v}_{3}, \mathbf{v}_{4} \in \bm{\eta}_1\).
## Entries in the EFIM with a Unitary RIS Correlation Matrix
In this subsection, we analyze the structure of the EFIM when the correlation matrix is a unitary matrix. More specifically, we analyze the structure of the FIM when \(\mathbf{D}_{\mathrm{\gamma}}^{\mathrm{H}} \mathbf{D}_{\mathrm{\gamma}} = \mathbf{I}_{{M}_1}\). A unitary correlation matrix can help establish orthogonality among the paths received from different RISs, which allows the structure of the Bayesian FIM and EFIM to be analyzed.
To investigate the reduction in information due to uncertainty about the nuisance parameters, we analyze the term concerning information loss in the Bayesian EFIM \(\mathbf{J}_{\bm{\eta}_{1} \bm{\eta}_{1}}^{nu}\). This Bayesian EFIM is obtained by applying Definition [\[definition_EFIM_2\]](#definition_EFIM_2){reference-type="ref" reference="definition_EFIM_2"} to the Bayesian FIM given by \(\mathbf{J}_{\bm{\eta}}^{}\). First, we consider the information loss concerning \(\bm{\theta}_{\mathrm{r},\mathrm{u}}\) due to uncertainty in the nuisance parameters. The information loss due to the nuisance parameters which concerns the cross correlation between the receive elevation angle and the elevation angle of reflection can be written as \[\label{equ:FIM_nuis_equivalent_submat_theta_r_u_theta_t_l} \begin{aligned} \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{\mathrm{nu}} = \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{R}}} \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \bm{\beta}_{\mathrm{R}}}^{\mathrm{T}} + \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{I}}} \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{I}} \bm{\beta}_{\mathrm{I}}}^{-1} \mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \bm{\beta}_{\mathrm{I}}}^{\mathrm{T}}= \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1}\left[ \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{R}}} \mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \bm{\beta}_{\mathrm{R}}}^{\mathrm{T}} + \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{I}}} \mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \bm{\beta}_{\mathrm{I}}}^{\mathrm{T}} \right], \\ \end{aligned}\] which is a consequence of Remark [\[remark:nusiance_params_diagonal\]](#remark:nusiance_params_diagonal){reference-type="ref" reference="remark:nusiance_params_diagonal"}. Now, applying basic complex analysis, \(\Im{(\nu_1)}\Im{(\nu_2)} + \Re{(\nu_1)}\Re{(\nu_2)} = \Re{(\nu_1 \nu_2^{\mathrm{H}})}= \Re{(\nu_1^{\mathrm{H}} \nu_2^{})}\), we have \[\label{equ:FIM_nuis_equivalent_submat_theta_r_u_theta_t_l_1} \begin{aligned} \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{\mathrm{nu}} =\Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1}\left[ ( \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{I}}}+ j\mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\beta}_{\mathrm{R}}}) ( \mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \bm{\beta}_{\mathrm{I}}}^{\mathrm{}} +j\mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \bm{\beta}_{\mathrm{R}}}^{\mathrm{}} )^{\mathrm{H}} \right]. \\ \end{aligned}\] Now, substituting ([\[equ:FIM_exact_submat_theta_r\_u_beta\]](#equ:FIM_exact_submat_theta_r_u_beta){reference-type="ref" reference="equ:FIM_exact_submat_theta_r_u_beta"}) and ([\[equ:FIM_exact_submat_theta_t\_l_beta\]](#equ:FIM_exact_submat_theta_t_l_beta){reference-type="ref" reference="equ:FIM_exact_submat_theta_t_l_beta"}) in the above equation gives \[\label{equ:FIM_nuis_equivalent_submat_theta_r_u_theta_t_l_3} \begin{aligned} &\mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{\mathrm{nu}} = \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1}\left[-\frac{4}{\sigma^4}\Re\left\{\left(\mathbf{B}^{\mathrm{H}} \mathbf{K}_{\mathrm{r},\mathrm{u}}^{\mathrm{H}} \mathbf{A}_{\mathrm{r},\mathrm{u}} \right) \odot\left(\mathbf{k}_{\mathrm{l}} \mathbf{k}_{\mathrm{l}}^{\mathrm{H}}\right)\odot \left(\mathbf{D}_{\mathrm{\gamma}}^{\mathrm{H}} \mathbf{D}_{\mathrm{\gamma}}\right) \odot\left(\mathbf{A}_{\mathrm{t},\mathrm{u}}^{\mathrm{H}} \mathbf{F} \mathbf{F}^{\mathrm{H}} \mathbf{A}_{\mathrm{t},\mathrm{u}}^{}\right)^{\mathrm{T}} \odot \mathbf{R}_{0} \right. \right.\\ &\left.\left. \odot \left( \mathbf{A}_{\mathrm{r},\mathrm{u}}^{\mathrm{H}} \mathbf{A}_{\mathrm{r},\mathrm{u}}\mathbf{B}^{\mathrm{}} \right) \odot\left(\mathbf{k}_{\mathrm{l}}^{\mathrm{}}\mathbf{k}_{\mathrm{t},\mathrm{l}}^{\mathrm{H}} \right) \odot\left(\mathbf{D}_{\mathrm{\gamma}}^{\mathrm{H}} \mathbf{D}_{\mathrm{\gamma}}\right)^{\mathrm{}} \odot \left(\mathbf{A}_{\mathrm{t},\mathrm{u}}^{\mathrm{H}} \mathbf{F} \mathbf{F}^{\mathrm{H}} \mathbf{A}_{\mathrm{t},\mathrm{u}}^{}\right)^{\mathrm{T}} \odot \mathbf{R}_{0} \right\} \right]. \end{aligned}\] Rearranging the terms and applying the properties relating Hadamard products with diagonal matrices produces \[\label{equ:FIM_nuis_equivalent_submat_theta_r_u_theta_t_l_4} \begin{aligned} &\mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{\mathrm{nu}} = \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1}\left[-\frac{4}{\sigma^4}\Re\left\{\left(\mathbf{B}^{\mathrm{H}} \mathbf{K}_{\mathrm{r},\mathrm{u}}^{\mathrm{H}} \mathbf{A}_{\mathrm{r},\mathrm{u}}\mathbf{B}^{\mathrm{}} \right) \odot\left(\mathbf{k}_{\mathrm{l}}^{\mathrm{}}\mathbf{k}_{\mathrm{t},\mathrm{l}}^{\mathrm{H}} \right)\odot \left(\mathbf{D}_{\mathrm{\gamma}}^{\mathrm{H}} \mathbf{D}_{\mathrm{\gamma}}\right) \odot\left(\mathbf{A}_{\mathrm{t},\mathrm{u}}^{\mathrm{H}} \mathbf{F} \mathbf{F}^{\mathrm{H}} \mathbf{A}_{\mathrm{t},\mathrm{u}}^{}\right)^{\mathrm{T}} \odot \mathbf{R}_{0}\right. \right.\\ &\left.\left. \odot \left( \mathbf{A}_{\mathrm{r},\mathrm{u}}^{\mathrm{H}} \mathbf{A}_{\mathrm{r},\mathrm{u}} \right) \odot \left(\mathbf{k}_{\mathrm{l}} \mathbf{k}_{\mathrm{l}}^{\mathrm{H}}\right) \odot\left(\mathbf{D}_{\mathrm{\gamma}}^{\mathrm{H}} \mathbf{D}_{\mathrm{\gamma}}\right)^{\mathrm{}} \odot \left(\mathbf{A}_{\mathrm{t},\mathrm{u}}^{\mathrm{H}} \mathbf{F} \mathbf{F}^{\mathrm{H}} \mathbf{A}_{\mathrm{t},\mathrm{u}}^{}\right)^{\mathrm{T}} \odot \mathbf{R}_{0} \right\} \right]. \end{aligned}\] Using Remark [\[remark:nusiance_params_diagonal\]](#remark:nusiance_params_diagonal){reference-type="ref" reference="remark:nusiance_params_diagonal"} yields \[\label{equ:FIM_nuis_equivalent_submat_theta_r_u_theta_t_l_5} \begin{aligned} &\mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{\mathrm{nu}} = \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} {\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\left[-\frac{2}{\sigma^2}\Re\left\{\left(\mathbf{B}^{\mathrm{H}} \mathbf{K}_{\mathrm{r},\mathrm{u}}^{\mathrm{H}} \mathbf{A}_{\mathrm{r},\mathrm{u}}\mathbf{B}^{\mathrm{}} \right) \odot\left(\mathbf{k}_{\mathrm{l}}^{\mathrm{}}\mathbf{k}_{\mathrm{t},\mathrm{l}}^{\mathrm{H}} \right)\odot \left(\mathbf{D}_{\mathrm{\gamma}}^{\mathrm{H}} \mathbf{D}_{\mathrm{\gamma}}\right) \odot\left(\mathbf{A}_{\mathrm{t},\mathrm{u}}^{\mathrm{H}} \mathbf{F} \mathbf{F}^{\mathrm{H}} \mathbf{A}_{\mathrm{t},\mathrm{u}}^{}\right)^{\mathrm{T}} \odot \mathbf{R}_{0} \right\} \right]. \end{aligned}\] Now, recognizing that the term in the square brackets is \(\mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{\mathrm{}}\), we have \[\label{equ:FIM_nuis_equivalent_submat_theta_r_u_theta_t_l_6} \begin{aligned} \mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{\mathrm{nu}} &=\Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \bm{\theta}_{\mathrm{r},\mathrm{u}} \bm{\theta}_{\mathrm{t},\mathrm{l}}}^{}. \end{aligned}\] Similar expressions can be derived to quantify the information concerning the cross correlation between the receive angles and other geometric channel parameters due to uncertainty in the nuisance parameters, hence we have \[\label{equ:FIM_nuis_equivalent_submat_theta_r_u_theta_r_l_phi_r_l} \begin{aligned} \mathbf{J}_{ \mathbf{v}_{1} \mathbf{v}_{2}}^{\mathrm{nu}} =\Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \mathbf{v}_{1} \mathbf{v}_{2}}^{}, \end{aligned}\] where \(\mathbf{v}_{1} \in \{\bm{\theta}_{\mathrm{\mathrm{r}, {\mathrm{u}}}}, \bm{\phi}_{\mathrm{\mathrm{r}, {\mathrm{u}}}}\}\) and \(\mathbf{v}_{2} \in \bm{\eta} / \{\bm{\theta}_{\mathrm{\mathrm{r}, {\mathrm{u}}}}, \bm{\phi}_{\mathrm{\mathrm{r}, {\mathrm{u}}}} \}\). The information loss terms related to the receive angles that do not follow the above structure are presented in Appendix [\[appendix:FIM_nuis_remark_structure\]](#appendix:FIM_nuis_remark_structure){reference-type="ref" reference="appendix:FIM_nuis_remark_structure"}. All information loss terms in the EFIM expression related to the RIS parameters have an identical structure. Specifically, \[\label{equ:FIM_nuis_equivalent_submat_ris} \begin{array}{llll} \mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \mathbf{v}}^{\mathrm{nu}} & \triangleq \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \bm{\theta}_{\mathrm{t},\mathrm{l}} \mathbf{v} }^{\mathrm{}}, \mathbf{J}_{ \bm{\phi}_{\mathrm{t},\mathrm{l}} \mathbf{v}}^{\mathrm{nu}} & \triangleq \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \bm{\phi}_{\mathrm{t},\mathrm{l}} \mathbf{v} }^{\mathrm{}}, \\ \mathbf{J}_{ \bm{\theta}_{\mathrm{\mathrm{r},\mathrm{l}}} \mathbf{v}}^{\mathrm{nu}} & \triangleq \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \bm{\theta}_{\mathrm{\mathrm{r},\mathrm{l}}} \mathbf{v} }^{\mathrm{}}, \mathbf{J}_{ \bm{\phi}_{\mathrm{\mathrm{r},\mathrm{l}}} \mathbf{v}}^{\mathrm{nu}} & \triangleq \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \bm{\phi}_{\mathrm{\mathrm{r},\mathrm{l}}} \mathbf{v} }^{\mathrm{}},\\ \end{array}\] where \(\mathbf{v} \in \bm{\eta}\). Similarly, the information loss terms concerning the transmit angles can be expressed as \[\label{equ:FIM_nuis_equivalent_submat_tx_angle} \begin{array}{llll} \mathbf{J}_{ \bm{\theta}_{\mathrm{\mathrm{t}, {\mathrm{u}}}} \mathbf{v}_{1}}^{\mathrm{nu}} & \triangleq \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \bm{\theta}_{\mathrm{\mathrm{t}, {\mathrm{u}}}} \mathbf{v_1}_{}}^{\mathrm{}}, \mathbf{J}_{ \bm{\phi}_{\mathrm{\mathrm{t}, {\mathrm{u}}}} \mathbf{v}_{2}}^{\mathrm{nu}} & \triangleq \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{}\mathbf{J}_{ \bm{\phi}_{\mathrm{\mathrm{t}, {\mathrm{u}}}} \mathbf{v}_{2}}^{\mathrm{}}, \end{array}\] where \(\mathbf{v}_{1} \in \bm{\eta} / \{\bm{\theta}_{\mathrm{\mathrm{t}, {\mathrm{u}}}} \}\) and \(\mathbf{v}_{2} \in \bm{\eta} / \{\bm{\phi}_{\mathrm{\mathrm{t}, {\mathrm{u}}}} \}\). Also, the information loss terms related to the ToAs in the EFIM expression is written as \[\label{equ:FIM_nuis_equivalent_submat_toa} \begin{aligned} \mathbf{J}_{ \bm{\tau}_{} \mathbf{v}_{}}^{\mathrm{nu}} & \triangleq \Tilde{\mathbf{J}}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{-1} \mathbf{J}_{ \bm{\beta}_{\mathrm{R}} \bm{\beta}_{\mathrm{R}}}^{} \mathbf{J}_{ \bm{\tau}_{} \mathbf{v}_{}}^{\mathrm{}}, \end{aligned}\] where \(\mathbf{v} \in \bm{\eta} / \{\bm{\tau}_{}\}\).
## RIS Related Angle Definitions and Relationships
To analyze both the angle relationships and derive the FIM for positioning, we define the rotation matrix \(\mathbf{Q}\left(\theta_{0}, \phi_{0}\right)\) given by \(\mathbf{Q}\left(\theta_{0}, \phi_{0}\right) =\mathbf{Q}_{z}\left(\phi_{0}\right) \mathbf{Q}_{-x^{}}\left(\theta_{0}\right)\), where \(\mathbf{Q}_{z}\left(\phi_{0}\right)\) and \(\mathbf{Q}_{-x^{}}\left(\theta_{0}\right)\) define the counter-clockwise rotation around the \(z\)-axis and the clockwise rotation around the \(x\)-axis respectively. We define \(\mathbf{g}^{[{m}]}_{{}}=( \mathbf{p}^{[{m}]}-\mathbf{p}_{\mathrm{BS}})\), and specify the AoD at the BS as \(\theta_{\mathrm{t}, {\mathrm{u}}}^{[{m}]} =\cos ^{-1}\left({g}^{[{m}]}_{z_{}} /\|\mathbf{g}^{[{m}]}_{{}}\|\right), \; \; \phi_{\mathrm{t}, {\mathrm{u}}}^{[{m}]} =\tan ^{-1}\left({g}^{[{m}]}_{y_{}} /{g}^{[{m}]}_{x_{}}\right)\). Next, we translate the \({m}^{th}\) RIS to the origin, and the new coordinates of the BS can be written as \(\mathbf{c}^{[{m}]}_{{}}=(\mathbf{p}_{\mathrm{BS}}-\mathbf{p}^{[{m}]}), \; \; \Tilde{\mathbf{c}}^{[{m}]}_{}=\mathbf{Q}^{-1}\left(\theta_{0}^{[{m}]}, \phi_{0}^{[{m}]}\right) \mathbf{c}^{[{m}]}\). With respect to these new coordinates, we can write \(\theta_{\mathrm{r},\mathrm{l}}^{[{m}]} =\cos ^{-1}\left({c}^{[{m}]}_{\Tilde{z}} /\|\Tilde{\mathbf{c}}^{[{m}]}_{}\|\right), \; \; \phi_{\mathrm{r},\mathrm{l}}^{[{m}]} =\tan ^{-1}\left({c}^{[{m}]}_{\Tilde{y}} /{c}^{[{m}]}_{\Tilde{x}}\right)\). Subsequently, the translated coordinates allow the following definition \(\mathbf{v}^{[{m}]}_{{}}=(\mathbf{p}-\mathbf{p}^{[{m}]}), \; \; \Tilde{\mathbf{v}}^{[{m}]}_{}=\mathbf{Q}^{-1}\left(\theta_{0}^{[{m}]}, \phi_{0}^{[{m}]}\right) \mathbf{v}^{[{m}]}_{{}}\) and we can write \(\theta_{\mathrm{t},\mathrm{l}}^{[{m}]} =\cos ^{-1}\left({v}^{[{m}]}_{\Tilde{z}} /\|\Tilde{\mathbf{v}}^{[{m}]}_{}\|\right), \; \; \phi_{\mathrm{t},\mathrm{l}}^{[{m}]} =\tan ^{-1}\left({v}^{[{m}]}_{\Tilde{y}} /{v}^{[{m}]}_{\Tilde{x}}\right)\). Similarly, we obtain a new set of coordinates by translating the UE to the origin, and we write the following definitions \(\mathbf{e}^{[{m}]}_{{}}=-(\mathbf{p}-\mathbf{p}^{[{m}]}), \; \; \Tilde{\mathbf{e}}^{[{m}]}_{}=\mathbf{Q}^{-1}\left(\theta_{0}^{}, \phi_{0}^{}\right) \mathbf{e}^{[{m}]}_{{}}\). Hence, we can write \(\theta_{\mathrm{r},\mathrm{u}}^{[{m}]} =\cos ^{-1}\left({e}^{[{m}]}_{\Tilde{z}} /\|\Tilde{\mathbf{e}}^{[{m}]}_{}\|\right), \; \; \phi_{\mathrm{r},\mathrm{u}}^{[{m}]} =\tan ^{-1}\left({e}^{[{m}]}_{\Tilde{y}} /{e}^{[{m}]}_{\Tilde{x}}\right)\).
## Relationship between AoI and AoR with a Unitary RIS Correlation Matrix
Based on the angle relationships and the coordinate translations, we show the relationship between the AoI and AoR for an RIS deployed as a passive uniform rectangular array (URA).
The following corollaries establish relationships between the information provided by the FIMs of various channel parameters. The first corollary is a vital step in showing dependence among some of the FIMs of the geometric channel parameters. The first corollary establishes a relationship between the following: i) the derivative with respect to the elevation AoR of the exponent in the array response vector due to the reflected signal at the \({m}^{th}\) RIS specified by \(\mathbf{K}_{\mathrm{t},\mathrm{l}}^{[{m}]}\), ii) the derivative with respect to the azimuth AoR of the exponent in the array response vector due to reflected signal at the same RIS specified by \(\mathbf{P}_{\mathrm{t},\mathrm{l}}^{[{m}]}\), and iii) the derivative with respect to the elevation AoI of the exponent in the array response vector due to incident signal at the same RIS specified by \(\mathbf{K}_{\mathrm{\mathrm{r},\mathrm{l}}}^{[{m}]}\). More specifically, the first corollary indicates that there exists a linear combination of (i) and (ii) that produces (iii). Additionally, in the first corollary, a similar statement is made about the azimuth AoI.
Corollary [\[corollary:fim_ura_unitary\]](#corollary:fim_ura_unitary){reference-type="ref" reference="corollary:fim_ura_unitary"} shows that the information provided by the AoI can be obtained as linear combination of the information provided by the AoR.
## Fisher Information for RIS Paths Plus LOS
In this subsection, we analyze the FIM of the LOS plus RIS paths. We define the geometric LOS channel parameters \(\bm{\psi}_{1} \triangleq\left[\theta_{\mathrm{r}, {\mathrm{u}}}^{[0]}, \phi_{\mathrm{r}, {\mathrm{u}}}^{[0]}, \theta_{\mathrm{t}, {\mathrm{u}}}^{[0]}, \phi_{\mathrm{t}, {\mathrm{u}}}^{[0]}, \tau^{[\mathrm{0}]}_{} \right]^{\mathrm{T}}\) and the LOS nuisance parameter as \(\bm{\psi}_{2} \triangleq\left[{\beta}_{\mathrm{R}}^{\mathrm{[0]}}, {\beta}_{\mathrm{I}}^{\mathrm{[0]}} \right]^{\mathrm{T}}\). The LOS plus RIS geometric channel parameters are defined as \(\bm{\zeta}_{1} \triangleq\left[ \bm{\psi}_{1}^{\mathrm{T}}, \bm{\eta}_{1}^{\mathrm{T}} \right]^{\mathrm{T}}\) and the LOS plus RIS nuisance parameters is defined as \(\bm{\zeta}_{2} \triangleq\left[ \bm{\psi}_{2}^{\mathrm{T}}, \bm{\eta}_{2}^{\mathrm{T}} \right]^{\mathrm{T}}.\) Hence, the LOS plus RIS channel parameters can be written as \(\begin{aligned} \begin{array}{llll} \bm{\zeta}_{} \triangleq\left[ \bm{\zeta}_{1}^{\mathrm{T}}, \bm{\zeta}_{2}^{\mathrm{T}} \right]^{\mathrm{T}} \end{array} \end{aligned}\) To write the FIM of \(\bm{\zeta}_{}\), we define the PDF as \[\label{equ:los_pdf_joint_channel} \begin{aligned} \chi(\mathbf{r}_{t}[n]; \bm{\zeta}_{} ) = \chi(\mathbf{r}_{t}[n] | \bm{\zeta}_{} ) \chi(\bm{\zeta}_{} ), \end{aligned}\] where \(\chi(\bm{\zeta}_{} ) = \chi(\bm{\psi}_{1}) \chi( \bm{\psi}_{2})\chi(\bm{\eta}_{1}) \chi( \bm{\eta}_{2})\). The FIM of the LOS and the RIS channel parameters due to observation \(\bm{r}\) has the structure \[\mathbf{J}_{\zeta}^{\mathrm{D}} \triangleq\left[\begin{array}{c|ccc} \mathbf{J}_{\psi}^{\mathrm{D}} & \mathbf{J}_{\psi \eta}^{\mathrm{D}} \\ \hline \mathbf{J}_{ \eta \psi}^{\mathrm{D}} & \mathbf{J}_{\eta}^{\mathrm{D}} \\ \end{array}\right],\] where \(\mathbf{J}_{\zeta}^{\mathrm{D}} \in \mathbb{R}^{(11 {M}_1 +7) \times (11 {M}_1 +7)}\), \(\mathbf{J}_{\psi}^{\mathrm{D}} \in \mathbb{R}^{7 \times 7}\), and \(\mathbf{J}_{\psi \eta}^{\mathrm{D}} \in \mathbb{R}^{7 \times 11 {M}_1}\). The entries of the latter two matrices are written in Appendix [\[appendix:FIM_los_RIS\]](#appendix:FIM_los_RIS){reference-type="ref" reference="appendix:FIM_los_RIS"}. The Bayesian FIM \(\mathbf{J}_{ \bm{\zeta}_{}}\) can be written as described in Section [3](#section:Fisher_Information_for_RIS_Paths){reference-type="ref" reference="section:Fisher_Information_for_RIS_Paths"}. Likewise, the equivalent Bayesian EFIM \(\mathbf{J}_{ \bm{\zeta}^{}_{1}}^{\mathrm{e}}\) can be written using Definition [\[definition_EFIM_2\]](#definition_EFIM_2){reference-type="ref" reference="definition_EFIM_2"}.
## LOS Related Angle Definitions and Relationships
To analyze both the angle relationships and derive the FIM for positioning, we define \(\mathbf{g}^{}_{{}}=( \mathbf{p}^{}-\mathbf{p}_{\mathrm{BS}})\), and specify the angles of departure at the BS as \(\theta_{\mathrm{t}, {\mathrm{u}}}^{[\mathrm{0}]} =\cos ^{-1}\left({g}^{}_{z_{}} /\|\mathbf{g}^{}_{{}}\|\right), \; \; \phi_{\mathrm{t}, {\mathrm{u}}}^{[\mathrm{0}]} =\tan ^{-1}\left({g}^{}_{y_{}} /{g}^{}_{x_{}}\right)\). Next, we translate the UE to the origin, and the new coordinates of the BS can be written as \(\mathbf{e}^{}_{{}}=(\mathbf{p}_{\mathrm{BS}}-\mathbf{p}^{}), \; \; \Tilde{\mathbf{e}}^{}_{}=\mathbf{Q}^{-1}\left(\theta_{0}^{}, \phi_{0}^{}\right) \mathbf{e}^{}\). With respect to these new coordinates, we can write \(\theta_{\mathrm{r},\mathrm{u}}^{[\mathrm{0}]}=\cos ^{-1}\left({e}^{}_{\Tilde{z}} /\|\Tilde{\mathbf{e}}^{}_{}\|\right), \; \; \phi_{\mathrm{r},\mathrm{u}}^{[\mathrm{0}]} =\tan ^{-1}\left({e}^{}_{\Tilde{y}} /{e}^{}_{\Tilde{x}}\right)\).
# Fisher Information of Location Parameters
In this section, we derive the FIM and the EFIM of the location parameters. We derive these information matrices for both the case with an arbitrary RIS sequence matrix and the special case with RIS sequences that both generates a unitary correlation matrix and sum to zero (see Assumption [\[asumption_far_field_rf_control_1\]](#asumption_far_field_rf_control_1){reference-type="ref" reference="asumption_far_field_rf_control_1"}). Based on the FIMs and EFIMs of the location parameters, we also derive expressions for the PEB and OEB for the UE. The location parameters are defined as \[\bm{\eta}_{L} \triangleq\left[ \mathbf{o}_{}^{\mathrm{T}}, \mathbf{p}_{}^{\mathrm{T}}, \mathbf{{o}^{\mathrm{T}}}^{[1]}, \mathbf{{p}^{\mathrm{T}}}^{[1]},\mathbf{{o}^{\mathrm{T}}}^{[2]}, \mathbf{{p}^{\mathrm{T}}}^{[2]},\ldots,\mathbf{{o}^{\mathrm{T}}}^{[{M_1}]}, \mathbf{{p}^{\mathrm{T}}}^{[{M_1}]} \right]^{\mathrm{T}} \triangleq\left[ \mathbf{o}_{}^{\mathrm{T}}, \mathbf{p}_{}^{\mathrm{T}}, \mathbf{{q}^{\mathrm{T}}} \right]^{\mathrm{T}}.\] The PDF \(\chi(\mathbf{r}_{t}[n]; \mathbf{\eta}_{L} )\) is obtained as \[\label{equ:pdf_joint_positioning} \begin{aligned} \chi(\mathbf{r}_{t}[n]; \bm{\eta}_{L} ) = \chi(\mathbf{r}_{t}[n] | \bm{\eta}_{L} ) \chi(\bm{\eta}_{L} ) \end{aligned}\] where \[\label{equ:pdf_joint_positioning_1} \begin{aligned} \chi(\bm{\eta}_{L} ) &= \chi(\mathbf{o}_{}^{}; \mathbf{p}_{}^{}) \prod_{m \in \mathcal{M}_{1}} \chi(\mathbf{o}_{}^{[{m}]}; \mathbf{p}_{}^{[{m}]} | \mathbf{o}_{}^{}, \mathbf{p}_{}^{}) = \chi(\mathbf{o}_{}^{}) \chi( \mathbf{p}_{}^{}) \prod_{m \in \mathcal{M}_{1}} \chi(\mathbf{o}_{}^{[{m}]}| \mathbf{o}_{}^{}, \mathbf{p}_{}^{}) \chi(\mathbf{p}_{}^{[{m}]}| \mathbf{o}_{}^{}, \mathbf{p}_{}^{}). \end{aligned}\] Based on the PDF in ([\[equ:pdf_joint_positioning\]](#equ:pdf_joint_positioning){reference-type="ref" reference="equ:pdf_joint_positioning"}), the FIM can be written as \(\mathbf{J}_{ \bm{\eta}_{L}} = \mathbf{J}_{ \bm{\eta}_{L}} ^{\mathrm{D} } + \mathbf{J}_{ \bm{\eta}_{L}} ^{\mathrm{P} }\). The parameter vector \(\bm{\eta}_{L}\) has a nonlinear relationship with the geometric channel parameters \(\bm{\eta}_{L} = \Upsilon(\bm{\zeta}_{1}).\) In, it was shown that this nonlinear relationship allows the FIM to be written as \[\label{equ:FIM_position} \begin{aligned} \mathbf{J}_{\bm{\eta}_{\mathrm{L}}}^{\mathrm{D}} \triangleq \mathbf{\Upsilon} \mathbf{J}_{\bm{\zeta}_{1}}^{\mathrm{e}} \mathbf{\Upsilon}^{\mathrm{T}}, \end{aligned}\] where \(\mathbf{\Upsilon}\) is a transformation matrix obtained by finding the gradient of the relationship between the location parameters and the geometric channel parameters.
## Transformation Matrix from Geometric Channel Parameters to Location Parameters
The transformation matrix in ([\[equ:FIM_position\]](#equ:FIM_position){reference-type="ref" reference="equ:FIM_position"}) can be defined as \[\mathbf{\Upsilon} \triangleq \frac{\partial \bm{\zeta}_{\mathrm{1}}^{\mathrm{T}}}{\partial \bm{\eta}_{\mathrm{L}}}=\left[\begin{array}{lllll} \frac{\partial \bm{\psi}^{}_{1}}{\partial \mathbf{o}_{}} & \frac{\partial \bm{\psi}^{}_{1}}{\partial \mathbf{p}_{}} & \frac{\partial \bm{\psi}^{}_{1}}{\partial \mathbf{q}} \\ \frac{\partial \bm{\eta}^{}_{1}}{\partial \mathbf{o}_{}} & \frac{\partial \bm{\eta}^{}_{1}}{\partial \mathbf{p}_{}} & \frac{\partial \bm{\eta}^{}_{1}}{\partial \mathbf{q}} \\ \end{array}\right]^{\mathrm{T}}\] with the entries of this matrix presented in Appendix [\[appendix:entries_transformation\]](#appendix:entries_transformation){reference-type="ref" reference="appendix:entries_transformation"} and Appendix [\[appendix:los_entries_transformation\]](#appendix:los_entries_transformation){reference-type="ref" reference="appendix:los_entries_transformation"}.
## Bayesian PEB and OEB: Arbitrary Correlation Matrix
In this section, we derive the FIM for the location parameters with an arbitrary RIS correlation matrix. This FIM has the structure \[\label{equ:position_FIM_1} \begin{aligned} \mathbf{J}_{\bm{\eta}_{\mathrm{L}}}^{\mathrm{D}} \triangleq \mathbf{\Upsilon} \mathbf{J}_{\bm{\zeta}_{1}}^{\mathrm{e}} \mathbf{\Upsilon}^{\mathrm{T}} =\left[\begin{array}{cc} \mathbf{J}_{\bm{\eta}_{L_1}\bm{\eta}_{L_1}} & \mathbf{J}_{\bm{\eta}_{L_1}\bm{\eta}_{L_2}} \\ \mathbf{J}_{\bm{\eta}_{L_1}\bm{\eta}_{L_2}}^{\mathrm{T}} & \mathbf{J}_{\bm{\eta}_{L_2}\bm{\eta}_{L_2}} \end{array}\right]. \end{aligned}\] With the assumption of independent prior information about the UE and the RISs, the entries in the Bayesian FIM \(\mathbf{J}_{ \bm{\eta}_{L}}\) can be written as \(\Tilde{\mathbf{J}}_{ {}_{\mathbf{v}_1 \mathbf{v}_1 }} = {\mathbf{J}}_{ {}_{\mathbf{v}_1 \mathbf{v}_1 }} + {\mathbf{J}}_{ {}_{\mathbf{v}_1 \mathbf{v}_1 }}^{\mathrm{P}}, \Tilde{\mathbf{J}}_{ {}_{\mathbf{v}_1 \mathbf{v}_2 }} = {\mathbf{J}}_{ {}_{\mathbf{v}_1 \mathbf{v}_2 }}\) where \(\mathbf{v}_{1}, \mathbf{v}_{2} \in \bm{\eta}_L\). Using Definition [\[definition_EFIM_2\]](#definition_EFIM_2){reference-type="ref" reference="definition_EFIM_2"} and ([\[equ:position_FIM_1\]](#equ:position_FIM_1){reference-type="ref" reference="equ:position_FIM_1"}), we can write the Bayesian FIM as \[\label{equ:position_EFIM_position_1} \begin{aligned} \mathbf{J}_{\bm{\eta}_{L}}^{\mathrm{e}}&=\mathbf{J}_{_{\bm{\eta}_{L_1}\bm{\eta}_{L_1}}}-\mathbf{J}_{_{\bm{\eta}_{L_1}\bm{\eta}_{L_1}}}^{nu} =\mathbf{J}_{_{\bm{\eta}_{L_1}\bm{\eta}_{L_1}}}-\mathbf{J}_{_{\bm{\eta}_{L_1}\bm{\eta}_{L_2}}} \mathbf{J}_{_{\bm{\eta}_{L_2}\bm{\eta}_{L_2}}}^{-1} \mathbf{J}_{_{\bm{\eta}_{L_1}\bm{\eta}_{L_2}}}^{\mathrm{T}}. \end{aligned}\] The term \(\mathbf{J}_{_{\bm{\eta}_{L_1}\bm{\eta}_{L_1}}}^{nu}\) accounts for information loss about the UE location due to the uncertainty in the RISs orientation and position. The SPEB and the SOEB with arbitrary RIS correlation matrix are defined as \[\label{equ:position_EFIM_position_exact} \begin{aligned} \operatorname{SOEB} &=\left[\left(\mathbf{J}_{\bm{\eta}_{L}}^{\mathrm{e}}\right)^{-1}\right]_{1,1}+\left[\left(\mathbf{J}_{\bm{\eta}_{L}}^{\mathrm{e}}\right)^{-1}\right]_{2,2}, \\ \operatorname{SPEB} &=\left[\left(\mathbf{J}_{\bm{\eta}_{L}}^{\mathrm{e}}\right)^{-1}\right]_{3,3}+\left[\left(\mathbf{J}_{\bm{\eta}_{L}}^{\mathrm{e}}\right)^{-1}\right]_{4,4}+\left[\left(\mathbf{J}_{\bm{\eta}_{L}}^{\mathrm{e}}\right)^{-1}\right]_{5,5}. \end{aligned}\]
## Bayesian PEB and OEB: Unitary Correlation Matrices and Assumption [\[asumption_far_field_rf_control_1\]](#asumption_far_field_rf_control_1){reference-type="ref" reference="asumption_far_field_rf_control_1"}
Under the conditions of unitary RIS correlation matrices, the restriction in ([\[equ:far_field_rf_control_1\]](#equ:far_field_rf_control_1){reference-type="ref" reference="equ:far_field_rf_control_1"}), and independent RIS placements; there is no mutual information between RISs and each path conveys information independently. The restriction also implies that there is no mutual information between the RISs and the LOS path. Hence, the parameters \(\bm{\eta}_{1}\) can be rearranged according to paths such \(\Tilde{\bm{\eta}}_{1} \triangleq\left[ {\mathbf{\eta}}^{[1]}_{1}, {\mathbf{\eta}}^{[2]}_{1}, \cdots, {\mathbf{\eta}}^{[{M_1}]}_{1} \right]\) where \({\bm{\eta}}^{[{m}]}_{1} \triangleq\left[ {{\theta}_{\mathrm{r},\mathrm{u}}^{[{m}]}}^{\mathrm{T}}, {{\phi}_{\mathrm{r},\mathrm{u}}^{[{m}]}} ^{\mathrm{T}}, {{\theta}_{\mathrm{t},\mathrm{l}}^{[{m}]}} ^{\mathrm{T}}, {\mathbf{\phi}_{\mathrm{t},\mathrm{l}}^{[{m}]}} ^{\mathrm{T}}, { {\theta}_{\mathrm{t},\mathrm{u}}^{[{m}]}} ^{\mathrm{T}}, { {\phi}_{\mathrm{t},\mathrm{u}}^{[{m}]}} ^{\mathrm{T}}, { {\tau}^{[{m}]}} ^{\mathrm{T}}_{} \right]^{\mathrm{T}}\). The nuisance parameters can be arranged similarly \({\bm{\eta}_{2}^{[{m}]}} \triangleq\left[{{\beta}_{\mathrm{R}}^{[{m}]}}^{\mathrm{T}}, {{\beta}_{\mathrm{I}}^{[{m}]}}^{\mathrm{T}} \right]^{\mathrm{T}}\). The corresponding rearranged EFIM is \[\label{equ:position_EFIM_position_rearranged} \begin{aligned} \Bar{\mathbf{J}}_{ \Tilde{\bm{\eta}}_{1}}^{\mathrm{e}} \triangleq \text{diag}\left[ {\Bar{\mathbf{J}}_{ {{1}}^{}}^{\mathrm{e}}}, {\Bar{\mathbf{J}}_{ {{2}}^{}}^{\mathrm{e}}}, \cdots, {\Bar{\mathbf{J}}_{ {{{M}_1}}^{}}^{\mathrm{e}}} \right], \end{aligned}\] where \(\Bar{\mathbf{J}}_{ {{m}}^{}}^{\mathrm{e}} = \Bar{\mathbf{J}}_{ {{{\bm{\eta}}^{[{m}]}_{1} }} {\bm{\eta}}^{[{m}]}_{1}}-\Bar{\mathbf{J}}_{ {{{\bm{\eta}}^{[{m}]}_{1} }} {\bm{\eta}}^{[{m}]}_{2}} \Bar{\mathbf{J}}_{ {{{\bm{\eta}}^{[{m}]}_{2} }} {\bm{\eta}}^{[{m}]}_{2}}^{\mathrm{-1}} \Bar{\mathbf{J}}_{ {{{\bm{\eta}}^{[{m}]}_{1} }} {\bm{\eta}}^{[{m}]}_{2}}^{\mathrm{T}} \; \; {m} \in \mathcal{M}_1\), and we can write \[\Bar{\mathbf{J}}_{ \Tilde{\bm{\zeta}}_{1}}^{\mathrm{e}} \triangleq\left[\begin{array}{c|ccc} \Bar{\mathbf{J}}_{ {\bm{\psi}}_{1}}^{\mathrm{e}} & \mathbf{0} \\ \hline \mathbf{0} & \Bar{\mathbf{J}}_{ \Tilde{\bm{\eta}}_{1}}^{\mathrm{e}} \\ \end{array}\right]\] where \(\Bar{\mathbf{J}}_{ {\bm{\psi}}_{1}}^{\mathrm{e}}\) presented in the Appendix [\[appendix:EFIM_Los\]](#appendix:EFIM_Los){reference-type="ref" reference="appendix:EFIM_Los"} is the EFIM obtained by applying Definition [\[definition_EFIM_2\]](#definition_EFIM_2){reference-type="ref" reference="definition_EFIM_2"} to the Bayesian FIM \(\mathbf{J}_{ \bm{\psi}^{}_{}}^{\mathrm{}}\) of the channel parameters for the LOS path. Accordingly, the translation matrix \(\mathbf{\Upsilon}^{}\) can be written as \[\label{equ:position_translation_rearranged} \begin{aligned} \mathbf{\Upsilon} \triangleq\left[\begin{array}{ccccc} \overline{\mathbf{\Upsilon}}_{\mathrm{0}}& \overline{\mathbf{\Upsilon}}_{\mathrm{1}}& \overline{\mathbf{\Upsilon}}_{\mathrm{2}} & \cdots & \overline{\mathbf{\Upsilon}}_{{M_1}} \\ \mathbf{0} & \overline{\overline{\mathbf{\Upsilon}}}_{\mathrm{1}} & \cdots & \cdots & \mathbf{0} \\ \vdots & \vdots & \ddots& \ddots & \vdots \\ \mathbf{0} & \mathbf{0} & \cdots &\cdots & \overline{\overline{\mathbf{\Upsilon}}}_{{M_1}} \end{array}\right],\end{aligned}\] where \(\overline{\mathbf{\Upsilon}}_{\mathrm{0}}\) is a \(5 \times 7\) matrix relating the LOS path to the UEs orientation and position, \(\overline{\mathbf{\Upsilon}}_{m}\) is the \(5 \times 7\) matrix relating the location of the \({m}^{{th}}\) RIS to the UEs orientation and position, and \(\overline{\overline{\mathbf{\Upsilon}}}_{m}\) is the \(5 \times 7\) matrix related to \(\mathbf{{o}^{}}^{[{m}]}\) and \(\mathbf{{p}^{}}^{[{m}]}\). The correspondingly rearranged prior matrix is defined as \[\label{equ:positioning_FIM_1} \begin{aligned} {\mathbf{J}}_{ {\bm{\eta}}_{L}}^{\mathrm{P}} \triangleq \text{diag}\left[ {\mathbf{J}}_{{\mathrm{UE}}}^{\mathrm{P}}, {\mathbf{J}}_{{\mathrm{1}}}^{\mathrm{P}}, \cdots, {\mathbf{J}}_{{{M_1}}}^{\mathrm{P}} \right]. \end{aligned}\] Hence, the Bayesian FIM for the positioning parameters is \[\label{equ:positioning_FIM_2} \begin{aligned} \mathbf{J}_{ \bm{\eta}_{L}} &=\mathbf{\Upsilon}\Bar{\mathbf{J}}_{ \Tilde{\bm{\zeta}}_{1}}^{\mathrm{e}} \mathbf{\Upsilon}^{\mathrm{T}} + {\mathbf{J}}_{ {\mathbf{\eta}}_{L}}^{\mathrm{P}} \\ &=\left[\begin{array}{ccccc} \sum_{{m}=0}^{{M}_1} \overline{\mathbf{\Upsilon}}_{{m}} \Bar{\mathbf{J}}_{ {{m}}^{}}^{\mathrm{e}} \overline{\mathbf{\Upsilon}}_{{m}}^{\mathrm{T}} + {\mathbf{J}}_{{\mathrm{UE}}}^{\mathrm{P}}& \overline{\mathbf{\Upsilon}}_{\mathrm{1}} \Bar{\mathbf{J}}_{ {{1}}^{}}^{\mathrm{e}} \overline{\overline{\mathbf{\Upsilon}}}_{\mathrm{1}}^{\mathrm{T}} & \ldots & \overline{\mathbf{\Upsilon}}_{\mathrm{{M}_1}} \Bar{\mathbf{J}}_{ {{{M}_1}}^{}}^{\mathrm{e}} \overline{\overline{\mathbf{\Upsilon}}}_{{M_1}}^{\mathrm{T}} \\ \overline{\overline{\mathbf{\Upsilon}}}_{\mathrm{1}} \Bar{\mathbf{J}}_{ {{1}}^{}}^{\mathrm{e}} {\overline{\mathbf{\Upsilon}}}_{\mathrm{1}}^{\mathrm{T}} & \overline{\overline{\mathbf{\Upsilon}}}_{\mathrm{1}} \Bar{\mathbf{J}}_{ {{1}}^{}}^{\mathrm{e}} \overline{\overline{\mathbf{\Upsilon}}}_{\mathrm{1}}^{\mathrm{T}} + {\mathbf{J}}_{{\mathrm{1}}}^{\mathrm{P}} & \ldots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ \overline{\overline{\mathbf{\Upsilon}}}_{{M_1}} \Bar{\mathbf{J}}_{ {{{M}_1}}^{}}^{\mathrm{e}} {\overline{\mathbf{\Upsilon}}}_{{M_1}}^{\mathrm{T}} & 0 & \ldots & \overline{\overline{\mathbf{\Upsilon}}}_{{M_1}} \Bar{\mathbf{J}}_{ {{{M}_1}}^{}}^{\mathrm{e}} \overline{\overline{\mathbf{\Upsilon}}}_{{M_1}}^{\mathrm{T}} + {\mathbf{J}}_{{{M_1}}}^{\mathrm{P}} \end{array}\right]. \end{aligned}\] Using Definition [\[definition_EFIM_2\]](#definition_EFIM_2){reference-type="ref" reference="definition_EFIM_2"}, the EFIM can be written as \[\label{equ:positioning_FIM_3} \begin{aligned} \mathbf{J}_{\bm{\eta}_{L}}^{\mathrm{e}} &= \overline{\mathbf{\Upsilon}}_{\mathrm{0}} \Bar{\mathbf{J}}_{ {{0}}^{}}^{\mathrm{e}} \overline{\mathbf{\Upsilon}}_{\mathrm{0}}^{\mathrm{T}} + \sum_{m=1}^{{M}_1} \overline{\mathbf{\Upsilon}}_{{m}} \Bar{\mathbf{J}}_{ {{{m}}}^{}}^{\mathrm{e}} \overline{\mathbf{\Upsilon}}_{{m}}^{\mathrm{T}} + {\mathbf{J}}_{{\mathrm{UE}}}^{\mathrm{P}}-\sum_{m=1}^{{M}_1} \overline{\mathbf{\Upsilon}}_{{m}} \Bar{\mathbf{J}}_{ {{{m}}}^{}}^{\mathrm{e}} \overline{\overline{\mathbf{\Upsilon}}}_{{m}}^{\mathrm{T}} \left(\overline{\overline{\mathbf{\Upsilon}}}_{{m}} \Bar{\mathbf{J}}_{ {{{m}}}^{}}^{\mathrm{e}} \overline{\overline{\mathbf{\Upsilon}}}_{{m}}^{\mathrm{T}} + {\mathbf{J}}_{{{m}}}^{\mathrm{P}} \right)^{-1} \overline{\overline{\mathbf{\Upsilon}}}_{{m}} \Bar{\mathbf{J}}_{ {{{m}}}^{}}^{\mathrm{e}} {\overline{\mathbf{\Upsilon}}}_{{m}}^{\mathrm{T}}. \\ \end{aligned}\] Hence, the FIM is partly composed of the FIM provided by the LOS path plus the FIM provided by all RIS paths. The corresponding SPEB and the SOEB can be obtained using Equation ([\[equ:position_EFIM_position_exact\]](#equ:position_EFIM_position_exact){reference-type="ref" reference="equ:position_EFIM_position_exact"}).
# Numerical Results
In this section, we evaluate the derived localization bounds with Monte Carlo simulations under different scenarios. Without loss of generality, we assume that the slow-varying reflection coefficients, \(\mathbf{\Gamma}^{[{m}]}\), are randomly generated, and the sequence matrix is a unitary matrix. Note that when LOS is available, Assumption [\[asumption_far_field_rf_control_1\]](#asumption_far_field_rf_control_1){reference-type="ref" reference="asumption_far_field_rf_control_1"} will be used to ensure that the LOS is orthogonal to the paths generated by the RIS. We focus on the case with URAs at the BS, RIS, and the UE with their respective normal vectors in the \(z\) direction. Except stated otherwise, the UE is operating at a frequency of \(30\) GHz, the transmit antenna gain is \(6\) dB with a transmit power of \(5 \; \text{dBm}\), the UE antenna gain is \(2\) dB, \(N_0 =-174 \; \text{dBm /\ Hz}\) and there are \(N = 256\) subcarriers. There are \(N_{T} = 4\) transmit antennas, \(N_{\mathrm{B}} = N_{T}\) transmit beams, and the UE orientation offset is \((\theta_0^{}, \phi_0^{}) = (10^{\circ}, 0^{\circ})\). The considered area is \(100 \; \text{m} \times 100 \; \text{m}\), and the considered bandwidth is \(0.1 \; \text{GHz}\) with a non-specular reflection pathloss model of \(1/{\rho^{[{m}]}} = \frac{\lambda^4 (\cos{\theta^{[{m}]}_{\mathrm{t},\mathrm{l}}} \cos{\theta^{[{m}]}_{\mathrm{r},\mathrm{l}}})^{0.57}}{512\pi^2 (d^{[{m}]}_{\mathrm{r},\mathrm{l}})^2 (d^{[{m}]}_{\mathrm{t},\mathrm{l}})^2}\) for the RIS-paths and \(1/{\rho^{[0]}} = \frac{\lambda^2}{(4\pi d^{}_{\mathrm{r},\mathrm{u}} )^2}\) for the LOS path, where \(d^{}_{\mathrm{r},\mathrm{u}}\), \(d^{[{m}]}_{\mathrm{r},\mathrm{l}}\), and \(d^{[{m}]}_{\mathrm{t},\mathrm{l}}\) are the distances between the BS and the UE, between the BS and the \({m}^\text{th}\) RIS, and between the \({m}^\text{th}\) RIS and the UE, respectively. To illustrate the impact of both RIS uncertainty and the number of RISs, we assume that the complex path gains are completely known in the simulation results below [^2]. For the case studies in this section, we assume that there is no prior information about the UE position, \({\mathbf{J}}_{{\mathrm{UE}}}^{\mathrm{P}} = \mathbf{0}_{5}\), but the prior information concerning the perturbed RISs is given as \({\mathbf{J}}_{{{m}}}^{\mathrm{P}} = \frac{2}{\sigma^2_{ {m}_{\mathrm{}}}}\mathbf{I}_{_{5}}, \; \; m \in \mathcal{M}_1^{b}\). The orientation offset of the perturbed RIS is given by \((\theta_0^{[{m}]}, \phi_0^{[{m}]}) = (45^{\circ}, 35^{\circ}), \; \; m \in \mathcal{M}_1^{b}\). In the subsequent figures, "PEB\" and "OEB\" refer to the error bounds without any prior information on the perturbed RISs, and "B-PEB\" and "B-OEB\" refer to the error bounds with prior information on the perturbed RISs. Finally, "B-PEB-LOS\" and "B-OEB-LOS\" refer to the error bounds with prior information on the perturbed RISs in the presence of an LOS path. In Figures [\[Results:peb_oeb_varying_numb_of_receive_ris\]](#Results:peb_oeb_varying_numb_of_receive_ris){reference-type="ref" reference="Results:peb_oeb_varying_numb_of_receive_ris"} and [\[Results:peb_oeb_varying_numb_of_ris_ris\]](#Results:peb_oeb_varying_numb_of_ris_ris){reference-type="ref" reference="Results:peb_oeb_varying_numb_of_ris_ris"}, the BS is located at \((0,0, 40 \text{m})\), the RIS and the UE are at a height of \(35 \text{m}\) and \(5 \text{m}\) respectively. In both figures, \(|\mathcal{M}_1^{b}| = 1\). The \(({x},{y})\) coordinate of both RIS and UE are randomly generated. Note that in the massive MIMO literature, the PEB and OEB scale down tremendously with an increase in the number of receive antennas. This scaling is attributed to the granular spatial resolution available in the large receive antenna regime. However, in an RIS enabled multipath environment, more receive antennas help, but not as much. In fact, the PEB and OEB improve until a point after which they enter the regime of diminishing returns. This phenomena is due to the orthogonal sequences employed by the RIS. These sequences cause the paths at the UE to be orthogonal to each other by design, because of which the deployment of a very large number of receive antennas is rendered redundant in this context. In Figures [\[Results:peb_oeb_varying_numb_of_receive_ris\]](#Results:peb_oeb_varying_numb_of_receive_ris){reference-type="ref" reference="Results:peb_oeb_varying_numb_of_receive_ris"} and [\[Results:peb_oeb_varying_numb_of_ris_ris\]](#Results:peb_oeb_varying_numb_of_ris_ris){reference-type="ref" reference="Results:peb_oeb_varying_numb_of_ris_ris"}, the localization bounds for different numbers of RISs are also presented. We observe that the difference between the Bayesian and the non-Bayesian bounds is very large when \(M_1 \leq 2\). This indicates the importance of prior information when there is one uncertain RIS \(|\mathcal{M}_1^{b}| = 1\), and \(|\mathcal{M}_1^{}| = 2\). However, when \(|\mathcal{M}_1^{}| > 2\), the RISs in the set \(\mathcal{M}_1^{a}\) provide enough information for UE localization. Lastly, in Figures [\[Results:peb_oeb_varying_numb_of_receive_ris\]](#Results:peb_oeb_varying_numb_of_receive_ris){reference-type="ref" reference="Results:peb_oeb_varying_numb_of_receive_ris"} and [\[Results:peb_oeb_varying_numb_of_ris_ris\]](#Results:peb_oeb_varying_numb_of_ris_ris){reference-type="ref" reference="Results:peb_oeb_varying_numb_of_ris_ris"}, we also present results for the scenarios where LOS paths are present in addition to the RIS paths. These scenarios provide lower bounds for the scenarios without LOS. Note that the gap between the localization bounds in the RIS-only scenarios and the RIS plus LOS scenarios reduces as the number of RISs increases.
# Conclusion
In this paper, the effect of the multipath created due to multiple RISs on wireless-enabled localization has been investigated from a Bayesian perspective. This was achieved by viewing the position and orientation of the RISs as prior information to assist downlink UE localization. On our way to describing the Bayesian bounds on localization accuracy, we first derived the FIM for the channel parameters. We showed that this FIM is decomposable into i) information provided by the transmitter, ii) information provided by the RIS, and iii) information provided by the receiver. Next, we showed through the derivations of the Bayesian EFIM that all information about the RIS-related geometric channel parameters is lost when the complex path gains are not known. We noted that this loss of information has severe implications on any parametric channel estimator, severely hindering localization. Next, we transformed the Bayesian EFIM of the geometric channel parameters to the Bayesian FIM for localization. We obtained the Bayesian EFIM of localization by considering RISs, with position and orientation offsets. Finally, through numerical results, we provided further insights on the effect of having a set of RISs with perturbed position/orientation on the localization performance.
[^1]: This is without loss of generality as any prior information about \(\bm{\eta}_{1}\) can be added in a manner similar to the addition of prior information about the nuisance parameters as described in the later part of this subsection.
[^2]: The knowledge of the complex path gains does not change the number of RISs needed, however, if this knowledge is unavailable, the UE location cannot be estimated as discussed already in Remark [\[remark:EFIM_remark_structure\]](#remark:EFIM_remark_structure){reference-type="ref" reference="remark:EFIM_remark_structure"}. | {'timestamp': '2022-06-06T02:16:54', 'yymm': '2206', 'arxiv_id': '2206.01652', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01652'} |
null | null |
null | null |
# Extended Abstract {#extended-abstract .unnumbered}
Vaccination hesitancy has been a threat to public health since a long time , and it is currently being fueled by the large spreading of misinformation in many social media . Attitudes against vaccination usually rise around fake news or conspiracy theories, but according to moral psychology they are shaped by individual moral preferences. The *Moral Foundations Theory* (MFT) explains individual variations in moral preferences using five dimensions (foundations): harm, fairness, loyalty, authority and purity, and it has been shown that these foundations underlie human judgements on politics, religion and social cooperation . In this work we assess the moral preferences expressed in the pro-and anti-vax discourses in a large dataset of Facebook comments. We show that significant differences exist in some moral dimensions between these two groups, and we suggest that these differences should be taken into account when designing vaccination campaigns. Previous evidence that moral values are related to individual attitudes towards vaccination can be found in , where the authors used a sample of \(1,000\) parents in the US and found that parents with higher vaccine hesitancy tended to have higher levels of the purity and liberty foundations. Surprisingly, they also found that the harm/care foundation, traditionally addressed in vaccination campaigns, did not have a significant impact on vaccine hesitancy. A similar study by was performed in Australia based on an online questionnaire administered to \(296\) individuals. The authors found that vaccine rejecters showed significantly lower values of authority and significantly higher levels of purity, liberty and fairness. Regarding the vaccination discourse in Facebook, categorized a set of \(\approx 20,000\) comments about vaccination in Poland, and they found that the main categories related with vaccine hesitancy were--in decreasing order of importance--conspiracy theories, falsehoods, and concerns regarding safety and effectiveness of vaccines. Similar clusters were obtained by with a sample of \(197\) individuals in Facebook in the US. Both works discuss narrative strategies to raise awareness on the importance of immunization.
The aim of this work is to provide more evidence on the impact of moral traits in vaccine hesitancy as expressed in social media, by analyzing a large volume of Facebook pages, posts and comments about vaccination. Though Facebook is currently diminishing its share in the social media landscape, it is still the most used social media in the general population and is a well-known target of misinformation campaigns . Recent efforts by Facebook Inc. to tackle fake news have certainly decreased their proliferation by adding a quality control[^1] However, the large historical volume of posts and conversations remaining in many of its public pages constitutes a relevant source of data for studying the anti-vaccination discourse.\
Here we analyzed a set of \(607,105\) comments in \(186\) pro-and anti-vaccination pages in the period Jan, 2012-Jun 2019. A summary of the dataset is shown in the upper part of Table [1](#statistics){reference-type="ref" reference="statistics"}. As many posts bring up discussions between pro-vax and anti-vax users, we can find comments from both user groups interlaced among the replies. In order to tell apart both groups, we trained a classifier that, based on the comment's text, can predict the type of attitude that it denotes: pro-vax, anti-vax or non-specified. Our classifier is composed of three parallel stages: an LSTM neural network, a named entity recognition stage based on TAGME and a page class stage. The outputs are finally concatenated and used as input to a dense layer with a tanh activation function predicting which of the \(3\) groups the comment belongs to. The architecture, depicted in Figure FIG, was implemented in Keras , and its design and configuration is inspired by . The size of the LSTM hidden state was configured in 100 and the final dense layer has size 3. Words were transformed using the 100-dimensional pre-trained GloVe embeddings . The ground-truth for training was composed of \(3000\) comments which were manually tagged as pro-vax, anti-vax or non-specified. The cross-validated prediction results on this set are shown in Figure [\[figuras\]](#figuras){reference-type="ref" reference="figuras"} (top-left) in terms of the ROC curve (Receiver Operating Characteristic). We chose the AUC measure as a metric for our model due to its high unbalance; The baseline accuracy for the area under this curve is 0.50, which represents the accuracy of a random classifier (coin flip). We obtained a cross-validated AUC (area under the ROC curve) of \(0.84\)
[\[table_presence_valence\]]{#table_presence_valence label="table_presence_valence"}
We applied this trained model to the unlabeled part of the dataset, and performed a characterization of the moral traits expressed in pro-and anti-vax comments. For each pro-and anti-vax comment in the training set, we annotated the presence of each of the morals, and in case a moral was present, we annotated the polarity of its content (i.e., if the moral was positively and/or negatively expressed). For example, a comment expressing that "unvaccinated children shouldn't be allowed to go to public school" has a negative polarity of the liberty foundation. Instead, a comment pointing out that "each person has the right to choose if they should be vaccinated" expresses the same moral, but with a positive polarity. Using these samples, we trained one LSTM model for predicting the presence of each moral, and another one for predicting their positive and/or negative polarity. After applying the trained model to the full dataset, the percentage of comments predicted as expressing each moral is detailed in Table [\[table_presence_valence\]](#table_presence_valence){reference-type="ref" reference="table_presence_valence"}. The performance of the classifiers is shown in Table [\[table_rocs\]](#table_rocs){reference-type="ref" reference="table_rocs"} in terms of the AUC and illustrated in Figure [\[figuras\]](#figuras){reference-type="ref" reference="figuras"}. We observe that some of the AUC's are quite low (we consider that an AUC is acceptable if it is above 0.65, which represents a clear distinction from a random classifier). This might be due to the low number of training samples and the difficulty of the learning task (consider that the expression of a moral might be underlying and not quite explicit in a comment). However, among those morals whose prediction is acceptable, we observe that pro-vax comments are more commonly expressing anti-liberty arguments, while anti-vax comments express pro-liberty ones; also, these expressions are much more common than in comments predicted as non vaccination-related. We also observe that purity is expressed more frequently with a negative polarity by pro-vax comments than by anti-vax comments. Interestingly, these two facts were precisely the ones found by as the main differences in morals between pro-vax and anti-vax respondents in their study on vaccination hesitancy. Finally, and though not as significant as the previous ones, we find that pro-vax comments express more frequently a positive polarity of authority, while anti-vax comments express more frequently a negative value of it. Indeed, several studies have shown that persuasive appeals in line with the moral traits endorsed by an audience can shift their attitudes on several issues. This kind of experiments framed by the Moral Foundations Theory has been performed, e.g., in the context of environmental consciousness and donation to charities . We think that these results might help designing and framing vaccination campaigns by focusing on the liberty and authority traits (e.g., by remarking that vaccination gives people more freedom, by allowing them to perform their activities with safety, or by giving a clear picture of how scientific studies are validated and can be reproduced by peers). This can be of particular interest in the current COVID-19 pandemic context, in which misinformation campaigns are putting into risk the herd immunity goal and probably constitute the highest threat to global public health .
[^1]: In particular, in March 2019 Facebook took a series of measures to specifically address vaccine misinformation (<https://about.fb.com/news/2019/03/combatting-vaccine-misinformation/>). | {'timestamp': '2022-06-06T02:15:19', 'yymm': '2206', 'arxiv_id': '2206.01598', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01598'} |
null | null |
null | null |
null | null |
null | null |
# Introduction {#sec1}
A common experimental design in clinical studies, especially longitudinal ones, is the matched pairs design where observations are made from the same subjects under two different conditions, often at two points in time, before and after treatment, or after a fixed period of time from the baseline. Revealing the possible dependence structure of the observations is a major goal. However, a typical issue when dealing with paired data is the occurrence of missing data. The challenge is to exploit the available data to perform valid inference. The literature on matched pairs with missing data has primarily focused on one-dimensional, continuous, discrete or ordinal variables, aimed at detecting changes in location/mean, scale/variance, and distribution. Some of the proposals apply multiple imputation techniques, but they often require large sample sizes for being correct. Other proposals rely on specific model assumptions such as symmetry or bivariate normality, but they exhibit a non-robust behavior against deviations. The common approach recently adopted in literature results from combining in a non parametric approach separate test statistics for the paired and unpaired observations, by using either weighted test statistics, a multiplication combination test, or combined p-values The recent scientific and technological progress in measuring biological processes has enabled monitoring of patient's condition with a growing level of detail and complexity. Thus, beyond the ongoing identification of univariate biomarkers, new complex data structures are being incorporated into the analysis, as is the case of population ages and mortality distributions, distributions of functional connectivity patterns in the brain, post-intracerebral hemorrhage hematoma densities, graph-based representations of connectivity and functional brain activity, and glucose distributions from continuous monitoring. The aim of the present paper is to provide a statistical test for matched pairs with missing data which does not require any parametric assumptions and uses all observations available. We propose new maximum mean discrepancy (MMD) estimators to achieve this aim. The energy distance and the MMD are two equivalent statistical metrics with the ability to detect distributional differences between random samples. Moreover, MMD-based statistics can also be seen as a natural generalization of the ANOVA test to cases where the distributions are not necessary Gaussian. MMD overcomes Gaussian assumptions by representing distances between distributions as distances between mean embeddings in a reproducing kernel Hilbert space (RKHS). MMD has been successfully applied to independence testing, two-sample testing, survival analysis, or clustering analysis. Besides conducting an extensive simulation study, the new testing procedures are applied to the AEGIS diabetes dataset, resulting from a longitudinal population-based study. This dataset includes data from continuous glucose monitoring (CGM), performed at the beginning of the study and five years later. Importantly, there is a substantial loss to follow up. A distributional representation of glucose concentration summarizes several days of monitoring, providing a personal signature of glucose homeostasis. The present approach allows us to address some interesting questions related to the possible changes in CGM profile with ageing, or the relation between obesity and diabetes. Furthermore, an adaption of a previous clustering method to matched pairs with missing data allows us to find out specific patient phenotypes, with potential applications in patient stratification. The rest of this paper is outlined as follows. In Section [2](#sec:glucodensities){reference-type="ref" reference="sec:glucodensities"} we provide a motivation for the new methods from the distributional representation of CGM data. In Section [3](#sec:methods){reference-type="ref" reference="sec:methods"} we define the problem in general terms and introduce the statistical model based on the MMD metric, providing weighted test statistics for dealing with missing data under MCAR mechanism (Section [3.1](#sec:mcar){reference-type="ref" reference="sec:mcar"}) and under MAR mechanism (Section [3.2](#sec:mar){reference-type="ref" reference="sec:mar"}). A proof presenting theoretical guarantees of the proposed methods is delivered in the appendix. In Section [3.3](#sec:kernel_choice){reference-type="ref" reference="sec:kernel_choice"} the choice of kernel functions and corresponding hyperparameters is discussed. Then we present the results of an extensive simulation study in Section [3.4](#sec:simulation_study){reference-type="ref" reference="sec:simulation_study"}. In Section [3.5](#sec:cluster){reference-type="ref" reference="sec:cluster"} a previous clustering method is adapted to missing data under the MAR mechanism. We present in Section [4](#sec:aegis){reference-type="ref" reference="sec:aegis"} some applications of both hypothesis testing and clustering analysis to the AEGIS study, by exploiting the distributional representation of CGM data. We close with a discussion in Section [5](#sec:discussion){reference-type="ref" reference="sec:discussion"}.
# Motivation: a distributional representation of continuous glucose monitoring data {#sec:glucodensities}
Distributional data analysis is a novel methodology that has proved successful to manage biosensor data in different settings such as the connectivity analysis of the brain network, the diabetes management, and the physical activity analysis. In a previous paper, we introduce a novel distributional representation for CGM data, termed glucodensity, which allows us to obtain a functional profile of patient glucose homeostasis. Glucodensity is a natural extension of Time in Range (TIR) metrics, that measures the proportion of time a person spends with their blood glucose levels within the target range of \(70-180\) mg/dL . Although very intuitive, TIR metrics have two main disadvantages: first, the range fits poorly depending on the characteristics of the population examined; second, there is a loss of information caused by the discretization of the recorded data into intervals. Instead, glucodensity effectively measures the proportion of time each individual spends at a specific glucose concentration. Previous results for glucodensities show a better predictive performance as compared to common diabetes biomarkers. Given a series of CGM data \(\{Y_j\}_{j=1}^m\), the glucodensity can be modeled as a probability density function \(f(\cdot)\) that can be approached by kernel density estimation: \[\hat f(y)=\frac{1}{m}\sum _{j=1}^{m} \frac{1}{h} K\left(\frac{Y_j-y}{h}\right),\] where \(h>0\) is the smoothing parameter and \(k(\cdot)\) denotes a non-negative real-valued integrable function (Figure [\[fig:gluco_overview\]](#fig:gluco_overview){reference-type="ref" reference="fig:gluco_overview"}). Let \(\mathcal{D}\) be the space of probability density functions \(f\) such that \(\int _{\mathbb{R}}u^2f(u)du <\infty\). In order to measure the difference between two glucodensities, \(f\) and \(g\), a metric on \(\mathcal{D}\) is required. We use the \(2-\)Wasserstein distance: \[d^2_{\mathcal{W}_2}(f,g)= \int_{0}^{1} \left|Q_{f}\left(t\right)-Q_{g}\left(t\right)\right|^{2}dt, \quad f, g\in \mathcal{D}, \label{eq:wasserstein}\] where \(Q_{f}\) and \(Q_{g}\) denote the corresponding quantile functions. The \(2\)-Wasserstein distance in ([\[eq:wasserstein\]](#eq:wasserstein){reference-type="ref" reference="eq:wasserstein"}) depends only on quantile functions, so we can approximate it by computing empirical quantile functions \(\hat{Q}_f=\hat{F}^{-1}\) and \(\hat{Q}_g=\hat{G}^{-1}\), from the corresponding empirical cumulative distribution functions \(\hat{F}\) and \(\hat{G}\). Figure [\[fig:glucodensity_changes\]](#fig:glucodensity_changes){reference-type="ref" reference="fig:glucodensity_changes"} contains an example of the glucodensity representation for the continuous glucose monitoring performed on three different individuals, both in a prediabetes and later diabetes status. This figure immediately poses the challenge of defining new statistical methods to compare two sets of glucodensity measurements to assess whether some population statistics differ. This can be useful to compare the glucose homeostasis before and after a treatment, or after a certain period of time.
# Hypotheses and statistics {#sec:methods}
Let \(\mathcal{D}\) be a separable Hilbert space and \(\left(X_{1},X_{2}\right)^\top \in \mathcal{D}^2\) a random pair representing two different measurements on a subject at two different time points. Let us consider a general matched pairs design given by i.i.d. random variables \[\mathbf{X}_j=\begin{pmatrix} X_{1j} \\ X_{2j} \end{pmatrix}, ~ j=1...,n.\] To continue the example of our motivation, \(X_{1j}\) can represent the glucodensity at the beginning of a certain study for the j-th patient, and \(X_{2j}\) the glucodensity at the end of the study for the same patient. Let assume that both \(\{X_{1j}\}_{j=1}^n\) and \(\{X_{2j}\}_{j=1}^n\) are drawn from probability measures \(P_1\) and \(P_2\), respectively. We are interested in testing the equality of distributions as null hypothesis \(H_{0}:\{P_{1} = P_{2}\}\) against the alternative \(H_{1}:\{ P_{1} \neq P_{2}\}\), i.e., to check whether there are systematical differences between the outcomes at different time points.
## Missing Completely at Random (MCAR) mechanism {#sec:mcar}
When some of the elements of the matched pairs are missing completely at random the available data can be sorted as: \[\mathbf{X}=\underbrace{\begin{pmatrix} X_{11} \\ X_{21} \end{pmatrix} \cdots \begin{pmatrix} X_{1n_1} \\ X_{2n_1} \end{pmatrix}}_{ \substack{\text{Complete data \(\mathbf{X}^{\text{com}}\)} \\ \text{\(n_1\) observations}}} \hspace{0.2cm} \underbrace{\begin{pmatrix} X_{1n_{1+1}} \\-\end{pmatrix} \cdots \begin{pmatrix} X_{1n_1+n_2} \\-\end{pmatrix}}_{ \substack{\text{Incomplete data \(\mathbf{X}^{\text{inc}}_1\)} \\ \text{\(n_2\) observations}}} \hspace{0.2cm} \underbrace{\begin{pmatrix}-\\ X_{2 n_1+n_2+1} \end{pmatrix} \cdots \begin{pmatrix} -\\ X_{2 n_1+n_2+n_3} \end{pmatrix}}_{ \substack{\text{Incomplete data \(\mathbf{X}^{\text{inc}}_2\)} \\ \text{\(n_3\) observations}}}, \\ \] where \(n=n_1+n_2+n_3\). For ease of notation, we denote \(\mathbf{X}^{\text{com}}_1=\{X_{1j}\}_{j=1}^{n_1}\), \(\mathbf{X}^{\text{com}}_2=\{X_{2j}\}_{j=1}^{n_1}\), \(\mathbf{X}^{\text{inc}}_1=\{X_{1j}\}_{j=n_1+1}^{n_1+n_2}\) and \(\mathbf{X}^{\text{inc}}_2=\{X_{2j}\}_{j=n_1+n_2+1}^{n}\). Additionally, a missingness status variable can be defined \(\delta_{ij}\in \{0,1\}\), \(i=1,2\), \(j=1,\dots, n\), so \(\delta_{ij}=1\) if the element is missing and \(\delta_{ij}=0\) otherwise. A natural way of testing the equality of distributions is measuring the distance between them. We propose two test statistics: \(\mathcal{T}_1\) for the complete data sets \(\mathbf{X}^{\text{com}}_1\) and \(\mathbf{X}^{\text{com}}_2\), and \(\mathcal{T}_2\) for the incomplete data sets \(\mathbf{X}^{\text{inc}}_1\) and \(\mathbf{X}^{\text{inc}}_2\), which are then combined in one weighted test statistic: \[\mathcal{T}(\mathbf{X})= \alpha \mathcal{T}_1(\mathbf{X}^{\text{com}}_1,\mathbf{X}^{\text{com}}_2)+(1-\alpha) \mathcal{T}_2(\mathbf{X}^{\text{inc}}_1,\mathbf{X}^{\text{inc}}_2), \label{eq:test_statistic}\] for some weighting parameter \(\alpha \in \left[0,1\right]\). Both \(\mathcal{T}_1\) and \(\mathcal{T}_2\) are based on the maximum mean discrepancy (MMD) to measure the empirical distance between the marginal distributions. Let \(k: \mathcal{D}\times \mathcal{D} \to \mathbb{R}^{+}\) be a symmetric definite positive kernel. The existence of a dot product space \(\mathcal{H}\) and feature mapping \(\phi: \mathcal{D} \to \mathcal{H}\) is guaranteed, such that \(k(X,X')=\langle \phi(X),\phi(X')\rangle_{\mathcal{H}}\). A reproducing kernel of \(\mathcal{H}\) is a kernel function that satisfies (1) \(\forall X\in \mathcal{D}\), \(k(\cdot,X)\in \mathcal{H}\), and (2) \(\forall X\in \mathcal{D}\), \(\forall g\in \mathcal{H}\), \(\langle g,k(\cdot,X) \rangle_{\mathcal{H}}=g(X)\). \(\mathcal{H}\) is then said to be a reproducing kernel Hilbert space (RKHS). Kernel mean embedding results from extending the mapping \(\phi\) to the space of probability distributions by representing each distribution as a mean function \(\phi(F) = \mathbf{E}[k(\cdot,X)] = \int_{\mathcal{D}} k(\cdot,X)dP\). The kernel mean embedding can be empirically estimated by \(\widetilde{\phi}= \frac{1}{n}\sum_{i=1}^{n}k\left(\cdot, X\right)\). Then, we can measure the distance between random samples as follows: \[\begin{aligned}
\mathcal{T}_1\left(\mathbf{X}^{\text{com}}_1,\mathbf{X}^{\text{com}}_2\right)&= \norm{\widetilde{\phi}^{\text{com}}_{1}-\widetilde{\phi}^{\text{com}}_{2}}_{\mathcal{H}}^{2}\\ &= \langle \frac{1}{n_1}\sum_{i=1}^{n_1}k\left(\cdot, X_{1i}\right)-\frac{1}{n_1}\sum_{i=1}^{n_1}k\left(\cdot, X_{2i}\right), \frac{1}{n_1}\sum_{i=1}^{n_1}k\left(\cdot, X_{1i}\right)-\frac{1}{n_1}\sum_{i=1}^{n_1}k\left(\cdot, X_{2i}\right) \rangle \\ &=\frac{1}{n_1^{2}} \sum_{i=1}^{n_1}\sum_{j=1}^{n_1} k\left(X_{1i},X_{1j}\right)+\frac{1}{n_1^{2}} \sum_{i=1}^{n_1}\sum_{j=1}^{n_1} k\left(X_{2i},X_{2j}\right)-\frac{2}{n_1^{2}} \sum_{i=1}^{n_1}\sum_{j=1}^{n_1}k\left(X_{1i},X_{2j}\right). \end{aligned}\] Analogously, \[\begin{aligned}
\mathcal{T}_2\left(\mathbf{X}^{\text{inc}}_1,\mathbf{X}^{\text{inc}}_2\right) &= \norm{\widetilde{\phi}^{\text{inc}}_{1}-\widetilde{\phi}^{\text{inc}}_{2}}_{\mathcal{H}}^{2} \\ &=\frac{1}{n_2^{2}} \sum_{i=n_1+1}^{n_1+n_2} \sum_{j=n_1+1}^{n_1+n_2} k\left(X_{1i}, X_{1j}\right)+\frac{1}{n_3^{2}} \sum_{i=n_1+n_2+1}^{n_1+n_2+n_3} \sum_{j=n_1+n_2+1}^{n_1+n_2+n_3} k\left(X_{2i}, X_{2j}\right)\\ &\phantom{=}\,\,-\frac{2}{n_2n_3} \sum_{i=n_1+1}^{n_1+n_2} \sum_{j=n_1+n_2+1}^{n_1+n_2+n_3} k\left(X_{1i},X_{2j}\right). \end{aligned}\] Importantly, for the class of *characteristic* kernels, the embeddings are injective, and hence \(\norm{P_{1}-P_{2}}_{\mathcal{H}}^{2} = 0\), if and only if \(P_1 = P_2\). In order to calibrate the tests under the null hypothesis it should be pointed out that both \(\mathcal{T}_1\) and \(\mathcal{T}_2\) do not follow a free asymptotic distribution. The empirical estimate of MMD is a one-sample V-statistics and hence asymptotic distribution is difficult to obtain due to the degeneracy of V-statistics, which incorporates a correlation structure for the complete paired observations \(\mathbf{X}^{\text{com}}\). To address this issue we propose a wild bootstrap procedure for the first \(n_1\) observations, while the remaining \(n_2+n_3\) observations can be properly handled by permutations methods, that can achieve an exact type I error control. For each \(b=1,\dots,B\), it proceeds as follows:
1. For the first \(n_1\) complete paired observations, take random weights \(w^b_{i}\), \(i=1,\dots, n_1\), with \[\begin{aligned}
w^b_{i}= e^{-1/l_{n_1}}w^{b}_{i-1}+\sqrt{1-e^{-2/l_{n_1}}}\epsilon_i, \end{aligned}\] where \(w^b_0,\epsilon_1,\cdots, \epsilon_{n_1}\) are independent standard normal variables, and \(l_{n_1}\) is a bootstrap parameter used to mimic the dependence structure, such that \(l_{n_1}= o\left(n_1\right)\) but \(\lim_{n_1\to \infty} l_{n_1}=\infty\). Then, \[\begin{aligned}
\mathcal{T}^{b}_1\left(\mathbf{X}^{\text{com}}_1,\mathbf{X}^{\text{com}}_2\right)= \frac{1}{n_1^{2}} \sum_{i=1}^{n_1}\sum_{j=1}^{n_1} w^b_{i} w^b_{j} \left[ k\left(X_{1i},X_{1j}\right)+ k\left(X_{2i},X_{2j}\right)-2k\left(X_{1i},X_{2j}\right) \right]. \end{aligned}\]
2. The remaining \(n_2+n_3\) observations belonging to \(\mathbf{X}^{\text{inc}}_1\) and \(\mathbf{X}^{\text{inc}}_2\) are randomly permuted, i.e. each observation is randomly assigned to new \(\mathbf{X}^{\text{inc},\pi}_1\) or \(\mathbf{X}^{\text{inc},\pi}_2\) sets, resulting in new \(\mathcal{T}_2^b(\mathbf{X}^{\text{inc},\pi}_1,\mathbf{X}^{\text{inc},\pi}_2)\).
3. Then, calculate \[\begin{aligned}
\mathcal{T}^{b}= \alpha \mathcal{T}^{b}_1 \left(\mathbf{X}^{\text{com}}_1,\mathbf{X}^{\text{com}}_2 \right)+(1-\alpha)\mathcal{T}_2^b\left(\mathbf{X}^{\text{inc},\pi}_1,\mathbf{X}^{\text{inc},\pi}_2\right) \end{aligned}\]
Finally, return \(p\)-value\(= \frac{1}{B} \sum_{b=1}^{B} 1\{ \mathcal{T}^{b} \geq \mathcal{T}(\mathbf{X})\}\).
## Missing at Random (MAR) mechanism {#sec:mar}
We assume a MAR mechanism where the probability of being missing on the second time point is based on the corresponding value on the first time point, which can be described as follows \[\mathbf{X}=\underbrace{\begin{pmatrix} X_{11} \\ X_{21} \end{pmatrix} \cdots \begin{pmatrix} X_{1n_1} \\ X_{2n_1} \end{pmatrix}}_{ \substack{\text{Complete data \(\mathbf{X}^{\text{com}}\)} \\ \text{\(n_1\) observations}}} \hspace{0.2cm} \underbrace{\begin{pmatrix} X_{1n_1+1} \\-\end{pmatrix} \cdots \begin{pmatrix} X_{1n_1+n_2} \\-\end{pmatrix}}_{ \substack{\text{Incomplete data \(\mathbf{X}^{\text{inc}}_1\)} \\ \text{\(n_2\) observations}}}, \\ \] where \(n=n_1 + n_2\). We denote by \(\pi\left(\cdot\right)= P\left(\delta_{2j}=1 |X_{1j}=\cdot\right)\), the conditional probability that the observation \(X_{2j}\) will be missing given \(X_{1j}\). A natural way to incorporate the missing data mechanism in the test statistic is to associate weight \(\omega_j\) with the \(j\)-th observation via an inverse probability weighting (IPW) estimator, given by \[\omega_j= \frac{\delta_{2j}}{n\pi\left(X_{1j}\right)},~ j = 1,\dots, n. \label{eq:weights}\] In practice, we estimate the probability \(\pi(\cdot)\) by means of a binary classification algorithm. We denote by \(\tilde{\omega}_j\) the estimated weight. We propose the following test statistic \[\begin{aligned}
\mathcal{T}\left(\mathbf{X}\right) &= \mathcal{T}\left(\mathbf{X}^{\text{com}}_1,\mathbf{X}^{\text{com}}_2\right) =\norm{\widetilde{\phi}^{\text{com}}_{1}-\widetilde{\phi}^{\text{com}}_{2}}_{\mathcal{H}}^{2}\\ &= \langle \sum_{j=1}^{n_1} \tilde{\omega}_j k\left(\cdot, X_{1j}\right)-\sum_{j=1}^{n_1} \tilde{\omega}_j k\left(\cdot, X_{2j}\right), \sum_{j=1}^{n_1} \tilde{\omega}_j k\left(\cdot, X_{1j}\right)-\sum_{j=1}^{n_1} \tilde{\omega}_j k\left(\cdot, X_{2j}\right) \rangle \\ &= \sum_{i=1}^{n_1}\sum_{j=1}^{n_1} \tilde{\omega}_i \tilde{\omega}_j k\left(X_{1i},X_{2j}\right) + \sum_{i=1}^{n_1}\sum_{j=1}^{n_1} \tilde{\omega}_i \tilde{\omega}_j k\left(X_{1i},X_{2j}\right)-2 \sum_{i=1}^{n_1}\sum_{j=1}^{n_1} \tilde{\omega}_i \tilde{\omega}_j k\left(X_{1i},X_{2j}\right). \end{aligned}\] In this scenario, we propose to calibrate the test under the null hypothesis in an analogous manner to the MCAR mechanism. Specifically, for each bootstrap iteration we propose to use the following estimator \[\begin{aligned}
\mathcal{T}^{b}\left(\mathbf{X}\right) = \frac{1}{n_1^{2}} \sum_{i=1}^{n_1}\sum_{j=1}^{n_1} w^{b}_{i} w^{b}_{j} \tilde{\omega}_i \tilde{\omega}_j \left[ k\left(X_{1i},X_{1j}\right)+ k\left(X_{2i},X_{2j}\right)-2k\left(X_{1i},X_{2j}\right) \right]. \end{aligned}\]
## Kernel choice and kernel hyperparameters {#sec:kernel_choice}
We propose using the Gaussian kernel \(k\left(X,Y\right)= e^{-\norm{X-Y}^{2}/\sigma^2}\) for \(X,Y \in \mathbb{R}\), and \(k\left(X,Y\right)= e^{-d^{2}_{\mathcal{W}_{2}}\left(X,Y\right)/ \sigma^2}\) for \(X,Y \in \mathcal{D}\), where \(\sigma>0\). Importantly, the Gaussian kernel is a characteristic kernel, and thus we can detect asymptotically any difference in distribution. The kernel bandwidth \(\sigma\) was estimated through the median heuristic \(\sigma ^2= median \{||X_i-X_j||^{2}:1\leq i< j\leq n \}\).
## Simulation study {#sec:simulation_study}
We investigate the finite sample behavior of the above methods in extensive simulations. A total of \(2,000\) simulations were performed for both MCAR and MAR scenarios. Methods were examined with respect to their Type-I error rate control at level 5%. A total of \(2,000\) bootstrap runs and permutation replicas were held. The wild bootstrap parameter \(l_{n_1}\) was selected according to \(l_{n_1}= \sqrt{n_1}\). The observations were generated by mimicking the sort of distributional representations commonly obtained from CGM data. Since the 2-Wasserstein distance depends only on quantile functions, observations were sampled from the following location-scale model on quantile functions: let \(Z\in \mathbb{R}^{p}\) be a random vector of predictor variables and let \(Q_0\) be a fixed quantile function; here we considered the age as the only predictor variable and fixed \(Q_0\left(t\right)= 70+240t\) in the range of glucose values expected from type-2 diabetes; let \(\eta\left(z\right)= a_0+a_1z_1\) and \(\tau\left(z\right)= b_0+b_1z_1\) be the location and scale components of the model, respectively, where \(a= \left(a_0,a_{1}\right)\) and \(b= \left(b_0, b_{1} \right)\) are the corresponding coefficients and we assume that \(\tau\left(Z\right)>0\) almost surely; let \(V_1\) and \(V_2\) two random variables that satisfy \(E\left(V_1|Z\right)=0, E\left(V_2|Z\right)=1\), and \(V_2>0\) almost surely; the model is given by \[Q\left(t\right)= V_1+V_2\eta\left(Z\right)+V_2\tau\left(Z\right)Q_0\left(t\right),\]
### MCAR scenario. {#sec:sim_mcar}
We fixed \(n_1=n_2=n_3=150\). In order to introduce correlation structure into the quantile functions for \(\mathbf{X}^{\text{com}}_1\) and \(\mathbf{X}^{\text{com}}_2\), we sampled variables \(V_1^*\) and \(V_2^*\) from bivariate uniform distributions with correlation given by \(\rho\in \{0.00,0.20,0.40,0.60,0.80\}\). The location-scale model is given by \(V_1=-20+40V_1^*\) and \(V_2= 0.8+0.4V_2^*\), and fixed parameters \(a_0=b_0=0\), \(a_1=0.3\) and \(b_1=0.005\). The observations for \(\mathbf{X}^{\text{inc}}_1\) and \(\mathbf{X}^{\text{inc}}_2\) were i.i.d. generated and then we applied the same location-scale model than before. A total of \(2,000\) simulations were performed assuming that the age was distributed as \(Z_1,Z_2\sim \mathcal{U}_{[30,50]}\) both at the beginning and at the end of the study, that is, for all the variables in \(\mathbf{X}\). Another \(2,000\) simulations were performed assuming that the age was distributed as \(Z_1\sim \mathcal{U}_{[30,50]}\) at the beginning of the study, that is, for all the variables in \(\mathbf{X}^{\text{com}}_1\) and \(\mathbf{X}^{\text{inc}}_1\), and was distributed as \(Z_2\sim \mathcal{U}_{[50,70]}\) at the end of the study, that is, for all the variables in \(\mathbf{X}^{\text{com}}_2\) and \(\mathbf{X}^{\text{inc}}_2\).
### MAR scenario. {#sec:sim_mar}
We fixed \(n=300\). The missing mechanism is given by \(P\left(\delta_{2j}=1|Y_1,Y_2\right)= (1+e^{-1+ Y_1+ Y_2})^{-1}\), \(j=1,\dots, n\), where \(Y_1,Y_2\sim \mathcal{N}\left(0,1\right)\) are two independent random variables. We introduced correlation structure into the quantile functions for \(\mathbf{X}^{\text{com}}_1\) and \(\mathbf{X}^{\text{com}}_2\) as we did in the MCAR scenario. We used the same location-scale model. The same methodology as in the MCAR scenario was applied for sampling the age.
### Results.
Table [1](#table:tabla1){reference-type="ref" reference="table:tabla1"} shows the results of the simulation study. We can see the test calibration under the null hypothesis is acceptable. However, there are some biases in the two situations due to missing data mechanisms. As the discrepancy of the null hypothesis increases, the test rejects the more null hypothesis, and in some cases, it is clear the consistency of the new methods with \(100\) percent of reject cases.
## Paired missing data clustering {#sec:cluster}
Let \(\mathbf{X}=\left\{\left(X_{1j}, X_{2j}, \delta_{2j}\right)\right\}^{n}_{j=1}\), be a dataset of i.i.d. random variables obtained under a MAR mechanism, where we denote again by \(\pi\left(\cdot\right)= P\left(\delta_{2j}=1 |X_{1j}=\cdot\right)\), the probability that the observation \(X_{2j}\) will be missing. We associate a weight \(\tilde{\omega}_j\) with the \(j\)-th observation via an IPW estimator, by applying equation ([\[eq:weights\]](#eq:weights){reference-type="ref" reference="eq:weights"}). Let \(\mathbf{X}_j=(X_{1j},X_{2j})^{\top}\), \(\mathbf{X}_h=(X_{1h},X_{2h})^{\top} \in \mathbf{X}^{\text{com}}\) be two different complete paired samples. We define the following bivariate kernel \(k\left(\mathbf{X}_j,\mathbf{X}_h\right)= e^{-\left(d^{2}_{\mathcal{W}_2}(X_{1j},X_{1h})+d^{2}_{\mathcal{W}_2}(X_{2j},X_{2h})\right)/\sigma^2}\), where \(\sigma^2= median\{d^{2}_{\mathcal{W}_2}\left(X_{1j},X_{1h}\right)+d^{2}_{\mathcal{W}_2}\left(X_{2j},X_{2h} \right): 1\leq j<h\leq n \}\). Consider a disjoint partition \(\mathbf{X}^{\text{com}}=\bigcup _{i=1}^k C_i\), with \(C_i\cap C_l=\emptyset\), for all \(i\neq l\). Following, we aim to build a new partition \(\tilde{C}_1,\dots, \tilde{C}_k\) by maximizing an objective function given by \[\left(\tilde{C}_1,\dots, \tilde{C}_k\right)= \arg \max_{\left(C_1,\dots,C_{k}\right)} \sum_{i=1}^{k} \frac{1}{v_i} \sum_{\mathbf{X}_{j},\mathbf{X}_{h} \in C_i} \tilde{\omega}_j \tilde{\omega}_h k\left(\mathbf{X}_{j},\mathbf{X}_{h}\right),\] where \(v_i= \sum_{\mathbf{X}_{j}\in \mathcal{C}_i} \tilde{\omega}_j\). We can iteratively solve this optimization problem by measuring the impact of moving each observation to another cluster. Let denote by \(S_i= \sum_{\mathbf{X}_{j},\mathbf{X}_{h}\in C_i} \tilde{\omega}_j \tilde{\omega}_h k(\mathbf{X}_{j},\mathbf{X}_{h})\) the internal similarity of cluster \(C_i\), and \(S_i\left(\mathbf{X}_{j}\right)= \sum_{\mathbf{X}_{h} \in C_i} \tilde{\omega}_j \tilde{\omega}_h k(\mathbf{X}_{j},\mathbf{X}_{h})\) the internal similarity with respect to the observation \(\mathbf{X}_{j}\). By moving the observation \(\mathbf{X}_{j}\) from cluster \(C_i\) to \(C_l\) we change the result of the objective function by \[\Delta S^{i\to l} \left(X_{2j}\right)= \frac{S^{+}_{l} }{v_l+ \tilde{\omega}_j} + \frac{S^{-}_{i}}{v_i-\tilde{\omega}_j}-\frac{S_{l}}{v_l}-\frac{S_{i}}{v_i},\] where \(S_l^{+}= S_l+2S_l\left(\mathbf{X}_{j}\right)+\tilde{\omega}_j \tilde{\omega}_j k\left(\mathbf{X}_{j},\mathbf{X}_{j}\right)\) is the internal similarity of the new cluster \(C_l\) after the addition of the observation \(\mathbf{X}_{j}\), and \(S_i^{-}= S_i-2S_i\left(\mathbf{X}_{j}\right)+\tilde{\omega}_j \tilde{\omega}_j k\left(\mathbf{X}_{j},\mathbf{X}_{j}\right)\) is the internal similarity of the new cluster \(C_i\) after removing the observation \(\mathbf{X}_{j}\). Ultimately, we compute \(i^{*}= \arg \max_{l=1\dots,k|l\neq i} \Delta Q^{i\to l} \left(\mathbf{X}_{j}\right)\), and if \(\Delta S^{i\to i^{*}}(\mathbf{X}_{j})>0\) we move \(\mathbf{X}_{j}\) to cluster \(C_{i^*}\), otherwise we keep it in \(C_i\).
# Illustrative data analysis {#sec:aegis}
As a practical application, we consider an ongoing longitudinal, population-based study by, aimed at analyzing the evolution of different clinical biomarkers related to circulating glucose in a initial random sample of \(1516\) patients over \(10\) years. In addition, a CGM are performed every five years on a randomized subset of patients. Specifically, at the beginning of the study, \(581\) participants were randomly selected for wearing a CGM device for \(3\)-\(7\) days. Out of the total of \(581\) participants, \(68\) were diagnosed with diabetes before the study and \(22\) during the first five years. Table [2](#tab:aegis_table){reference-type="ref" reference="tab:aegis_table"} shows the baseline characteristics of these \(581\) patients grouped by sex. After a five-year follow-up, only 161 participants agreed to perform a second glucose monitoring.
The AEGIS study raises some interesting questions that can be addressed with the present approach. **Changes in CGM profile with ageing**. Some recent works explore the important role of ageing in glucose dysregulation, and the difficulties inherent in maintaining glucose homeostasis as close to normal as possible. The proposed \(\mathcal{T}\)-test gives us the opportunity to examine if there exist statistical differences after five years at a distributional level. We estimate the missing data mechanism by means of logistic regression, using as predictors the age and glycaemic status (normoglycemic, prediabetes or type-2 diabetes) at the beginning of the study and sex of each participant. We applied the \(\mathcal{T}\)-test considering glucodensities at both time points to check the null hypothesis of equality of distributions. We obtained a p-value = \(0.048\), identifying significant differences at both time points. **Obesity in diabetes**. Obesity is a critical risk factor for the development of type-2 diabetes. In order to further characterize this risk subpopulation, we analyzed those normoglycemic subjects with overweight in the AEGIS dataset, by examining again if there exist statistical differences after five years at a distributional level. We applied the \(\mathcal{T}\)-test to check the null hypothesis in the following two subgroups of the normoglycemic population: i) individuals with a body mass index less than \(22 Kg/m^{2}\) (low body mass index); ii) individuals with a body mass index higher than \(22 Kg/m^{2}\) (overweight and obesity). In the first case we obtained a \(p\)-value = \(0.36\), providing no evidence against the null hypothesis, while in the second case we obtained a \(p\)-value = \(0.056\), which can be interpreted as borderline. Figure [\[fig:obesity\]](#fig:obesity){reference-type="ref" reference="fig:obesity"} shows the difference between the quantile curves in these two subgroups. **Patient stratification**. Clustering analysis can be a useful tool for providing distinctive and meaningful patient phenotypes and, consequently, in guiding patient stratification for delivering more personalized care. We applied a clustering analysis to those individuals for whom CGM has been performed at both time points. Figure [\[fig:clustering\]](#fig:clustering){reference-type="ref" reference="fig:clustering"} shows the resulting two clusters. The individuals in cluster \(1\) do not present significant changes between both time points, while some significant differences are noted in cluster \(2\). Table [3](#table:groups){reference-type="ref" reference="table:groups"} shows the baseline clinical characteristics of each cluster. Both groups of individuals have important differences in insulin resistance and glycaemic variability metrics. Importantly, in cluster 2 the average glycaemic characteristics in terms of glycated hemoglobin and fasting plasma glucose are consistent with prediabetes (5.7% \(\leq\) A1c \(\leq\) 6.4% or 100 mg/dl \(\leq\) FPG \(\leq\) 125 mg/dl according to American Diabetes Association guidelines). In contrast, cluster 1 is composed of normoglycemic individuals. Ultimately, clustering results effectively correlates with a significant change in the glycaemic status. Finally, we performed stepwise logistic regression with forward selection to identify which baseline characteristics independently predicted the corresponding group, resulting age, FPG and CONGA. We checked the null hypothesis that each coefficient is equal to zero. Table [4](#table:stepwise){reference-type="ref" reference="table:stepwise"} shows the results of this analysis, identifying FPG and CONGA as the subset of characteristics that best predicted the outcome.
:::
# Discussion {#sec:discussion}
The analysis of paired data with missing values is becoming critical in longitudinal studies, particularly when comparing the participants' condition across different time points. The available methods in the literature are not applicable when data adopt non-vectorial representations, better suited to capture functional, structural or other complex forms of information increasingly common in current medicine. To overcome this limitation we have provided novel methods for hypothesis testing in the presence of complex paired missing data under both MCAR and MAR mechanisms. They are not based on any parametric assumption and use all observations within the matched pairs design. The methods are based on the notion of maximum mean discrepancy, a metric between mean embeddings in a RKHS that can be applied to both Euclidean and non-Euclidean data, with different structured, functional and distributional representations, by an appropriate design of the reproducing kernel. Specifically, the space of probability density functions has been used throughout the text to test the feasibility of this approach. The asymptotic validity of the methods was proven and can be found in the appendix. In an extensive simulation study, the type-I error rate control of the tests has been examined under both MCAR and MAR mechanisms, performing well with different correlation coefficients. The sample size affects the behavior of the tests, since inference in a functional space customarily demands more data than in a vectorial space. Hence a worsening of performance is expected for very small sample sizes. The application of these methods to a real longitudinal, population-based, diabetes study has highlighted some of their capabilities and advantages to explore new clinical findings, by exploiting monitoring information along the continuous range of glucose values. It should be emphasized the robustness of the results, even in an scenario with an important proportion of missing data. Furthermore, a complementary clustering analysis has revealed the effectiveness of this approach to provide an early risk identification with the potential to enable a personalized strategy. In order to simplify the application of these methods, they are freely available on GitHub (https://github.com/). | {'timestamp': '2022-06-06T02:15:00', 'yymm': '2206', 'arxiv_id': '2206.01590', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01590'} |
null | null |
# Introduction {#sec:intro}
Modelling pedestrian behavior in traffic environments is a crucial step in the development and testing of autonomous vehicles (AV) and automated driving systems (ADS). As the environment's most vulnerable road users, misinterpretation of their behavior by an AV can lead to catastrophic consequences. Through rigorous testing of a wide range of traffic scenarios, safe interactions between AVs and pedestrians can be ascertained. A realistic and highly controllable pedestrian simulation model that supports the scenario-based testing process is a valuable tool in testing AV capabilities and responses to critical situations. In this work, we present a hierarchical pedestrian behavior model that incorporates behavior trees to handle high-level decision-making processes and an adapted Social Force Model to drive low-level motion. Our model caters to scenario-based testing as it provides an explicit representation of decision processes, allowing engineers to inject desired pedestrian behaviors into existing scenarios. Through this process, rare and critical situations, which may be scarce or absent in existing data, can be generated and tested on AV systems to evaluate their responses safely in a simulation environment. For example, consider the situation of a pedestrian running into the roadway directly in front of a vehicle, causing a collision. If the goal of this test scenario is to cause a collision independent of the agents' starting positions and velocities, recreation of this scenario is difficult with a simple trajectory-based pedestrian that only follows a set speed profile. However, with our model, we can dynamically adjust the pedestrian's speed to ensure a collision occurs, independent of the vehicle's approaching distance and speed. Pedestrians with pre-defined trajectories would require constant manual adjustments in their positioning and speed profiles whenever changes are made to the vehicle's approaching parameters. visualizes our model recreating this situation in a test simulation scenario. Within the context of traffic simulation, existing microscopic models have a heavy focus on specific interaction scenarios leaving little room for extensibility or they employ a "black-box" or non-deterministic approach where the pedestrian's actions and overall trajectory can not be derived given the scenario set up and inputs. There are existing models, including numerous Social Force and cellular automata models , that demonstrate higher-level behaviors and thought-processes, such as lane formation and bottle-necking; however, they largely do not allow engineers to customize these behaviors or inject specific desired actions into a test scenario. The presented simulation model provides a high degree of control over pedestrian actions and behaviors currently lacking in literature. Our model is explicitly decision-driven, through the use of behavior trees, while still producing realistic low-level movements. Such functionality can greatly benefit the scenario-based testing of AVs, since testers can force rare or possibly dangerous situations involving pedestrians in a safe simulation environment. Aside from singular-concept approaches, multi-layered models have been developed to address pedestrian behaviors in traffic. Conjunctive use of trajectory-planning, rule-based, and Social Force layers have been implementing to assist the design of shared spaces . Another common approach is to combine a popular microscopic model, such as a cellular automaton, with higher-level game theoretic concepts , to construct a multi-layered model in which high and low-level interactions can be handled by separate components. As before, such models lack the flexibility provided by behavior trees, which our model exposes to test engineers as a domain-specific language and allows them to create and customize diverse and realistic scenarios. In summary, our work makes the following contributions:
1. a novel model for pedestrian motion simulation, with (i) highly controllable and customizable decision making via behavior trees, including a catalog of reusable pedestrian maneuvers and decision conditions, and (ii) realistic motion via an adapted Social Force Model;
2. an evaluation of the decision-making realism and motion fidelity of the model on two traffic trajectory data sets, showing the ability to replicate (i) the high-level decisions of the empirical pedestrians with a 98% or better accuracy and (ii) the motion trajectories with an average deviation of 1.36 m.
# Background {#sec:background}
In this section we briefly describe some of the background technology and concepts used in our approach.
## Social Force Model {#subsec:sfm}
The Social Force model (SFM) treats pedestrian motion as if it is based on attractive and repulsive physical forces. Our model incorporates a variation of the classical SFM with additional forces. In its classical version, the SFM applies three main collections of forces to the agent: an attracting force drawing the agent towards their destination, a repelling force from each of the other agents in the scene, and a repelling force from each wall or border in the environment. These three forces are summed into an acceleration equation that describes the pedestrian's change in velocity throughout the scenario's duration. The attracting force, \(\textbf{f}_\text{adapt}\), is responsible for propelling pedestrian \(i\) in the direction \(\textbf{e}_i^0(t)\), pointing towards their destination point. Given a current velocity, \(\textbf{v}_i(t)\), and a desired speed, \(v_i^0(t)\), the attracting force acting on pedestrian \(i\) with mass \(m_i\) is \[\textbf{f}_\text{adapt} = m_i \frac{v_i^0(t)\textbf{e}_i^0(t)-\textbf{v}_i(t)}{\tau_i}\] where their acceleration is over characteristic time \(\tau_i\). Repelling forces acting on the pedestrian are divided into three groups of forces: \(\textbf{f}_\text{otherPeds}\), \(\textbf{f}_\text{vehicles}\), and \(\textbf{f}_\text{borders}\). Each of these forces is composed of a sum of sub-forces directed away from an object (pedestrian, vehicle, or wall). They describe the human tendency to avoid collision with and maintain a comfortable distance from other objects and agents in their surroundings. Further details on the \(\textbf{f}_\text{otherPeds}\) and \(\textbf{f}_\text{borders}\) forces can be found in Helbing et al.'s model and \(\textbf{f}_\text{vehicles}\) is presented by Anvari et al.Ṫhe resulting sum of forces that drives each pedestrian's motion is given by \[\label{eq:sfm} m_i \, \frac{d\textbf{v}_i}{dt} = \textbf{f}_\text{adapt} + \textbf{f}_\text{otherPeds} + \textbf{f}_\text{vehicles} + \textbf{f}_\text{borders}\]
## Behavior Trees {#BT_background}
Behavior trees can be used to concisely and explicitly model a wide range of decision making processes . Within our model, behavior trees are used to select an appropriate maneuver for each pedestrian at each simulation cycle. Each pedestrian contains a personal behavior tree which, at each time step in a given scenario, is "ticked". The ticking process traverses the tree with a certain path and ultimately outputs a selected maneuver. The behavior trees used in the presented model are composed of four types of nodes: *selector*, *sequence*, *maneuver*, and *condition*. Selectors and sequences are internal nodes and control the path of the tree traversal, while maneuvers and conditions are the tree's leaf nodes. Each leaf node, after being evaluated, returns a status to their parent node. Valid statuses are **Success**, **Failure**, and **Running**. These statuses affect which nodes are visited next by the tick. Selector nodes, denoted by a question mark (?), are analogous to a short-circuit OR in that they tick their child nodes sequentially from left to right until a status of **Success** or **Running** is received. This status is then returned to the selector's parent node. If all of the child nodes return **Failure**, then **Failure** is returned to the selector's parent. On the other hand, sequence nodes, denoted by an arrow (\(\rightarrow\)), are similar to a short-circuit AND. They tick their child nodes sequentially from left to right until a status of **Failure** is received (or they run out of child nodes). Maneuvers and conditions are leaf nodes that are evaluated and return one of the three statuses to their internal parent node. The maneuver node that is visited last before the tick returns from the entire tree becomes the selected maneuver. As a simple example, shows a valid behavior tree an agent could use to determine whether to enter a crosswalk. The process flow of this tree first checks if the crossing signal is green. If so, the pedestrian enters the crosswalk and waits at the entrance otherwise. In the case that the crossing signal is not green, the tick returns from the left side of the tree with a status of **Failure** (from the diamond condition node). Since the root is a *selector* node, it proceeds to tick the right side, consisting solely of a single maneuver node which will return a non-failure status and become the selected maneuver.
# Model Design {#sec:model-design}
To introduce our model, we first discuss the design requirements necessary for a practical pedestrian simulation model.
## Facilitation of Scenario Creation
An essential component of scenario-based testing is a straight-forward and clear scenario creation process . Test scenarios involving pedestrians running our model are expressed in GeoScenario . GeoScenario is a domain-specific language for simple and extensible representation of traffic scenarios, built using the Open Street Map (OSM) standard. All of the essential components of a scenario can be represented in GeoScenario, including dynamic agents, static objects, and the underlying Lanelet2 map, among others. Through GeoScenario, pedestrian agents are represented by nodes containing tags to describe their individual attributes. As GeoScenario abides by the Open Street Map standard, we employ the OSM scenario editor tool, JOSM[^1], to create and edit scenarios. JOSM provides engineers with a simple visual tool to quickly add, move, and tag nodes and ways. Pedestrian-specific agents must include tags defining their identifying name, destination point, and the file containing their personal behavior tree.
## Customizable Pedestrian Behavior
Behavior trees are an integral part of each scenario as they provide the tester with a fine-level of control over how each pedestrian behaves in a given context. Each pedestrian is assigned a behavior tree file in the scenario definition file which dictates their decision process at each simulation cycle. Our model comes with a library of reusable behavior trees that represent common behaviors, such as walking along a sidewalk and traversing signalized and unsignalized crosswalks. The explicit nature of behavior trees in representing discrete decision-making processes lends itself well to forcing desired behaviors within scenarios. Testers can intentionally trigger particular actions that can lead to critical or dangerous situations; for example, pedestrians unexpectedly running in front of a moving vehicle. The behavior trees designed for our model are also augmented with a number of tunable parameters that define the different ways pedestrians may execute the same maneuver. When importing behavior trees from a library, their parameters and sub-trees can be overridden to define the desired behavior.
## Dynamic Interactions Between Agents
Dynamic interactions between agents within a scenario are handled by both the Social Force model and behavior tree components of our model, though in different ways. The Social Force model (SFM) is responsible for immediate, reactionary interactions with vehicles and other pedestrians. At the trajectory level, a collection of repulsive forces is applied to the pedestrian when another agent is in its proximity that causes the pedestrian to naturally avoid collisions. Behavior trees are designed to handle higher-level interactions proactively. Depending on their composition, behavior trees can check the relative states of other agents, such as their distance and speed, to output an informed response to another agent's actions. If required by a scenario, behavior trees can selectively and conditionally modify the forces for specific pedestrians by manipulating SFM parameters.
## Realistic Human Movements and Decisions
A practical pedestrian behavior model needs to be able to simulate realistic movements and rational decision-making processes. We evaluate this requirement on our model against two naturalistic data sets with different road structures. Our hierarchical approach of handling high-level decisions with customizable behavior trees that inform low-level trajectory movements is shown to be effective in producing realistic movements and decisions that map to real-world scenarios.
# Model Architecture
The model structure is composed of three layers: the Behavior layer, Maneuver layer, and Motion Planner layer. As an overview, the Behavior layer receives the environment state representation and decides on an appropriate maneuver to execute. This maneuver is then passed to the Maneuver layer, which plans how best to execute the selected maneuver. The layer forms instructions on how to adapt the current trajectory in the form of a vector containing the pedestrian's updated waypoint, direction vector, and desired speed, to pass to the Motion Planner layer. When the Motion Planner layer receives these instructions, it feeds the passed vector into the Social Force Model to determine the state information of the pedestrian for the next time step. Finally, this state information is updated and reflected in the environment. This process flow is visualized in . The simulation traffic environment is represented in the two dimensional Cartesian coordinate frame and all computations for pedestrian movements are calculated within this frame. The physical traffic structure of the world is represented by Lanelet2 compatible map files. Dynamic elements of the environment include the changing states of vehicles, other pedestrians, and traffic lights for both vehicle lanes and crosswalks. We define two types of pedestrian agents: Empirical Pedestrians (EP) and Simulated Pedestrians (SP). Empirical Pedestrians follow a predefined sequence of trajectory points with preset time intervals and do not use our model to drive their motion. Simulated pedestrians do apply the presented model to determine their movements and are the focus of this paper. Agents' state information contains two dimensional positional, velocity, and acceleration components as well as heading in the vector \(|(x, \dot{x}, \ddot{x}, y, \dot{y}, \ddot{y}, \theta)|_t\). In our evaluation, Empirical Vehicle (EV) agents are used in the experiments; however, our pedestrian model does not depend on the particular implementation of vehicle agents, which allows for targeted testing of a desired ADS or AV system. The state vector format is shared between all types of pedestrian and vehicle agents.
## Behavior Layer
The Behavior layer is tasked with determining an appropriate maneuver to execute given the current environmental context. At scenario creation, each SP agent is assigned a text-based behavior tree file representing their personal tree. The Behavior layer consumes the current traffic state estimation as well as configuration parameters for the overall tree and each possible maneuver. The output of this layer is a selected maneuver with its specific configuration parameters, which is subsequently passed to the Maneuver layer. contains the available maneuvers and conditions currently implemented in the model. Behavior trees are constructed with a subset of these lists as leaf nodes. This maneuver and condition catalog was derived manually in the process of developing the behavior trees required to simulate each of the pedestrians in the data sets discussed in Section [6](#sec:evaluation){reference-type="ref" reference="sec:evaluation"}.
As explained and visualized in Section [2.2](#BT_background){reference-type="ref" reference="BT_background"}, the internal nodes of a behavior tree are sequence and selector nodes and the leaf nodes are maneuver, condition, and, optionally, sub-tree nodes. The internal nodes dictate the traversal path of the "tick" based on the return statuses of leaf nodes. A sub-tree is itself a complete behavior tree that can be "plugged in" in place of a leaf node of a different tree. Sub-trees are simply an extension of their base tree that are useful to maintain modularity and, in our case, add varying levels of behavior to control how different pedestrians react to the same situation. In our model, we have designed three sub-trees, representing three levels of aggressiveness, to define different decision-making processes when a pedestrian is planning to enter a crosswalk. We label the three levels of aggressiveness as Low, Medium, and High. Briefly, low aggressiveness pedestrians only enter the crosswalk when the crossing signal is green (or there is no crossing signal at all). A medium aggressiveness pedestrian enters on green, but also on a yellow signal when they judge that they can sufficiently cross before the red signal. Finally, the pedestrians with a high level of aggressiveness choose to enter the crosswalk regardless of the signal state as long as they are not put in danger by a vehicle in doing so. At each simulation cycle, the behavior tree of each pedestrian is ticked and an appropriate maneuver is selected to be executed by the following layers. This maneuver is passed as input to the Maneuver layer for interpretation.
## Maneuver Layer
The job of the Maneuver layer is to translate a maneuver received from the Behavior layer into instructions on how the pedestrian should adjust their trajectory. The instructions must be interpretable by the subsequent Motion Planner layer. Each received maneuver is converted into low-level instructions containing the following three components: a waypoint, a desired direction unit vector, and a desired speed. A waypoint is defined as an intermediate goal point the pedestrian visits before its final destination. As an example of translating a selected maneuver, if the *Enter Crosswalk* maneuver is received, the Maneuver layer determines the updated waypoint to be a point at the end of the crosswalk, the desired direction vector to be the unit vector pointing to the new waypoint, and the desired speed to be the default desired speed of the pedestrian. These three components are passed to the next layer, the Motion Planner layer, which handles the execution of the selected maneuver.
## Motion Planner Layer
The Motion Planner layer is driven by an adaptation of the Social Force Model. It receives the three components passed by the Maneuver layer describing changes to the pedestrian's trajectory. This layer also has access to the traffic state estimation and therefore all positions and velocities of the other agents in the scene. Our model implements the classic SFM described in Section [2.1](#subsec:sfm){reference-type="ref" reference="subsec:sfm"} extended by Anvari's method for handling pedestrian-vehicle interactions. The parameters of our SFM were manually calibrated through simulation testing. Our traffic environments in Section [6](#sec:evaluation){reference-type="ref" reference="sec:evaluation"}) did not contain any walls, and thus the wall forces are currently not included. Walking within a sidewalk or crosswalk's bounds is handled through the *Keep in Lane* maneuver, which uses a desired walking direction vector (that may not necessarily point to the waypoint) to guide the pedestrian within the sidewalk or crosswalk (i.e., a pedestrian lane) and bias it along the right or left lane-boundary. Wall forces may be added for future environments containing physical walls. The Motion Planner layer is tasked with running an iteration of the Social Force model's formula [\[eq:sfm\]](#eq:sfm){reference-type="eqref" reference="eq:sfm"} to determine the model pedestrian's change in velocity at each simulation time step. Subsequently, this layer directly updates the pedestrian's state information before the next simulation cycle.
# Implementation
The presented model is written in Python and integrated into GeoScenario Server, a full scenario simulation environment capable of running traffic scenarios as a standalone application. The Server provides the necessary infrastructure for full scenario simulation. It parses GeoScenario scenario files and initializes the necessary elements (pedestrians and vehicle agents, agent goal points, traffic lights, etc.), reads and loads the Lanelet2 map file to provide the underlying road network structure for the scenario, maintains the static and dynamic environment and facilitates the information flow between agents and their surrounding environmental context, and finally runs the scenario by iterating through each of its agents to update their state in the environment at each simulation cycle. Vehicles are simulated with the SDV model from Queiroz et al. . Pedestrians originally used a simple model with Pre-defined Trajectories (PDTs). We extended the server by adding our pedestrian model as an alternative to replace the PDT-based model. GeoScenario Server provides the optional integration with WISE Sim, a simulator based on UnrealEngine[^2] to run the WISE Automated Driving System. The server incorporates a shared-memory interface with WISE Sim to integrate its dynamic agents and scenario environment into the simulator, but it also features an experimental integration with the CARLA simulator . More details on the GeoScenario Server and how to use the pedestrian model can be found at <https://geoscenario2.readthedocs.io>.
# Evaluation {#sec:evaluation}
We assess our model in terms of how well it can reproduce low-level trajectories and high-level decisions observed in a naturalistic data set, as well as its extensibility when it is applied to environments with different road structures and geometries.
## Evaluation Scenario Configuration
To approach and assess our evaluation criteria, we need a standardized process for comparing a simulated pedestrian generated by our model against a real-world pedestrian from a naturalistic data set. We devise a process to create *evaluation scenarios*. These generated scenarios assist in validating our model against real-world data and provide a standard process that can be applied to any data set in a compatible format. The idea of an evaluation scenario is to replace a single pedestrian in a given traffic recording with a simulated pedestrian and observe how it interacts with other pedestrians and whether it follows the same trajectory as the empirical pedestrian it replaced. In terms of a GeoScenario scenario, one pedestrian is selected as the *evaluation pedestrian* and is created as an SP agent while all other pedestrians and vehicles in the recording are created as EP and EV agents respectively, and follow their corresponding trajectories from the data set. We refer to the data set pedestrian that the SP agent replaced as the *empirical pedestrian* or alternatively, the evaluation pedestrian's *empirical counterpart*. The scenario begins when the evaluation pedestrian enters the recording and ends when they exit. The SP agent is initialized with three pieces of knowledge about its empirical counterpart: its starting position, its last position, and its average walking speed. In terms of the SFM component of the model, the last position is set as the SP's destination point and the average walking speed is set as the SP's desired speed. The evaluation scenario creation process is repeated for each individual pedestrian in the data set, resulting in a unique evaluation scenario for each real-world pedestrian. As a result, each pedestrian in the data set produces their own evaluation scenario in which they are replaced by a model pedestrian that dynamically interacts with the other agents. This process compiles a suite of evaluation scenarios on which we can perform analysis and draw conclusions in terms of our evaluation criteria. Two separate naturalistic data sets from different locations in Ontario, Canada, were used in the evaluation process.[^3] Each data set contains video files recorded by an overhead drone. The video files were then analyzed and relevant information, such as road user trajectory tracking and traffic light timings, was extracted and saved into a database. The first data set, referred to as the *intersection* data set, was recorded at a busy four-way intersection with four signalized pedestrian crosswalks and two additional unsignalized crosswalks, each one across a right-turn merge lane (or slip lane). The second, referred to as the *single crosswalk* data set, contains a single unsignalized crosswalk across a two-way two-lane road at a university. During evaluation, we noted that the scenarios at the intersection location had relatively longer durations with an average of 66.82 seconds. Due to the minimal knowledge about the empirical trajectory, a concern arose that small deviations in the trajectories early in the scenario may amplify and compound into large deviations further into the scenario. These large deviations may not be representative of the model's performance at each moment in time and may be consequences of previous error. To mitigate this, we introduce segmented scenarios, in which each full evaluation scenario is subdivided into multiple segmented scenarios. Each segmented scenario represents one section of the pedestrian's journey spanning, for example, a single crosswalk or a single segment of sidewalk. For each evaluation scenario from the intersection data set, one or more additional segmented evaluation scenarios are created and grouped separately.
## Realism of Low-Level Trajectories {#subsec:trajectory_metrics}
It is crucial that our behavior model produces pedestrian movements that are as natural and human-like as possible in order to be relied upon as a realistic representation of pedestrians in test scenarios. To evaluate our model's effectiveness at generating realistic motion, we run the evaluation scenarios and record each evaluation pedestrian's generated trajectory. We compare this simulated trajectory generated by our model with the trajectory of the corresponding empirical pedestrian by two trajectory-matching metrics: Euclidean distance and discrete Fréchet distance. Given the dynamic nature of humans, it is unreasonable to expect a single configuration of our model to completely cover the varying behaviors of all the pedestrians in the data set. To accommodate this, we manually search for a custom configuration of parameter values for each evaluation scenario that best matches the corresponding real-world pedestrian's actions and behaviors. The Euclidean distance (ED) measures the average distance between each pair of corresponding points between the simulated and empirical trajectories. The ED metric we use is also known as the spatio-temporal Euclidean distance (STED) . The points of each trajectory are equi-timed and recorded at each simulation time step in the evaluation scenario. On the other hand, the Fréchet distance provides a measure of geometric similarity between the trajectories or paths. It does not depend on the velocity profiles as the Euclidean distance does and instead judges the similarity in "shape" of the two compared trajectories. The majority of test cases during development of our model were derived from the intersection data set. While this data set provides a substantial number of pedestrians exhibiting a wide range of behaviors, a goal of any useful pedestrian model should be extensibility to differing road structures and geometries. We test the generalization and extensibility of our model by extracting evaluation scenarios from the second data set (single crosswalk). We evaluate the model's ability to navigate this new road structure with the same trajectory-matching metrics. Table [2](#tab:trajectory_results){reference-type="ref" reference="tab:trajectory_results"} displays the average ED, the maximum ED, and the average FD over the scenarios on each data set location. To give context to our model's results, we introduce a method to generate a baseline trajectory-approximation for each empirical pedestrian. We produce a baseline trajectory for each evaluation scenario based on the empirical path. We first define a set of points on the intersection map that can be connected in various orders to approximate the empirical paths. The selected map points closely resemble the waypoints a model pedestrian may select to navigate through the intersection. A baseline trajectory is constructed by manually selecting an ordered set of points, beginning and ending with the start and end points, respectively, that best approximates the data set pedestrian's path. Equi-distanced and equi-timed points are then linearly interpolated between these points to form a trajectory with the same number of points as the corresponding empirical trajectory.
It is interesting to note that the baseline method has a slightly improved Fréchet distance over the model's full intersection scenarios. A reasonable explanation for this is that the baseline does not react to other pedestrians along the path, as it is simply a set of distances between waypoints. The model, on the other hand, interacts with other pedestrians and vehicles in their vicinity as do the empirical pedestrians. This introduces the risk of the model choosing a different course of action for their interaction than their empirical counterpart, for example, avoiding oncoming pedestrians on the right side instead of left, resulting in larger points of error. The baseline method, being non-reactive, averages out these deviations and may result in an improved geometric-based metric over a long duration scenario. To visually confirm our model's trajectory-matching effectiveness, we trace both the simulated and empirical pedestrians' trajectories overlaid on a outline of the map file. shows traces of evaluation scenarios at both the intersection location () and the single crosswalk location (). We note that the individual layers of the model are not evaluated independently. The Social Force model and its variations have been shown to effectively model both the individual movements of agents in sparse groups and the crowd dynamics of dense groups . We focus on the conjunctive multi-layer use of high and low-level planners to produce realistic motion.
## Realism of High-Level Decisions
For a pedestrian behavior model to properly represent the behaviors of real-world pedestrians, not only must it model the low-level trajectory movements, but it must also be able to replicate high-level decision-making processes. In order to measure this criterion, we need to define a metric by which we can conclusively declare that a decision made by a model pedestrian is the same decision made by its empirical counterpart. First, we must define a list of decisions to be observed from the naturalistic data. It was determined that there are two notable decisions pedestrians make while navigating an intersection.
1. Given a set of accessible crosswalks and the requirement that at least one crosswalk must be crossed to reach the destination, which crosswalk is selected to cross
2. Given a target crosswalk that the pedestrian has already decided to take which displays a red/yellow crossing light state, will the pedestrian begin to cross or wait until the next green state
The above decisions were designed to ensure a binary response can be recorded for each instance of the decision across evaluation scenarios. The model pedestrian either makes the same decision as the data set pedestrian or they do not. In the rest of this section, we refer to these decisions as Decision 1 and 2 respectively. Note that this metric is only considered on the intersection data set due to the overly simplified road structure of the second data set location. For each full evaluation scenario, we noted the decision points where one of the two listed decisions were made. If the model pedestrian makes the same decision, a *Same* data point is recorded, otherwise we record a *Not Same* data point. displays the results of this process.
We see that our model pedestrians selected the same crosswalk when presented with multiple options as the real-world pedestrian 100% of the time. However, there are three instances where the model pedestrian failed to enter the crosswalk at the same time as defined by Decision 2. After investigation, all three of these instances were due to the real-world pedestrian waiting at a red signal for most of its duration, then, when the vehicles' traffic lights show a two-way advanced green signal, they decide to enter the crosswalk as soon as there are no more conflicting left-turning vehicles. An adjustment to the High level of aggressiveness can account for this specific situation. A range of behaviors can be implemented, including a behavior where the probability of entering the crosswalk on a red signal increases with waiting time.
## Vehicle-Pedestrian Collision Scenario
To showcase our model's practicality in creating critical scenarios, we revisit the scenario outlined in . The goal of this scenario is to demonstrate a plausible yet dangerous scenario that cannot be replicated with simple pedestrians that follow a constant speed profile. We created a scenario with one vehicle agent and one pedestrian agent in which the vehicle wants to pass through an unsignalized crosswalk that the pedestrian wants to cross. The pedestrian dynamically adjusts their walking or running speed to ensure a collision occurs at the crosswalk. At a technical level, this is achieved with our model by including a parameter in the **Keep in Lane** maneuver: *collision_vehicle = \[vehicle_id\]*. Provided that the *vehicle_id* exists in the scenario and there is a crosswalk at which the agents can meet, the pedestrian dynamically determines a collision point based on the point of intersection between the crosswalk and the vehicle's heading vector. With assumed perfect perception of the vehicle's distance and speed, the pedestrian adjusts their own speed to ensure they reach the collision point at the same time as the vehicle. A key benefit to using our model for this scenario is that the pedestrian will ensure a collision occurs regardless of the starting positions and velocities of the agents. With preset speed-profile pedestrians, the tester would need to manually reset the scenario's parameters with either trial and error or complex search-based computations to achieve the same result. Instead, with our approach, no changes need to be applied in order to force a collision scenario at every run.
# Discussion and Limitations {#sec:limitations}
In the current version of our model, we do not consider the implicit or explicit communication between human drivers and pedestrians, such as changes in speed or hand gestures. Such interactions are important considerations in modelling pedestrian movement and behavior. In our evaluation, we do not have access to this communication due to the data sets being recorded by overhead drones. Previous works, such as TrafficSim, have learned pedestrian behavior taking into account all forms of communication and external factors through data-driven approaches. Given the black-box nature of such approaches, we differentiate our work by providing finer control over individual pedestrian behaviors and interactions. An extension of our model could incorporate human-driver interactions through behavior trees, specifically through more complex condition nodes. As long as the communication can be perceived by each agent, the pedestrian can evaluate the interaction and choose an appropriate action accordingly. Adding to the set of available maneuvers and conditions can enhance the complexity and accuracy of the pedestrian's behavior. However, more maneuvers and conditions may also lead to a higher risk of incorrect or dangerous decisions. We must mediate the trade off between a manageable set of behaviors and their coverage and accuracy of real-world decisions, with evaluation on the available real-world data. Engineers must also be cautious of overfitting the behavior trees when designing critical scenarios not found in data. However, some scenario-based testing applications may require forcing a specific behavior to create a desired test scenario. In these cases, it is not so important to construct generalized behavior trees and it is sufficient to simply create the tree that generates the isolated behavior. Since behavior trees allow for any number of parameters within their maneuver and condition nodes, it may be desirable to automate the process of tuning these parameters on real-world data. With separate sets of scenarios for tuning the trees and testing the model, the accuracy and robustness of the behavior trees can be improved. Using real-world data, the tuning process can also introduce the influence of implicit communication methods not previously incorporated into the model. Another limitation of our model is the point-mass representation of pedestrians. The SP model abstracts characteristics from the agents, such as demographic information (age, gender, etc.) and body pose, which may be relevant to decisions for both drivers and pedestrians. For example, an elderly person will tend to move at different speeds and make different crossing decisions to an adult or a child in the same situation. We anticipate more sophisticated representations of pedestrians in future iterations of our model. Finally, our model does not consider the uncertainty and noise of human motion. This is especially relevant in interactions between pedestrians and autonomous vehicles. Though not included in this iteration of the SP model, uncertainty can be injected into behavior trees. Noise can be used in condition nodes to express perception or judgement errors made by humans and in maneuver nodes to produce unexpected movements or decisions that may be difficult for an AV to predict.
# Conclusions {#sec:conclusions}
In this paper, we presented a novel hierarchical pedestrian behavior model that is capable of producing realistic trajectories through different traffic environments while following the rules of the road and making rational real-time decisions, but also allowing for misbehaviors. A multi-layer approach was applied to this problem that incorporates a high-level Behavior layer that determines an appropriate maneuver to be processed by the Maneuver layer, which informs the Motion Planner layer on how to adjust the low-level trajectory movements. The conjunctive use of behavior trees with an adapted Social Force Model ensures the model's agents are an accurate representation of real-world pedestrians. They were shown to make the same decisions when faced with multiple options and also display natural movements and interactions with other pedestrians and vehicles in the scene. Our presented model offers benefit to the scenario-based testing of autonomous vehicles. Since pedestrian decisions and actions can be explicitly represented with behavior trees, engineers are able to inject desired behaviors into scenarios to test the AV's responses to critical situations. We provide an implementation of a set of basic maneuvers and conditions shown to sufficiently cover real-world behaviors. The flexibility and modularity of behavior trees allow for extensions of this list and configurations of trees that cover a wide range of conceivable behaviors. We evaluated our model in terms of its ability to produce realistic low-level movements in two environments with different road structures and to replicate high-level decisions made by real-world pedestrians observed in a naturalistic data set. The results of our evaluation confirm that we present a viable pedestrian simulation model capable of producing realistic pedestrian decisions and movements.
[^1]: <https://josm.openstreetmap.de>
[^2]: <https://www.unrealengine.com>
[^3]: <https://wiselab.uwaterloo.ca/waterloo-multi-agent-traffic-dataset/> | {'timestamp': '2022-06-06T02:15:24', 'yymm': '2206', 'arxiv_id': '2206.01601', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01601'} |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction[\[intro\]]{#intro label="intro"}
TBC
# Governing equations in Eulerian frame of reference.
The filter width is \(\tilde{\Delta}\), which is the discretisation lengthscale (in SPH terms, \(\tilde{\Delta}=2h\)). Following , Lakehal and others, the filtered, Eulerian-Eulerian equations are \[\frac{\partial}{\partial{t}}\left(\alpha_{0}\rho_{0}\right)+\tilde{\bm{u}}_{0}\cdot\nabla\left(\alpha_{0}\rho_{0}\right)=\alpha_{0}\rho_{0}\nabla\cdot\tilde{\bm{u}}_{0}\label{eq:mass_le}\] \[\frac{\partial}{\partial{t}}\left(\alpha_{0}\tilde{\bm{u}}_{0}\right)+\tilde{\bm{u}}_{0}\cdot\nabla\left(\alpha_{0}\tilde{\bm{u}}_{0}\right)=-\frac{1}{\rho_{0}}\nabla\left(\alpha_{0}\tilde{p}\right)+\alpha_{0}\bm{g}+\frac{1}{\rho_{0}}\nabla\cdot\left(\alpha_{0}\bm{\tau}\right)+\bm{M}_{0}\label{eq:mom_le}\] in which \(\bm{\tau}\) is the viscous stress, comprised of both resolved and modelled terms. Discussed later. The dispered bubble phase is represented by a discrete set of \(N_{G}\) groups, each covering a range of bubble radii. Each group \(k\in\left[1,N_{G}\right]\) has characteristic bubble diameter \(d_{k}\) and mass \(m_{k}\). The density of the air is \(\rho_{a}\). The volume fraction of all phases sums to unity: \[\alpha_{0}+\displaystyle\sum_{k=1}^{N_{G}}\alpha_{k}=1\] The volume fraction of the \(k^{th}\) bubble group obeys the evolution equation \[\frac{\partial\alpha_{k}}{\partial{t}}+\tilde{\bm{u}}_{k}\cdot\nabla\alpha_{k}+\alpha_{k}\nabla\cdot\tilde{\bm{u}}_{k}=\frac{m_{k}}{\rho_{a}}\left\{B_{k} + S_{k} + D_{k}\right\};\quad\forall{k}\in\left[1,N_{G}\right]\label{eq:vf_be},\] where XX the terms on the RHS are for bubble production, intergroup mass transfer, and diffusion. The equation for the momentum of the \(k^{th}\) bubble group is \[\frac{\partial}{\partial{t}}\left(\alpha_{k}\tilde{\bm{u}}_{k}\right)+\tilde{\bm{u}}_{k}\cdot\nabla\left(\alpha_{k}\tilde{\bm{u}}_{k}\right)=-\frac{1}{\rho_{a}}\nabla\left(\alpha_{k}\tilde{p}\right)+\alpha_{k}\bm{g}+\bm{M}_{k}\label{eq:mom_be}\] Closed with an equation of state: \(\tilde{p}=c^{2}\rho_{0}\). Note that key in the derivation of the multi-fluid filtered equations in is the criteria that the filter width must be larger than the length scale of the dispersed phase. In the present context, this indicates that the finest resolution of our smoothed particle hydrodynamics scheme must be coarser than the largest bubbles we wish to represent.
## Sub-particle-scale model
Following , we use the dynamic model of , with modification (e.g. ..). We denote the resolved strain rate tensor \(\tilde{\mathcal{S}}=\left(\nabla\tilde{\bm{u}}_{0}+\left(\nabla\tilde{\bm{u}}_{0}\right)^{T}\right)/2\), and express the viscous stress tensor as \[\bm{\tau}=2\mu_{eff}\tilde{\mathcal{S}}\] in which \(\mu_{eff}=\mu+\mu_{S}+\mu_{B}\), with \(\mu_{S}\) the shear induced viscosity, determined via the smagorinsky model, and \(\mu_{B}\) the bubble induced viscosity, from XX. For the divergence free velocity field \(\tilde{\bm{u}}_{0}\), the viscous term \(\frac{1}{\rho_{0}}\nabla\cdot\left(\alpha_{0}\bm{\tau}\right)\) in [\[eq:mom_l\]](#eq:mom_l){reference-type="eqref" reference="eq:mom_l"} becomes \[\frac{1}{\rho_{0}}\nabla\cdot\left(\alpha_{0}\mu_{eff}\nabla\tilde{\bm{u}}_{0}\right)\] The shear induced viscosity is obtained from \[\mu_{S}=\left(C_{s}\tilde{\Delta}\right)^{2}\left\lvert\tilde{\mathcal{S}}\right\rvert,\] with \(C_{s}\) the Smagorinsky coefficient and \(\left\lvert\tilde{\mathcal{S}}\right\rvert\) is the norm of the strain rate tensor. Following XX, a test filter is introduced, with width \(\widehat\Delta\), where \(\widehat\Delta/\tilde{\Delta}>1\), and a \(\widehat{~}\) denotes the operation of the test filter. We introduce the tensors \[L_{ij}=\widehat{\tilde{u}_{0,i}\tilde{u}_{0,j}}-\widehat{\tilde{u}_{0,i}}\widehat{\tilde{u}_{0,i}}\qquad\text{and}\qquad{M}_{ij}=\left(\frac{\widehat\Delta}{\tilde{\Delta}}\right)^{2}\widehat{\left\lvert\tilde{\mathcal{S}}\right\rvert}\widehat{\tilde{\mathcal{S}}_{ij}}-\widehat{\left\lvert\tilde{\mathcal{S}}\right\rvert\tilde{\mathcal{S}}_{ij}}\] with the subscripts \(i\) and \(j\) here denoting the elements of the tensor. The Smagorinsky coefficient is then \[C_{s}^{2}=-\frac{L_{ij}M_{ij}}{2\tilde{\Delta}^{2}M_{ij}M_{ij}}\label{eq:germano}\] The values obtained from [\[eq:germano\]](#eq:germano){reference-type="eqref" reference="eq:germano"} contain large fluctuations. Ways around this: averaging in the homogeneous direction (no good for waves). Interesting to read paper on SPH from an LES perspective. Additional terms due to Lagrangian nature?
### Local averaging
The values of \(C_{s}\) obtained from [\[eq:germano\]](#eq:germano){reference-type="eqref" reference="eq:germano"} are averaged locally using a Shephard filter (e.g. SPH filtering).
### Lagrangian averaging
Following XX, consider the error of the something XX along streamlines. Introducing a weighting function \(\exp\left(-\left(t-t^{\prime}\right)/T\right)\), an expression for the error with respect to \(C_{s}\) is obtained. This expresion can be differentiated, to obtain the value of \(C_{s}\) for which the error is a minimum: \[C_{s}^{2}=\frac{\mathcal{I}_{LM}}{\mathcal{I}_{MM}}\] in which
The choice of an exponential weighting function enables [\[eq:Ilmmm\]](#eq:Ilmmm){reference-type="eqref" reference="eq:Ilmmm"} to be expressed as advection-relaxation equations:
Following , there are some choices of the memory timescale. We choose to set \(T=\tilde{\Delta}\mathcal{I}_{LM}^{-1/4}/2\) in which the constant \(C_{\mu,B}=0.6\).
### Bubble-induced turbulence model
We use the model by Sato & Sekoguchi: \[\mu_{B}=\rho_{0}C_{\mu,B}\displaystyle\sum_{k=1}^{N_{G}}\alpha_{k}d_{k}\left\lvert\tilde{\bm{u}}_{k}-\tilde{\bm{u}}_{0}\right\rvert\]
## Bubble entrainment model
We use (following ), originally in \[B_{k}=\frac{c_{en}\rho_{0}}{4\pi\sigma}\alpha_{0}\left\{\frac{f\left(a_{k}\right)\Delta{a}_{k}}{\displaystyle\sum_{k=1}^{N_{G}}a^{2}_{k}f\left(a_{k}\right)\Delta{a}_{k}}\right\}\varepsilon_{0,sps,S}\label{eq:entrainment}\] where \(c_{en}\) is a tuned parameter, \(\sigma\) is surface tension, \(a_{k}=d_{k}/2\), \(\Delta{a}_{k}\) is the width of the group, and \(f\left(a_{k}\right)\) is the bubble size spectrum, given by \[f\left(a\right)\propto\begin{cases}a^{-10/3}&\text{if}\quad{a}>a_{h}\\a^{-3/2}&\text{if}\quad{a}\le{a}_{h}\end{cases}\] where \(a_{h}\) is the Hinze scale, taken to be \(10^{-3}\) ? XX. The shear-induced sub-particle-scale dissipation rate is \[\varepsilon_{sps,S}=\frac{\mu_{S}}{\rho_{0}}\left\lvert\tilde{\mathcal{S}}\right\rvert^{2}\]
### Derivation
For bubble group \(k\), let \(E_{k}\) be the energy required to create a single bubble, related to surface tension via \[E_{k}=4\pi{a}_{k}^{2}\sigma\] Assume the power involved in bubble production-\(E_{k}B_{k}\)-is proportional to the turbulent dissipation rate: \[E_{k}B_{k}=c_{en}\alpha_{0}\rho_{0}\varepsilon_{0,sps,S}\] Combining the above gives \[B_{k}=\frac{c_{en}\rho_{0}}{4\pi\sigma}\alpha_{0}\frac{1}{a_{k}^{2}}\varepsilon_{0,sps,S}\] The form of [\[eq:entrainment\]](#eq:entrainment){reference-type="eqref" reference="eq:entrainment"} arises by assuming the bubble size distribution varies according to \(f\left(a\right)\) (which originate from lab experiments of Dean&Stokes): \[B_{k}\propto{f}\left(a_{k}\right)\Delta{a}_{k}\label{eq:entrain_spec}\] then relating the total energy of entrainment to the SPS dissipation rate: \[\displaystyle\sum_{k=1}^{N_{G}}4\pi{a}_{k}^{2}\sigma{B}_{k}=c_{en}\rho_{0}\alpha_{0}\varepsilon_{0,sps,S}\label{eq:entrain_energy}\] Putting [\[eq:entrain_spec\]](#eq:entrain_spec){reference-type="eqref" reference="eq:entrain_spec"} into [\[eq:entrain_energy\]](#eq:entrain_energy){reference-type="eqref" reference="eq:entrain_energy"} we get a constant of proportionality \[f_{0}=\frac{c_{en}}{4\pi}\frac{\rho_{0}}{\sigma}\alpha_{0}\frac{1}{\displaystyle\sum_{k=1}^{N_{G}}a_{k}^{2}f\left(a_{k}\right)\Delta{a}_{k}}\varepsilon_{0,sps,S}\] which leads to [\[eq:entrainment\]](#eq:entrainment){reference-type="eqref" reference="eq:entrainment"}.
## Bubble phase diffusion
The bubble phase diffusion term is given by \[D_{k}=\frac{\rho_{a}}{m_{k}\rho_{L}}\nabla\cdot\left(\frac{\mu_{S}+\mu_{B}}{Sc_{a}}\nabla\alpha_{k}\right)\] which is different from Kirby, who may have a mistake, and this needs checking.
## Bubble breakup model
The bubble breakup model is thatproposed by Martinez-bazan et al (2010), and I haven't yet implemented it\...
## Momentum exchange
This bit referencing . The sum of the momentum transfer between phases is zero: \[\rho_{0}\bm{M}_{0}+\rho_{a}\displaystyle\sum_{k=1}^{N_{G}}\bm{M}_{k}=0\] The momentum exchange between the liquid and each bubble group is determined by the drag, lift and virtual mass forces: \[\bm{M}_{k}=\alpha_{k}\frac{\rho_{0}}{\rho_{a}}\left(\bm{f}_{k}^{VM}+\bm{f}_{k}^{D}+\bm{f}_{k}^{L}\right)\] which are given by
We set \(C_{VM}=C_{L}=0.5\) and \[C_{D}=\begin{cases}\frac{24}{Re_{k}}\left(1+0.15Re_{k}^{0.687}\right)&\text{for}\quad{Re}_{k}<1000\\0.44&\text{for}\quad{Re}_{k}\ge1000\end{cases}\] where the bubble Reynolds number is \(Re_{k}=d_{k}\left\lvert\tilde{\bm{u}}_{0}-\tilde{\bm{u}}_{k}\right\rvert\rho_{0}/\mu\).
# Equations in volume averaged frame of reference
The mixture averaged velocity (average w.r.t. volume) is \[\tilde{\bm{u}}_{m}=\alpha_{0}\tilde{\bm{u}}_{0}+\displaystyle\sum_{k=1}^{N_{G}}\alpha_{k}\tilde{\bm{u}}_{k}\] The relative velocity between the liquid phase \(0\) and phase \(k\) is \[\tilde{\bm{u}}_{r}^{k}=\tilde{\bm{u}}_{0}-\tilde{\bm{u}}_{k}\] The velocities for each phase are then
where \[\tilde{\bm{u}}_{r}^{L}=\displaystyle\sum_{k=1}^{N_{G}}\alpha_{k}\tilde{\bm{u}}_{r}^{k}\] We will solve the equations in FoR moving with the mixture velocity plus shifting: particles have velocity \(\tilde{\bm{u}}_{m}+\bm{u}_{s}\). The derivative moving with the particles is then \[\frac{d}{dt}=\frac{\partial}{\partial{t}}+\left(\tilde{\bm{u}}_{m}+\bm{u}_{s}\right)\cdot\nabla\] All the LES, bubble production, and interfacial momentum transfer models are unaffected by the change of F.o.R.
## Liquid phase
Conservation of mass and momentum for the liquid phase are: \[\frac{d}{dt}\left(\alpha_{0}\rho_{0}\right)+\left(\tilde{\bm{u}}_{r}^{L}-\bm{u}_{s}\right)\cdot\nabla\left(\alpha_{0}\rho_{0}\right)=-\alpha_{0}\rho_{0}\nabla\cdot\tilde{\bm{u}}_{0}\label{eq:mass_l}\] \[\frac{d}{dt}\left(\alpha_{0}\tilde{\bm{u}}_{0}\right)+\left(\tilde{\bm{u}}_{r}^{L}-\bm{u}_{s}\right)\cdot\nabla\left(\alpha_{0}\tilde{\bm{u}}_{0}\right)=-\frac{1}{\rho_{0}}\nabla\left(\alpha_{0}\tilde{p}\right)+\alpha_{0}\bm{g}+\frac{1}{\rho_{0}}\nabla\cdot\left(\alpha_{0}\bm{\tau}\right)+\bm{M}_{0}\label{eq:mom_l}\] in which \(\bm{\tau}\) is the viscous stress, comprised of both resolved and modelled terms. Discussed later. Closed with an equation of state: \(\tilde{p}=c^{2}\rho_{0}\).
## Gas phase
The volume fraction of the \(k^{th}\) bubble group obeys the evolution equation \[\frac{d\alpha_{k}}{dt}+\left(\tilde{\bm{u}}_{r}^{L}-\tilde{\bm{u}}_{r}^{k}-\bm{u}_{s}\right)\cdot\nabla\alpha_{k}+\alpha_{k}\nabla\cdot\tilde{\bm{u}}_{k}=\frac{m_{k}}{\rho_{a}}\left\{B_{k} + S_{k} + D_{k}\right\};\quad\forall{k}\in\left[1,N_{G}\right]\label{eq:vf_b},\] The equation for the momentum of the \(k^{th}\) bubble group is \[\frac{d}{dt}\left(\alpha_{k}\tilde{\bm{u}}_{k}\right)+\left(\tilde{\bm{u}}_{r}^{L}-\tilde{\bm{u}}_{r}^{k}-\bm{u}_{s}\right)\cdot\nabla\left(\alpha_{k}\tilde{\bm{u}}_{k}\right)=-\frac{1}{\rho_{a}}\nabla\left(\alpha_{k}\tilde{p}\right)+\alpha_{k}\bm{g}+\bm{M}_{k}\label{eq:mom_b}\] We neglect the LHS of [\[eq:mom_b\]](#eq:mom_b){reference-type="eqref" reference="eq:mom_b"} and bubble phase momentum can be expressed as an evolution equation for the relative velocity: \[\begin{gathered}
0=-\frac{1}{\rho_{a}}\nabla\left(\alpha_{k}\tilde{p}\right)+\alpha_{k}\bm{g}+\\\frac{\alpha_{k}\rho_{0}}{\rho_{a}}\left\{C_{VM}\left(\frac{d\tilde{\bm{u}}_{r}^{k}}{dt}+\left(\tilde{\bm{u}}_{r}^{L}+\bm{u_{s}}\right)\cdot\nabla\tilde{\bm{u}}_{0}-\left(\tilde{\bm{u}}_{r}^{L}-\tilde{\bm{u}}_{r}^{k}-\bm{u_{s}}\right)\cdot\nabla\tilde{\bm{u}}_{k}\right)+\right.\\\left.\frac{3}{4}\frac{C_{D}}{d_{k}}\tilde{\bm{u}}_{r}^{k}\left\lvert\tilde{\bm{u}}_{r}^{k}\right\rvert+C_{L}\tilde{\bm{u}}_{r}^{k}\times\nabla\times\tilde{\bm{u}}_{0}\right\} \end{gathered}\] rearranged to obtain: \[\begin{gathered}
\frac{d\tilde{\bm{u}}_{r}^{k}}{dt}+\left(\tilde{\bm{u}}_{r}^{L}+\bm{u_{s}}\right)\cdot\nabla\tilde{\bm{u}}_{0}-\left(\tilde{\bm{u}}_{r}^{L}-\tilde{\bm{u}}_{r}^{k}-\bm{u_{s}}\right)\cdot\nabla\tilde{\bm{u}}_{k}=\\ \frac{C_{VM}\rho_{a}}{\alpha_{k}\rho_{0}}\left(\frac{1}{\rho_{a}}\nabla\left(\alpha_{k}\tilde{p}\right)-\alpha_{k}\bm{g}\right)-C_{VM}\left(\frac{3}{4}\frac{C_{D}}{d_{k}}\tilde{\bm{u}}_{r}^{k}\left\lvert\tilde{\bm{u}}_{r}^{k}\right\rvert+C_{L}\tilde{\bm{u}}_{r}^{k}\times\nabla\times\tilde{\bm{u}}_{0}\right)\label{eq:nodiss} \end{gathered}\] We ought to update the particle volumes. From Reynolds Transport theorem: \[\frac{dV}{dt}=V\nabla\cdot\left(\tilde{\bm{u}}_{m}+\bm{u}_{s}\right)\] where \(V\) is the volume of a particle. Or should we evolve \(\ln{V}\) to guarantee non-negative volumes? \[\frac{d\ln{V}}{dt}=\nabla\cdot\left(\tilde{\bm{u}}_{m}+\bm{u}_{s}\right)\] For single-phase flows, \(\tilde{\bm{u}}_{m}\) is the liquid velocity, and the volumes are (for incompressible flow) constant.
- Equation [\[eq:nodiss\]](#eq:nodiss){reference-type="eqref" reference="eq:nodiss"} contains no dissipation!
- \(\nabla\cdot\tilde{\bm{u}}_{k}\) in [\[eq:vf_b\]](#eq:vf_b){reference-type="eqref" reference="eq:vf_b"} is a source of bubbles, even when we prescribe \(\tilde{\bm{u}}_{r}^{k}=0\) and have an initially uniform \(\alpha_{0}\). Maybe Renato is right-we're not conservative, and just have to accept it.
- I was having issues integrating [\[eq:nodiss\]](#eq:nodiss){reference-type="eqref" reference="eq:nodiss"}. If \(C_{D}=C_{L}=0\), and it still tends to lead to instability. I was testing for steady Poiseuille flow, where the time deriv and advective terms should be zero-virtual mass term still causing instability.
- Setting \(C_{VM}=C_{L}=0\) and equating drag to pressure/grav terms to obtain velocity can be made to work (within limits).
- I mentioned there's no limit in [\[eq:vf_b\]](#eq:vf_b){reference-type="eqref" reference="eq:vf_b"} to the values of \(\alpha_{k}\) can take. One option is to artificially set a max and min value. Another option is to evolve \(\ln\alpha_{k}\) to prevent negative volume fractions (easy to rearrange [\[eq:vf_b\]](#eq:vf_b){reference-type="eqref" reference="eq:vf_b"} to do this. Would be nice if there was a transformation which imposed a similar upper limit-actually a tangent or cotangent transform would be a very elegant way to do this\... Another option is to put in some penalty functions on the RHS to impose limits.
# Numerical implementation
## SPH discretisation
In SPH, the fluid is represented by set of discrete particles, each of which we label \(i\in\left[1,N\right]\), where \(N\) is the total number of particles. In the present work the core of the SPH scheme closely follows . Each particle \(i\) has a volume \(\delta{V}\) (uniform and constant in the present formulation) and a position vector \(\bm{r}_{i}=\left[x_{i},y_{i}\right]^{T}\). Properties of particle \(i\) are denoted by the subscript \(i\). We denote the difference in the property \(\left(\cdot\right)\) of two particles \(i\) and \(j\) as \(\left(\cdot\right)_{ij}=\left(\cdot\right)_{i}-\left(\cdot\right)_{j}=-\left(\cdot\right)_{ji}\). In SPH, gradients of field variables at the location of particle \(i\) are calculated using a weighted sum of the values of the field variables at the neighbouring particles \(j\in\mathcal{N}_{i}\) where the weights are obtained from a kernel function \(W\left(\left\lvert\bm{r}_{ij}\right\rvert\right)=W_{ij}\) and its derivatives. The set of neighbours \(\mathcal{N}_{i}\) contains all particles \(j\) with \(\left\lvert\bm{r}_{ij}\right\rvert\le{r}_{s}\), where \(r_{s}\) is the support radius of the kernel. For a derivation and analysis of SPH fundamentals, we refer the reader to . Throughout this work we use the Wendland C2 kernel with a support radius of \(2h\) and an initial particle spacing of \(\delta{r}=h/1.3\). First derivatives are discretized according to \[\langle\nabla\tilde{\phi}\rangle_{i}=\displaystyle\sum_{j}\left(\tilde{\phi}_{j}-\tilde{\phi}_{i}\right)\nabla{W}^{\star}_{ij}\delta{V}\qquad\langle\nabla\cdot\tilde{\bm{u}}\rangle_{i}=\displaystyle\sum_{j}\left(\tilde{\bm{u}}_{j}-\tilde{\bm{u}}_{i}\right)\cdot\nabla{W}^{\star}_{ij}\delta{V}\label{eq:grad}.\] The Laplacian is approximated using the formulation of as \[\langle\nabla^{2}\tilde{\phi}\rangle_{i}=\displaystyle\sum_{j}\frac{2\tilde{\phi}_{ij}}{\lvert\bm{r_{ij}}\rvert^{2}}\bm{r_{ij}}\cdot\nabla{W}^{\star}_{ij}\delta{V}.\label{eq:lap}\] The corrected kernel gradient \(\nabla{W}^{\star}_{ij}\) due to is used, which provides first-order consistency for first derivatives, and ensures zero-order consistency for the Laplacian, and is given by \(\nabla{W}^{\star}_{ij}=\bm{L}_{i}\nabla{W}_{ij}\), with \(\nabla{W}\) the uncorrected kernel gradient, and the correction tensor \[\bm{L}_{i}=\begin{bmatrix}\displaystyle\sum{x}_{ji}\nabla{W}_{ij}\cdot\bm{e_{x}}\delta{V}&\displaystyle\sum{x}_{ji}\nabla{W}_{ij}\cdot\bm{e_{y}}\delta{V}\\\displaystyle\sum{y}_{ji}\nabla{W}_{ij}\cdot\bm{e_{x}}\delta{V}&\displaystyle\sum{y}_{ji}\nabla{W}_{ij}\cdot\bm{e_{y}}\delta{V}\end{bmatrix}^{-1},\] with \(\bm{e_{x}}\) and \(\bm{e_{y}}\) the unit vectors in the \(x\)-and \(y\)-directions respectively. XX Write above eqn concisely or in 3D\... The value of \(\delta{t}\) is set adaptively according to criteria for the Courant condition and viscous diffusion, as in , but with additional constraint for sound speed: \[\delta{t}=\min\left\{0.2\min\left(\frac{h}{\max_{i}\left(\lvert\tilde{\bm{u}}_{0}\rvert\right)},\frac{\rho_{0}{h}^{2}}{\mu}\right),8\frac{h}{c}\right\},\] in which \(\max_{i}\) is the maximum value over all particles \(i\in\left[1,N\right]\). Because acoustic waves are treated implicitly, the coefficient for the acoustic time step constraint can be as big as we want. Smaller is less dissipative, but more costly\... \(8\) is already pretty dissipative.
## Algorithm
Temporal integration of the fluid phase largely follows , with a modification to the Poisson equation to account for compressibility. The algorithm is based on Chorin's projection method , initially introduced to SPH by . The algorithm is presented in an Arbitrary-Lagrangian-Eulerian (ALE) framework, in which particles move with velocity \(\tilde{\bm{u}}_{m}+\bm{u_{s}}\). We set \(\bm{u}_{s}\) using the particle shifting procedure as described in , originally developed by . In the following, each operation is applied to every particle \(i\in\left[1,N\right]\), and we have dropped the subscripts \(i\) for clarity. We introduce the superscripts \(n\), \(\star\) and \(n+1\) to represent properties at the current, intermediate, and next time step. The algorithm is as follows. Throughout simulation we store \(\tilde{\bm{u}}_{0}\) and \(\tilde{\bm{u}}_{r}^{k}\). Other velocities calculated only when needed.
1. Advect particles \(\bm{r}^{\star}=\bm{r}^{n}+\delta{t}\tilde{\bm{u}}_{m}\)
2. Create mirror particles, neighbour lists and interparticle weights. Calculate \(\bm{u}_{s}\) based on a modified form of the Fickian shifting introduced by , described in detail in .
3. 1st order update of vol fractions-store \(\alpha^{n}\) and \(\alpha^{n+1}\).
4. Prediction step to get \(\tilde{\bm{u}}_{0}^{\star}\): \[\tilde{\bm{u}}_{0}^{\star}\alpha_{0}^{n+1}=\tilde{\bm{u}}_{0}^{n}\alpha_{0}^{n}+\delta{t}\left[\left(\bm{u_{s}}-\tilde{\bm{u}}_{r}^{L}\right)\cdot\nabla\tilde{\bm{u}}_{0}^{n}+\frac{1}{\rho_{0}}\nabla\cdot\left(\mu_{eff}\alpha_{0}^{n}\nabla\tilde{\bm{u}}_{0}^{n}\right) + \bm{M}_{0}\right]\] All LES modelling done here. Velocity gradients stored for use later.
5. Solve Poisson/Helmholtz/Elliptic equation [\[eq:icompppe\]](#eq:icompppe){reference-type="eqref" reference="eq:icompppe"} or [\[eq:isoTppe\]](#eq:isoTppe){reference-type="eqref" reference="eq:isoTppe"} (see [\[sec:incomp\]](#sec:incomp){reference-type="ref" reference="sec:incomp"} or [\[sec:isoT\]](#sec:isoT){reference-type="ref" reference="sec:isoT"}) to get \(\alpha_{0}\tilde{p}^{n+1}\), and hence \(\tilde{p}^{n+1}\) Solved using a BiCGStab algorithm with Jacobi preconditioning. subject to boundary conditions \(\bm{n}\cdot\nabla{\alpha_{0}\tilde{p}}^{n+1}=\bm{n}\alpha_{0}^{n+1}\cdot\bm{g}\) on solid boundaries (where \(\bm{n}\) is the local unit boundary normal vector), and \(\alpha_{0}^{n+1}\tilde{p}^{n+1}=0\) on free-surfaces.
6. Project \(\tilde{\bm{u}}_{0}^{\star}\) onto \(\tilde{\bm{u}}_{0}^{n+1}\) using pressure gradient: \[\tilde{\bm{u}}_{0}^{n+1}=\tilde{\bm{u}}_{0}^{\star}-\frac{\delta{t}}{\rho_{0}}\nabla\alpha_{0}^{n+1}{\tilde{p}}^{n+1}+\delta{t}\alpha_{0}^{n+1}\bm{g},\]
7. Update bubble phase velocities (we'll actually just update \(\tilde{\bm{u}}_{r}^{k}\))
8. Calculate \(\tilde{\bm{u}}_{m}\) and update volumes.
9. Final advect particles: \(\bm{r}^{n+1}=\bm{r}^{n}+\delta{t}\left(\tilde{\bm{u}}_{m}^{n}+\tilde{\bm{u}}_{m}^{n+1}+2\bm{u}_{s}\right)/2\)
### Fractional step approach-isothermal compressibility[\[sec:isoT\]]{#sec:isoT label="sec:isoT"}
Splitting the RHS of [\[eq:mom_l\]](#eq:mom_l){reference-type="eqref" reference="eq:mom_l"} between pressure and gravity, and viscous and interfacial terms, and using the speed of sound to convert [\[eq:mass_l\]](#eq:mass_l){reference-type="eqref" reference="eq:mass_l"} into an evolution equation for \(\tilde{p}\), we get the following equation for pressure: \[\begin{gathered}
\frac{1}{\rho_{0}}\nabla^{2}\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)-\frac{\alpha_{0}^{n+1}\tilde{p}^{n+1}}{\delta{t}^{2}c^{2}\rho_{0}}-\frac{\left(\tilde{\bm{u}}_{r}^{L}-\bm{u}_{s}\right)}{\delta{t}c^{2}\rho_{0}}\cdot\nabla\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)+\frac{1}{\delta{t}}\tilde{\bm{u}}_{0}^{n+1}\cdot\nabla\alpha_{0}^{n+1}\\=\frac{1}{\delta{t}}\nabla\cdot\left(\alpha_{0}^{n+1}\tilde{\bm{u}}_{0}^{\star}\right)+\bm{g}\cdot\nabla\alpha_{0}^{n+1}-\frac{\alpha_{0}^{n}\tilde{p}^{n}}{\delta{t}^{2}c^{2}\rho_{0}}\end{gathered}\] Problematic, with \(\tilde{\bm{u}}_{0}^{n+1}\) explicitly in there. We use \[\tilde{\bm{u}}_{0}^{n+1}=\tilde{\bm{u}}_{0}^{\star}+\delta{t}\bm{g}-\frac{\delta{t}}{\alpha_{0}^{n+1}\rho_{0}}\nabla\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)\] which then gives: \[\begin{gathered}
\frac{1}{\rho_{0}}\nabla^{2}\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)-\frac{\alpha_{0}^{n+1}\tilde{p}^{n+1}}{\delta{t}^{2}c^{2}\rho_{0}}-\frac{\left(\tilde{\bm{u}}_{r}^{L}-\bm{u}_{s}\right)}{\delta{t}c^{2}\rho_{0}}\cdot\nabla\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)-\frac{1}{\rho_{0}\alpha_{0}^{n+1}}\nabla\alpha_{0}^{n+1}\cdot\nabla\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)\\=\frac{\alpha_{0}^{n+1}}{\delta{t}}\nabla\cdot\tilde{\bm{u}}_{0}^{\star}-\frac{\alpha_{0}^{n}\tilde{p}^{n}}{\delta{t}^{2}c^{2}\rho_{0}}\label{eq:isoTppe}\end{gathered}\]
### Fractional step approach-incompressible[\[sec:incomp\]]{#sec:incomp label="sec:incomp"}
If incompressible, we let \(c^{2}\) tend to infinity, and the isothermally compressible equation [\[eq:isoTppe\]](#eq:isoTppe){reference-type="eqref" reference="eq:isoTppe"} collapses to: \[\frac{1}{\rho_{0}}\nabla^{2}\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)-\frac{1}{\rho_{0}\alpha_{0}^{n+1}}\nabla\alpha_{0}^{n+1}\cdot\nabla\left(\alpha_{0}^{n+1}\tilde{p}^{n+1}\right)=\frac{\alpha_{0}^{n+1}}{\delta{t}}\nabla\cdot\tilde{\bm{u}}_{0}^{\star}\label{eq:icompppe}\] If \(\alpha_{0}^{n+1}\) is uniform, the final term on the LHS is zero, and it collapses to the single phase PPE. N.B. equation [\[eq:icompppe\]](#eq:icompppe){reference-type="eqref" reference="eq:icompppe"} can also be derived by setting \(\nabla\cdot\tilde{\bm{u}}_{0}^{n+1}=0\).
- I showed results for 2D Taylor Green with \(\tilde{\bm{u}}_{r}^{k}=0\) and fixed \(\alpha_{0}=0.8+0.2\sin\left(2x\right)\sin\left(2y\right)\). Liquid phase doesn't "see" gas phase. Interesting, and led to discussion of what \(\alpha_{0}\) is doing here.
- Renato suggested \(\overline{\rho}=\alpha_{0}\rho_{0}\) is of interest. I agree it is, but it is not quite the same as a density, because [\[eq:mom_l\]](#eq:mom_l){reference-type="eqref" reference="eq:mom_l"} is coupled to [\[eq:mass_l\]](#eq:mass_l){reference-type="eqref" reference="eq:mass_l"} by \(\alpha_{0}\) and a constant \(\rho_{0}\), whilst [\[eq:mass_l\]](#eq:mass_l){reference-type="eqref" reference="eq:mass_l"} takes a prescribed \(\alpha_{0}\) and allows \(\rho_{0}\) to vary (actually evolving \(\tilde{p}=c^{2}\rho_{0}\)). Maybe it is the same as a density if \(c\to\infty\)?
- I see \(\alpha_{0}\) as sort of a field which alters the momentum without affecting the density, which is why I tried to describe it as warping the background space in which we solve the equations.
- Looking for nice tests to validate. I will check Elliptic eqn against an exact solution.
- Poiseuille flow might be promising, as variable density/viscosity solutions still have uniform pressure.
# A new sub-particle-scale model for SPH
Consider three filters: \(\tilde{\Delta}=h\) is the "grid" filter, \(\widehat{\Delta}=2h\) is a Shepard filter, and \(\overline{\Delta}=4h\) is a box filter over the kernel support. The strain rate tensor is \(S\). First we consider the grid and Shepard filter scales, and we denote this pair of scales with a superscript \(\alpha\). The unresolved momentum fluxes at the grid and Shepard filter scales are:
Between these scales, Germano's identity reads \[\mathcal{L}_{ij}^{\alpha}=\widehat{\tilde{u}_{i}\tilde{u}_{j}}-\widehat{\tilde{u}}_{i}\widehat{\tilde{u}}_{j}=T^{\alpha}_{ij}-\widehat{\mathcal{T}}^{\alpha}_{ij}\label{eq:germ1}\] The Smagorinsky model at these scales gives:
Substituting the Germano identity [\[eq:germ1\]](#eq:germ1){reference-type="eqref" reference="eq:germ1"} into [\[eq:smag1\]](#eq:smag1){reference-type="eqref" reference="eq:smag1"} we obtain: \[\mathcal{L}_{ij}^{\alpha}-\frac{\delta_{ij}}{3}\mathcal{L}_{kk}^{\alpha}=2C_{\alpha}\left\{\widehat{\tilde{\Delta}^{2}\left\lvert\tilde{S}\right\rvert\tilde{S}_{ij}}-\widehat{\Delta}^{2}\left\lvert\widehat{\tilde{S}}\right\rvert\widehat{\tilde{S}}_{ij}\right\}\label{eq:dyn1}\] Next we consider the same process, but at the Shepard and box filter scales, denoting this pair of scales with a superscript \(\beta\). The unresolved momentum fluxes at the Shepard and box filter scales are:
Between these scales, Germano's identity reads \[\mathcal{L}_{ij}^{\beta}=\overline{\widehat{\tilde{u}}_{i}\widehat{\tilde{u}}_{j}}-\overline{\widehat{\tilde{u}}}_{i}\overline{\widehat{\tilde{u}}}_{j}=T^{\beta}_{ij}-\overline{\mathcal{T}}^{\beta}_{ij}\label{eq:germ2}\] The Smagorinsky model at these scales gives:
We have introduced the notation \(\breve{\left(\cdot\right)}=\widehat{\tilde{\left(\cdot\right)}}\), and we note that \(\breve{S}\) is interpreted as having the grid and then Shepard filter applied to the velocity field, after which the gradient is evaluated (i.e., \(\breve{S}\) is different from the \(\widehat{\tilde{S}}\) which appears in [\[eq:smag1\]](#eq:smag1){reference-type="eqref" reference="eq:smag1"}). Substituting the Germano identity [\[eq:germ2\]](#eq:germ2){reference-type="eqref" reference="eq:germ2"} into [\[eq:smag2\]](#eq:smag2){reference-type="eqref" reference="eq:smag2"} we obtain: \[\mathcal{L}_{ij}^{\beta}-\frac{\delta_{ij}}{3}\mathcal{L}_{kk}^{\beta}=2C_{\beta}\left\{\overline{\widehat{\Delta}^{2}\left\lvert\breve{S}\right\rvert\breve{S}_{ij}}-\overline{\Delta}^{2}\left\lvert\overline{\breve{S}}\right\rvert\overline{\breve{S}}_{ij}\right\}\label{eq:dyn2}\] In [\[eq:dyn1\]](#eq:dyn1){reference-type="eqref" reference="eq:dyn1"} and [\[eq:dyn2\]](#eq:dyn2){reference-type="eqref" reference="eq:dyn2"} we have two means of evaluating a dynamic Smagorinsky coefficient, based on different scales. The former is based on the grid filter, using the Shepard filter as a test filter. The latter filters the velocity fields with the Shepard filter, then uses this as the grid filter and the box filter as a test filter. The idea is that the two approaches will give different results, due to the SPH gradient operator being better able to resolve the gradients of the grid-filtered-Shepard-filtered velocity field than it can for the grid-filtered velocity field. The next observation is that because the gradient operator is linear \[\widehat{\frac{\partial\tilde{u}_{i}}{\partial{x}_{j}}}=\frac{\partial\widehat{\tilde{u}}_{i}}{\partial{x}_{j}}\] i.e. \(\widehat{\tilde{S}}=\breve{S}\). However, this is only true in the continuous case, where the modal response of the discrete gradient operator is linear. In reality the response of the discrete gradient operator is non-linear (in wavenumber space), and \[\widehat{\left\langle\frac{\partial\tilde{u}_{i}}{\partial{x}_{j}}\right\rangle}\ne\left\langle\frac{\partial\widehat{\tilde{u}}_{i}}{\partial{x}_{j}}\right\rangle\] where \(\left\langle\cdot\right\rangle\) is the discrete (SPH) approximation. We will somehow use this inequality as a measure of the accuracy of the gradient operators, and to inform how the Smagorinsky coefficient is calculated. We can evaluate \(C_{\alpha}\) and \(C_{\beta}\) to separately satisfy [\[eq:dyn1\]](#eq:dyn1){reference-type="eqref" reference="eq:dyn1"} and [\[eq:dyn2\]](#eq:dyn2){reference-type="eqref" reference="eq:dyn2"} in a least squares sense as in the traditional dynamic model. The next step is to work out what coefficient to actually choose? Another point: [\[eq:dyn2\]](#eq:dyn2){reference-type="eqref" reference="eq:dyn2"} interprets the cut-off scale as the Shepard filter scale, and we expect it to result in greater sub-grid-dissipation rates (I think). I got a bit confused here\...
## Demonstration of something
Verify this by considering (in one dimension) \[\tilde{u}=a_{0}+\displaystyle\sum_{n=1}^{\infty}a_{n}\sin\left(nx\right)\] We then have \[\frac{\partial\tilde{u}}{\partial{x}}=\displaystyle\sum_{n=1}^{\infty}na_{n}\cos\left(nx\right)\] Let the amplitude response of the filter at wavenumber \(n\) be \(b_{n}\), and we see that \[\widehat{\frac{\partial\tilde{u}}{\partial{x}}}=\displaystyle\sum_{n=1}^{\infty}b_{n}na_{n}\cos\left(nx\right)=\frac{\partial\widehat{\tilde{u}}}{\partial{x}}\]
# Results
# Conclusions {#conc}
# Introduction[\[intro\]]{#intro label="intro"}
Flows involving complex, dynamic free-surface motion are found widely in industry and nature, with fuel sloshing in aircraft wings and wave impacts on coastal and offshore structures being prime examples. For waves in particular, the violent motion of the free-surface often results in the entrainment of bubbles at the free-surface, which can have significant effects on the overall dynamics and peak loads, and plays a major role in the exchange of gas between the ocean and atmosphere. With a greater understanding of the effect of bubbles in breaking waves as our motiviation, we seek improved approaches to their numerical simulation. The topic of bubble entrainment in breaking waves has been the subject of considerable experimental (e.g. ) and numerical research, and with advances in computational resources and mesh-adaptivity, in recent years, researchers have begun to conduct multi-phase simulations with the aim of resolving even the smallest bubble and droplet scales . This work has elucidated fundamental aspects of the process of wave breaking, including the degree of three-dimensionality in the flow , non-locality in the bubble break-up cascade , and the underlying physical mechanisms controlling bubble break-up . Whilst high-fidelity simulations are desirable for obtaining fundamental insight, their applicability in more industrially relevent settings is limited due to computational costs. Even with adaptive mesh refinement, the computational cost of multi-phase simulations as in and is significant: of the order of a month. An alternative numerical approach is to model the presence of bubbles (rather than resolving individual bubbles), with some form of population balance equation and, typically, an assumption that the bubbles are spherical. With this approach, there are two options for the treatment of the dispersed phase. Firstly, it may be modelled as a continuum, through a bubble volume fraction, or number density field, which is subject to an evolution equation. Here the evolution equations for the dispersed phase are partial differential equations (PDEs), which are integrated in the same numerical framework as the equations of motion of the continuum liquid phase. In the mesh-based literature, such schemes are referred to as *Eulerian-Eulerian*, and have been developed for the simulation of bubble plume dynamics , air entrainment in breaking waves , and liquid jet breakup . Since the dynamics of bubbles dispersed in a liquid have a strong dependence on the bubble size, the treatment of polydisperse bubble distributions here is problematic. Typically, a population of bubbles is segregated into groups of similar sizes, each group requiring two additional evolution equations. Bubble breakup and coalescence may then be incorporated by source and sink terms exchanging mass (or number density) between bubble groups. An additional limitation is that this approach does not allow the tracking of individual bubbles over their lifetimes, but only of statistical averages, and this constrains potential additional physical models, such as those for bubble break up. Although numerically this approach allows bubbles larger than the discretisation lengthscale of the continuous liquid phase (but with small number density), the assumptions used in the derivation of such models limit bubble sizes to smaller than the resolution of the continuous phase . The second option for treating the dispersed phase is to model the bubbles as a set of discrete Lagrangian particles. Again, such approaches have been widely developed for mesh-based schemes, for example the work by focussing on bubble plumes, or the studies on turbulence-bubble interactions , and cavitation bubble clouds . In the mesh-based community, these methods are described as *Eulerian-Lagrangian* schemes. The particles representing individual bubbles interact with the continuous phase through exchanges of momentum (and sometimes volume, as in , although in many cases where the concentration of the dispersed phase is small, volume exchanges are neglected). Schemes of this type allow for the tracking of individual bubbles, and a continuous polydisperse bubble distribution poses no additional challenge. Temporarily setting aside the presence of bubbles, mesh-free methods have shown significant promise for simulations of breaking waves in recent decades. Smoothed Particle Hydrodynamics (SPH) is one mesh-free method, originally developed for astrophysical simulations , and since applied with considerable success to a range of terrestrial flows, including those with dynamically evolving free-surfaces . The fluid is discretised by a set of Lagrangian particles, and spatial derivatives are approximated by weighted sums of fluid properties at neighbouring particles. Whilst tracking a deforming surface undergoing topological changes is a complex task in mesh-based methods, for SPH, little additional effort is required. There are now a wide variety of SPH schemes and related methods capable of simulating breaking waves, including weakly compressible SPH , incompressible schemes , and the Moving Particle Semi-Implicit method . Although multi-phase SPH schemes are well established (e.g. ), and capable of simulating multiple bubbles , and bubble-free-surface interactions , we seek to avoid the cost of explicitly resolving both phases. We observe that the terms "Eulerian-Eulerian" and "Eulerian-Lagrangian" used above are misnomers (and somewhat ambiguous) in the context of SPH-based methods, while appropriate for Eulerian mesh-based numerical methods. In this work we refer to the two approaches as "continuous-continuous" and "continuous-discrete", descriptions which remain clear even when used to describe schemes with non-Eulerian methods for the continuous phase. Of the bubble modelling approaches described above, developments in SPH lag behind mesh-based methods. A model based on the continuous-continuous approach has only recently been introduced to SPH , but with very promising results for air entrainment in flow over a spillway , though the method is currently limited to simplified closure models for interphase momentum exchange. We are not aware of any continuous-discrete SPH models for bubbly flows, although the SPH implementations closest in philosophy to this approach are perhaps the multi-phase dusty gas simulations conducted in astrophysical settings . Herein, we present an SPH implementation of the continuous-discrete approach for bubbly free-surface flows. The liquid is resolved via large eddy simulations (LES) using a semi-implicit isothermally compressible SPH framework, whilst bubbles are represented as discrete Lagrangian particles which interact with the liquid via exchanges of momentum, volume, and sub-resolution turbulence closures. We particularly focus on integral models for bubble entrainment, break-up and free-surface interaction, with application to breaking waves. The remainder of the paper is set out as follows. In Section [9](#sec:ge){reference-type="ref" reference="sec:ge"} we introduce the governing equations of our model. Section [10](#sec:ni){reference-type="ref" reference="sec:ni"} presents details of the numerical implementation. In Section [11](#sec:plume){reference-type="ref" reference="sec:plume"} we test our simulation framework against numerical and experimental data for bubble plumes, and in Section [12](#sec:bw){reference-type="ref" reference="sec:bw"} we use our model to simulate the air entrainment in breaking waves. Section [13](#sec:conc){reference-type="ref" reference="sec:conc"} is a summary of our conclusions. Further validation of our LES model is provided in Appendix [\[les\]](#les){reference-type="ref" reference="les"}. A table of all symbols used in the work is given in Table [\[los\]](#los){reference-type="ref" reference="los"}.
# Governing equations {#sec:ge}
The system we consider is a continuous liquid phase, containing a dispersed bubble phase, as illustrated in Figure [\[fig:elschem\]](#fig:elschem){reference-type="ref" reference="fig:elschem"}. The liquid phase is governed by the isothermal compressible Navier-Stokes equations, whilst the bubbles are treated as discrete particles which obey Newton's second law. In the following, the subscripts \(l\) and \(b\) indicate properties in the liquid (continuous) and dispersed bubble phases respectively. Where the subscript \(l,b\) appears, it denotes a liquid property evaluated at a bubble. We consider the liquid to be isothermally compressible, with sound speed \(c\), density \(\rho_{l}\) and viscosity \(\mu_{l}\). The bubbles are assumed to be spherical, comprised of gas with density \(\rho_{b}\), and with a constant liquid-bubble surface tension of \(\gamma\). The entire system is subject to a gravitational acceleration of \(g\bm{e_{g}}\), where \(\bm{e_{g}}\) is a unit vector. Following non-dimensionalisation by suitable integral length-and velocity-scales, the problem is parameterised by the Reynolds number \(Re\), the Mach number \(Ma\), the Froude number \(Fr\), the (integral scale) Weber number \(We\), and the density ratio \(\beta=\rho_{l}/\rho{b}\). Although our numerical framework is able to capture acoustic signals, in the present work these are not of interest and are damped out, and hence \(Ma\) is treated as a numerical, rather than physical, parameter.
## Liquid phase
The liquid phase is modelled with a Large Eddy Simmulation (LES) scheme, with an implicit filter of width \(\widetilde{\Delta}\). Due to the assumption of isothermal compressibility, we can write \[\frac{d\tilde{p}}{d\rho_{l}}=c^{2},\] where \(\tilde{p}\) is the (implicitly) filtered pressure. The filtered continuity equation may be then expressed, in an arbitrary frame of reference, as an evolution equation for the pressure \[\frac{d}{d{t}}\left(\alpha\tilde{p}\right)-\bm{u}_{ps}\cdot\nabla\left(\alpha\tilde{p}\right)=-\frac{\alpha}{Ma^{2}}\nabla\cdot\tilde{\bm{u}}_{l},\label{eq:mass_le}\] where \(\alpha\) is the liquid volume fraction, and \(\tilde{\bm{u}}_{l}\) is the filtered liquid velocity. Following , the transport velocity \(\bm{u}_{ps}\) is the relative velocity between the frame of reference and a Lagrangian frame of reference. Hence, for perfectly Lagrangian scheme, we have \(\bm{u}_{ps}=\bm{0}\), whilst for an Eulerian scheme, \(\bm{u}_{ps}=-\tilde{\bm{u}}_{l}\). We next assume that the liquid is weakly compressible; the compressibility \(1/\rho_{l}c^{2}\ll1\), allowing us to neglect terms involving the density variation in the (filtered) momentum equation, which is written as \[\frac{d}{d{t}}\left(\alpha\tilde{\bm{u}}_{l}\right)-\bm{u}_{ps}\cdot\nabla\left(\alpha\tilde{\bm{u}}_{l}\right)=-\nabla\left(\alpha\tilde{p}\right)+\frac{\alpha}{Fr^{2}}\bm{e_{g}}+\frac{1}{Re}\nabla\cdot\left(\alpha\left(1+\nu_{srs}\right)\nabla\tilde{\bm{u}}_{l}\right)+\bm{M},\label{eq:mom_le}\] where \(\nu_{srs}\) is a dimensionless sub-resolution viscosity, determined by the LES closure model and the term \(\bm{M}\) represents the momentum exchange between the liquid and bubble phases.
### LES model
The filtered equations are closed with a model for the sub-resolution viscosity, which is comprised of a shear-induced and bubble-induced eddy viscosity component: \(\nu_{srs}=\nu_{S}+\nu_{B}\). The bubble-induced viscosity \(\nu_{B}\) accounts of the production of sub-resolution turbulence by bubbles following the model of , and is described in Section [9.2.2](#sec:bit){reference-type="ref" reference="sec:bit"}. For the shear-induced turbulence, we use the mixed-scale model of , with \[\nu_{S}=Re{C}_{M}\widetilde{\Delta}^{1+ where \(\widetilde{\mathcal{S}}=\left(\nabla\tilde{\bm{u}}_{l}+\left(\nabla\tilde{\bm{u}}_{l}\right)^{T}\right)/2\) is the resolved strain rate tensor, \(q_{c}^{2}\) is the test-filtered kinetic energy, $ \begin{equation}\varepsilon=\frac{1+\nu_{srs}}{Re}\left\lvert\tilde{\mathcal{S}}\right\rvert^{2}.\] In the development of our method, we explored several additional closure models, including a standard Smagorinsky model, the dynamic Germano model , with both local (via Shepard filtering) and Lagrangian (along streamlines) averaging. We choose the mixed-scale model for our simulations because it is known to yield good results in flows with deforming free-surfaces , and in tests of the decay of single phase isotropic turbulence (included in Appendix [\[les\]](#les){reference-type="ref" reference="les"}), it yielded improved results *in our numerical framework* compared with the other closure models.
## Dispersed bubble phase
The dispersed phase is represented by a discrete set of \(N_{b}\) Lagrangian bubbles \(\mathcal{B}\), each with a velocity \(\bm{u}_{b}\), position \(\bm{r}_{b}\), radius \(a_{b}\), and volume \(V_{b}=4\pi{a}_{b}^{3}/3\). The system of bubbles is governed by
for each bubble in \(\mathcal{B}\). Here \(\bm{F}_{d}\), \(\bm{F}_{l}\), \(\bm{F}_{vm}\) and \(\bm{F}_{g}\) are the drag, lift, virtual mass and buoyancy forces acting on the bubble due to the surrounding liquid. These forces are evaluated for each bubble \(i\in\mathcal{B}\) through closure models as in (e.g.) and , with
where the relative velocity between the bubble and liquid phase at bubble \(i\) is given by \(\bm{u}_{rel,i}=\bm{u}_{l}\left(\bm{r}_{b,i}\right)-\bm{u}_{b,i}\), and \(C_{d}\), \(C_{l}\) and \(C_{vm}\) are drag, lift and virtual mass coefficients respectively. Note the absence of the tilde in the liquid velocity appearing in the definition of relative velocity, and in [\[eq:fl\]](#eq:fl){reference-type="eqref" reference="eq:fl"} and [\[eq:fvm\]](#eq:fvm){reference-type="eqref" reference="eq:fvm"}. Following , the sub-resolution fluctuating part of the liquid velocity is modelled stochastically, with \[\bm{u}_{l}\left(\bm{r}_{b,i}\right)=\tilde{\bm{u}}_{l}\left(\bm{r}_{b,i}\right)+\bm{u}^{\prime}_{l}\left(\bm{r}_{b,i}\right)\] comprising the LES filtered velocity and a sub-resolution fluctuating part. The calculation of \(\bm{u}^{\prime}_{l}\left(\bm{r}_{b,i}\right)\) is described in Section [9.2.1](#sec:langevin){reference-type="ref" reference="sec:langevin"}. The drag coefficient in [\[eq:fd\]](#eq:fd){reference-type="eqref" reference="eq:fd"} is \[{C}_{d}=\begin{cases}0.44&{Re}_{b,i}>1000;\\\frac{24}{Re_{b,i}}\left(1+0.15{Re}_{b,i}^{0.687}\right)&Re_{b,i}\le1000,\end{cases}\] where the relative bubble Reynolds number is related to integral-scale Reynolds number by \[Re_{b,i}=2a_{b,i}\left\lvert\bm{u}_{rel,i}\right\rvert{Re}.\] Following the coefficient of virtual mass and the lift coefficient are set to \(C_{vm}=C_{l}=0.5\). The momentum exchange from bubble \(i\) back to the liquid phase is \[\bm{M}_{b,i}=\left(-\bm{F}_{d,i}-\bm{F}_{l,i}-\bm{F}_{vm,i}\right)/\beta.\]
### Langevin model {#sec:langevin}
The fluctuating velocity component "felt" by the bubbles is calculated through the integration of a stochastic model following work by and . The fluctuation velocity obeys the Langevin equation \[d\bm{u}_{l}^{\prime}=-\bm{\mathsf{G}}\bm{u}_{l}^{\prime}dt+\sqrt{2\sigma_{srs}^{2}}\bm{\mathsf{B}}\bm{dW},\label{eq:langevin1}\] where \(\bm{dW}\) is a Wiener process, and \(\bm{\mathsf{G}}\) and \(\bm{\mathsf{B}}\) are drift and diffusion matrices respectively. The quantity \(\sigma_{srs}\) is the standard deviation of the fluctuating velocity, and related to the sub-resolution turbulent kinetic energy \(k_{srs}\) by \[\sigma_{srs}=\sqrt{\frac{2}{3}k_{srs}}.\] The sub-resolution turbulent kinetic energy is estimated from the double filtered velocity, as \[k_{srs}=\frac{1}{2}\left\lvert\bm{\tilde{u}}_{l}-\bm{\hat{\tilde{u}}}_{l}\right\rvert^{2},\] where the \(\hat{}\) indicates explicit filtering with a test filter, descibed in Section [10.1.2](#sec:sph){reference-type="ref" reference="sec:sph"}. Following , by taking advantage of the fact that a Langevin equation may be analytically integrated, we transform [\[eq:langevin1\]](#eq:langevin1){reference-type="eqref" reference="eq:langevin1"} into the recursion equation \[\bm{u}_{l}^{\prime}\left(\bm{r}_{b,i},t+\delta{t}\right)=\bm{\mathsf{E}}\bm{u}_{l}^{\prime}\left(\bm{r}_{b,i},t\right)+\bm{\mathsf{W}}\bm{\zeta},\label{eq:langevin2}\] where \(\delta{t}\) is a time increment, \(\bm{\mathsf{E}}\) is the exponential of the drift matrix \(\bm{\mathsf{G}}\), \(\bm{\mathsf{W}}\) is the square root of the velocity fluctuation covariance matrix, and \(\bm{\zeta}\) is a random vector, whose components are normally distributed. Denoting the filtered relative velocity (excluding the fluctuations) as \(\bm{\tilde{u}}_{rel}\), we define the matrix \[\bm{\mathsf{R}}=\frac{1}{\left\lvert\bm{\tilde{u}}_{rel}\right\rvert^{2}}\bm{\tilde{u}}_{rel}\otimes\bm{\tilde{u}}_{rel},\] and the exponential of the drift matrix is then given by \[\bm{\mathsf{E}}=\left(E_{\parallel}-E_{\perp}\right)\bm{\mathsf{R}}+E_{\perp}\bm{\mathsf{I}},\] where \(\bm{\mathsf{I}}\) is the identity matrix, and
are the sub-resolution timescales associated with fluctuations parallel and perpendicular to the filtered velocity, and the sub-resolution timescale \(\tau_{srs}\) is related to the velocity fluctuations by \[\tau_{srs}=\frac{C\tilde{\Delta}}{\sigma_{srs}},\] with the constant \(C=1\). The factors relating \(\tau_{\parallel}\) and \(\tau_{\perp}\) to the sub-resolution timescale \(\tau_{srs}\) account for the crossing trajectory and continuity effects . The square root of the covariance matrix is given by \[\bm{\mathsf{W}}=\left(W_{\parallel}-W_{\perp}\right)\bm{\mathsf{R}}+W_{\perp}\bm{\mathsf{I}},\] where
### Bubble-induced turbulence model {#sec:bit}
As mentioned above, we evaluate the bubble-induced turbulence, following , based on the model of , where the contribution of an individual bubble to the turbulent viscosity is proportional to the product of the bubble diameter and the relative velocity. In our discrete bubble framework, the contribution of an individual bubble \(j\) is \[\nu_{B,b,j}=C_{\nu,B}2a_{b,j}\left\lvert\bm{u}_{rel,j}\right\rvert,\] where the constant \(C_{\nu,B}=0.6\) as in . To obtain the bubble-induced turbulent viscosity in the liquid \(\nu_{B}\), we interpolate \(\nu_{B,b}\) from the bubbles to the liquid phase, as described in Section [10.1.1](#sec:interp){reference-type="ref" reference="sec:interp"}.
### Bubble entrainment and breakup
Our intention is to simulate flows where bubbles are entrained at the free-surface. We use an entrainment model, similar in principle to that of and , in which a fraction of the turbulent kinetic energy of the liquid is assumed to be converted into surface energy as bubbles are created (or entrained) at the free-surface. We further include a model for bubble breakup, which is based on the imbalance between the restoring pressure on a bubble due to surface tension, and the deforming stress due to the turbulent motion of the liquid, based on the models of and . These models cannot be clearly explained without reference to our discretisation scheme, and we defer detailed description of them to later, in Section [10.4](#sec:be){reference-type="ref" reference="sec:be"}.
# Numerical implementation {#sec:ni}
## SPH discretisation
The liquid phase is represented by set of discrete particles \(\mathcal{P}\), each of which we label \(i\in\left[1,N\right]\), where \(N\) is the total number of particles. In the present work the core of the SPH scheme closely follows and . All particles carry a (constant) liquid mass \(m\), and with the assumption of a weakly compressible liquid, the associated liquid volume \(V_{l}\) is assumed constant. The position of particle \(i\) is denoted \(\bm{r}_{i}\), its smoothing length is \(h_{i}\), and the total volume associated with the particle is \(V_{i}\). We denote the difference in the property \(\left(\cdot\right)\) of two particles \(i\) and \(j\) as \(\left(\cdot\right)_{ij}=\left(\cdot\right)_{i}-\left(\cdot\right)_{j}=-\left(\cdot\right)_{ji}\). In SPH, values and derivatives of field variables at the location of particle \(i\) are calculated using a weighted sum of the values of the field variables at the neighbouring particles \(j\in\mathcal{P}_{i}\), where the weights are obtained from a kernel function \(W\left(\left\lvert\bm{r}_{ij},h_{i}\right\rvert\right)=W_{ij}\) and its derivatives. Here, \(\mathcal{P}_{i}\) is the set of neighbours of particle \(i\), and contains all particles \(j\) with \(\left\lvert\bm{r}_{ij}\right\rvert\le{r}_{s,i}\), where \(r_{s,i}\) is the support radius of the kernel of particle \(i\). Throughout this work we use the Wendland C2 kernel for which the support radius is \(r_{s,i}=2h_{i}\). We use an initial particle spacing of \(\delta{r}=h_{0}/1.3\), where \(h_{0}\) is the smoothing length when \(\alpha=1\). The implicit filter width is taken as \(\widetilde{\Delta}=h\). For a derivation and analysis of SPH fundamentals, we refer the reader to and . In a perfectly Lagrangian framework, particles follow streamlines, which can result in highly anisotropic particle distributions, particularly around stagnation points, degrading the accuracy of the simulation. To regularise the particle distribution, we use the particle shifting technique of to set \(\bm{u}_{ps}\) as in .
### Interpolation between phases {#sec:interp}
The interaction between the liquid and bubbles occurs both through the modification of the liquid volume fraction \(\alpha\) due to the presence of the bubbles, the momentum exchange \(\bm{M}\) between the phases, and the bubble-induced turbulent viscosity \(\nu_{B}\). To achieve this, it is necessary to interpolate between the phases: that is, to calculate the value of a liquid property at a bubble location, and vice versa. Figure [\[fig:schem\]](#fig:schem){reference-type="ref" reference="fig:schem"} provides a diagrammatic overview of the framework, including those properties which are exchanged between phases. Numerically, the interphase interpolation is as follows. The total volume associated with SPH particle \(i\) is defined as \[V_{i}=V_{l,i}\left(1+\displaystyle\sum_{j\in\mathcal{B}_{i}}W\left(\bm{r}_{b,j}-\bm{r}_{i},h_{i}\right)V_{b,j}\right),\label{eq:bvols}\] where \(\mathcal{B}_{i}\) is the set of all bubbles within the support radius of particle \(i\). The liquid volume fraction of particle \(i\) is then \[\alpha_{i}=\frac{V_{l,i}}{V_{i}}\label{eq:alp}.\] As the volume of the bubbles is accounted for in the liquid phase, the SPH particles effectively expand in the proximity of gas bubbles. To retain an accurate SPH approximation, the smoothing length of each SPH particle must be adjusted accordingly, to maintain \[\displaystyle\sum_{j\in\mathcal{P}_{i}}W\left(\bm{r}_{j}-\bm{r}_{i},h_{i}\right)V_{j}\approx1\qquad\forall{i}\in\mathcal{P}\label{eq:pou}.\] With both \(V_{i}\) (through [\[eq:bvols\]](#eq:bvols){reference-type="eqref" reference="eq:bvols"}) and the partition of unity [\[eq:pou\]](#eq:pou){reference-type="eqref" reference="eq:pou"} having a non-linear dependence on \(h_{i}\), we cannot choose \(h_{i}\) explicitly to satisfy [\[eq:pou\]](#eq:pou){reference-type="eqref" reference="eq:pou"}. However, by setting \[h_{i}=h_{0}\left(\frac{V_{i}}{V_{l,i}}\right)^{\frac{1}{d}},\label{eq:seth}\] in which \(d=3\) is the number of spatial dimensions, we obtain a system in which the SPH particle distribution expands and contracts in response to the bubble volumes. The response is not instantaneous, but occurs over a finite time, as the volume effects propagate through the particle distribution. However, when averaged over time, this system results in a discretisation for which [\[eq:pou\]](#eq:pou){reference-type="eqref" reference="eq:pou"} is satisfied. The momentum exchange between the bubble and liquid phases is evaluated at each bubble (denoted \(\bm{M}_{b}\)), and then interpolated back to the liquid phase through \[\bm{M}_{i}=V_{i}\displaystyle\sum_{j\in\mathcal{B}_{i}}\bm{M}_{b,j}W\left(\bm{r}_{b,j}-\bm{r}_{i},h_{i}\right)\label{eq:b2l}.\] The bubble-induced turbulence at each bubble \(\nu_{B,b}\) is interpolated back to each SPH particle in the same manner to obtain \(\nu_{B}\). Evaluation of the lift, drag and virtual mass forces for each bubble require the knowledge of the filtered liquid velocity and liquid velocity gradients, and the turbulent kinetic energy, at each bubble location. Additionally, the bubble entrainment model described in Section [10.4](#sec:be){reference-type="ref" reference="sec:be"} requires the turbulent dissipation rate to be interpolated from the liquid to each bubble location. These properties are interpolated from SPH particles to bubble locations through \[\phi_{b,j}=\displaystyle\sum_{i\in{P}}\phi_{i}W\left(\bm{r}_{b,j}-\bm{r}_{i},h_{i}\right)V_{i}\qquad\forall{j\in\mathcal{B}},\label{eq:l2b}\] where \(\phi_{i}\) is the value at particle \(i\) of the property to be interpolated, and \(\phi_{b,j}\) is the interpolated property at bubble \(j\). In our code, we construct an array for each particle \(i\) containing the indices \(\mathcal{P}_{i}\), and a global array containing the indices of the bubble neighbours of all particles: \(\left[\mathcal{B}_{1}\dots\mathcal{B}_{i}\dots\mathcal{B}_{N}\right]\).
### SPH operators {#sec:sph}
For the test filter used to evaluate \(k_{srs}\), and \(q_{c}\) in the LES closure model we use a normalised Shepard filter \[\hat{\tilde{\phi}}_{i}=\frac{\displaystyle\sum_{j\in\mathcal{P}_{i}}\tilde{\phi}_{j}W_{ij}V_{j}}{\displaystyle\sum_{j\in\mathcal{P}_{i}}W_{ij}V_{j}}.\label{eq:shep}\] First derivatives are discretized according to \[\langle\nabla\tilde{\phi}\rangle_{i}=\displaystyle\sum_{j\in\mathcal{P}_{i}}\left(\tilde{\phi}_{j}-\tilde{\phi}_{i}\right)\nabla{W}^{\star}_{ij}{V}_{j}\qquad\langle\nabla\cdot\tilde{\bm{u}}\rangle_{i}=\displaystyle\sum_{j\in\mathcal{P}_{i}}\left(\tilde{\bm{u}}_{j}-\tilde{\bm{u}}_{i}\right)\cdot\nabla{W}^{\star}_{ij}{V}_{j}\label{eq:grad},\] where the corrected kernel gradient \(\nabla{W}^{\star}_{ij}\) due to is used, as detailed in . This provides first order consistency for first derivatives. The Laplacian is approximated using the formulation of as \[\langle\nabla^{2}\tilde{\phi}\rangle_{i}=\displaystyle\sum_{j\in\mathcal{P}_{i}}\frac{2\tilde{\phi}_{ij}}{\lvert\bm{r_{ij}}\rvert^{2}}\bm{r_{ij}}\cdot\nabla{W}_{ij}{V}_{j},\label{eq:lap}\] and for the inhomogeneous "div-grad" operator with spatially varying coefficient \(\kappa\), we use \[\langle\nabla\cdot\left(\kappa\nabla\phi\right)\rangle_{i}=\displaystyle\sum_{j\in\mathcal{P}_{i}}\frac{2\bar{\kappa}_{ji}\tilde{\phi}_{ij}}{\lvert\bm{r_{ij}}\rvert^{2}}\bm{r_{ij}}\cdot\nabla{W}_{ij}{V}_{j},\] with \(\bar{\kappa}_{ji}={2\kappa_{i}\kappa_{j}}/\left(\kappa_{i}+\kappa_{j}\right)\) the harmonic mean. To evaluate the transport velocity \(\bm{u}_{ps}\), we calculate the gradient of the particle number density as \[\nabla{\rho}_{N,i}=\displaystyle\sum_{j\in\mathcal{P}_{i}}\left[1+\frac{1}{4}\left(\frac{W_{ij}}{W_{ii}}\right)^{4}\right]\nabla{W}_{ij}V_{j},\] and then set the transport velocity as \[\bm{u}_{ps,i}=\frac{h_{i}^{2}}{4\delta{t}}\begin{cases}\nabla{\rho}_{N,i}&\forall{i}\in\mathcal{P}_{I}\\ \left(\bm{n}_{i}\cdot\nabla{\rho}_{N,i}\right)\bm{n}_{i}&\forall{i}\in\mathcal{P}_{FS},\label{eq:ups}\end{cases}\] where \(\bm{n}_{i}\) is the unit vector normal to the surface at particle \(i\), \(\mathcal{P}_{I}\) is the set of internal particles, and \(\mathcal{P}_{FS}\) is the set of free-surface particles, identified as in .
## Fractional step approach for isothermally compressible liquid[\[sec:isoT\]]{#sec:isoT label="sec:isoT"}
Our approach to the time integration of [\[eq:mass_le\]](#eq:mass_le){reference-type="eqref" reference="eq:mass_le"} and [\[eq:mom_le\]](#eq:mom_le){reference-type="eqref" reference="eq:mom_le"} combines the methods used in , and the approach described in . We use a fractional step algorithm to solve [\[eq:mass_le\]](#eq:mass_le){reference-type="eqref" reference="eq:mass_le"} and [\[eq:mom_le\]](#eq:mom_le){reference-type="eqref" reference="eq:mom_le"} in our SPH framework. Following the classic projection method of , initially introduced to SPH in , the right hand side of [\[eq:mom_le\]](#eq:mom_le){reference-type="eqref" reference="eq:mom_le"} is split, with viscous and advective terms being applied in a predictor step, and the pressure gradient and any divergence-free body forces (e.g. gravity) being used in a projection step to obtain a velocity field which satisfies the continuity equation [\[eq:mass_le\]](#eq:mass_le){reference-type="eqref" reference="eq:mass_le"}. Splitting [\[eq:mom_le\]](#eq:mom_le){reference-type="eqref" reference="eq:mom_le"} as described above, we obtain
where \(\tilde{\bm{u}}^{\star}_{l}\) is an intermediate velocity, which is not required to be compatible with the continuity equation [\[eq:mass_le\]](#eq:mass_le){reference-type="eqref" reference="eq:mass_le"}. We require the velocity field at the end of the time step \(\tilde{\bm{u}}_{l}^{n+1}\) to satisfy [\[eq:mass_le\]](#eq:mass_le){reference-type="eqref" reference="eq:mass_le"}, and we take the divergence of [\[eq:fs2\]](#eq:fs2){reference-type="eqref" reference="eq:fs2"}, to get \[\begin{gathered}
\alpha^{n+1}\nabla\cdot\tilde{\bm{u}}_{l}^{n+1}=\\-\tilde{\bm{u}}_{l}^{n+1}\cdot\nabla\alpha^{n+1}+\nabla\cdot\left(\alpha^{n+1}\tilde{\bm{u}}_{l}^{\star}\right)-\delta{t}\nabla^{2}\left(\alpha^{n+1}\tilde{p}^{n+1}\right)+\frac{\delta{t}}{Fr^{2}}\bm{e_{g}}\cdot\nabla\alpha^{n+1}.\label{eq:divfs2}\end{gathered}\] Substituting the right hand side of [\[eq:divfs2\]](#eq:divfs2){reference-type="eqref" reference="eq:divfs2"} into the final term of [\[eq:mass_le\]](#eq:mass_le){reference-type="eqref" reference="eq:mass_le"} evaluated at time step \(n+1\), and replacing the time derivative with a backwards Euler difference equation, yields \[\begin{gathered}
\frac{\left(\alpha^{n+1}\tilde{p}^{n+1}-\alpha^{n}\tilde{p}^{n}\right)}{\delta{t}}-\bm{u}_{ps}\cdot\nabla\left(\alpha^{n+1}\tilde{p}^{n+1}\right)=\\\frac{1}{Ma^{2}}\left[\tilde{\bm{u}}_{l}^{n+1}\cdot\nabla\alpha^{n+1}-\nabla\cdot\left(\alpha^{n+1}\tilde{\bm{u}}_{l}^{\star}\right)+\delta{t}\nabla^{2}\left(\alpha^{n+1}\tilde{p}^{n+1}\right)-\frac{\delta{t}}{Fr^{2}}\bm{e_{g}}\cdot\nabla\alpha^{n+1}\right].\label{eq:ppe1}\end{gathered}\] As [\[eq:ppe1\]](#eq:ppe1){reference-type="eqref" reference="eq:ppe1"} explicitly contains \(\tilde{\bm{u}}_{l}^{n+1}\), we substitute [\[eq:fs2\]](#eq:fs2){reference-type="eqref" reference="eq:fs2"} back into [\[eq:ppe1\]](#eq:ppe1){reference-type="eqref" reference="eq:ppe1"}, note that \(\left(1/\alpha\right)\nabla\alpha=\nabla\ln\alpha\), and collect terms containing \(\tilde{p}^{n+1}\), obtaining \[\begin{gathered}
\nabla^{2}\left(\alpha^{n+1}\tilde{p}^{n+1}\right)-\frac{Ma^{2}\alpha^{n+1}\tilde{p}^{n+1}}{\delta{t}^{2}}+\left\{\frac{Ma^{2}\bm{u}_{ps}}{\delta{t}}-\nabla\left(\ln\alpha^{n+1}\right)\right\}\cdot\nabla\left(\alpha^{n+1}\tilde{p}^{n+1}\right)\\=\frac{\alpha^{n+1}}{\delta{t}}\nabla\cdot\left(\tilde{\bm{u}}_{l}^{\star}\right)-\frac{Ma^{2}\alpha^{n}\tilde{p}^{n}}{\delta{t}^{2}}.\label{eq:ppe2}\end{gathered}\] Note that in the combined limits of single phase (\(\alpha=1\)) incompressible (\(Ma=0\)) flow, the standard Poission equation is recovered. In practice, we omit the terms containing \(\bm{u}_{ps}\) from [\[eq:ppe2\]](#eq:ppe2){reference-type="eqref" reference="eq:ppe2"} and [\[eq:fs1\]](#eq:fs1){reference-type="eqref" reference="eq:fs1"}, as their effect on the overall simulation is small for large \(Re\).
## Temporal evolution of both phases
We now describe the complete algorithm for the temporal evolution of the complete system. In the following, each operation is applied to every SPH particle \(i\in\mathcal{P}\) or bubble \(i\in\mathcal{B}\), and we have dropped the subscripts \(i\) for clarity. We introduce the superscripts \(n\), \(\star\) and \(n+1\) to represent properties at the current, intermediate, and next time steps. The algorithm is as follows.
1. Advect particles to intermediate positions according to \(\bm{r}^{\star}=\bm{r}^{n}+\delta{t}\tilde{\bm{u}}_{l}\), and bubbles to new positions according to \(\bm{r}_{b}^{n+1}=\bm{r}_{b}^{n}+\delta{t}\bm{u}_{b}^{n}\).
2. Construct boundary conditions via mirror particles, calculate bubble entrainment and breakup if any (as described in Section [10.4](#sec:be){reference-type="ref" reference="sec:be"}), and build neighbour lists \(\mathcal{P}_{i}\) and \(\mathcal{B}_{i}\).
3. Calculate the transport velocity \(\bm{u}_{ps}\) from [\[eq:ups\]](#eq:ups){reference-type="eqref" reference="eq:ups"}, based on a modified form of the Fickian shifting introduced by , described in detail in .
4. Interpolate bubble volumes into SPH particle locations using [\[eq:bvols\]](#eq:bvols){reference-type="eqref" reference="eq:bvols"} and [\[eq:alp\]](#eq:alp){reference-type="eqref" reference="eq:alp"} to obtain volume fractions \(\alpha^{n+1}\). Adjust smoothing lengths via [\[eq:seth\]](#eq:seth){reference-type="eqref" reference="eq:seth"}. Note that this is a first order approximation to \(\alpha^{n+1}\), but allows us to retain an explicit scheme for the volume fractions.
5. Evaluate \(\nabla\tilde{\bm{u}}_{l}\), \(\varepsilon\), and \(k_{srs}\), and interpolate to bubble positions through [\[eq:l2b\]](#eq:l2b){reference-type="eqref" reference="eq:l2b"}.
6. Evolve [\[eq:langevin2\]](#eq:langevin2){reference-type="eqref" reference="eq:langevin2"} to obtain \(\bm{u}_{l}^{\prime}\) at bubble locations, and evaluate forces in bubbles via [\[eq:bforces\]](#eq:bforces){reference-type="eqref" reference="eq:bforces"}.
7. Update bubble velocities by integrating [\[eq:mom_b\]](#eq:mom_b){reference-type="eqref" reference="eq:mom_b"}, evaluate \(\bm{M}_{b}\) and \(\nu_{B,b}\), and interpolate back to SPH particle positions through [\[eq:b2l\]](#eq:b2l){reference-type="eqref" reference="eq:b2l"}.
8. Evaluate remaining terms in [\[eq:fs1\]](#eq:fs1){reference-type="eqref" reference="eq:fs1"}, and evaluate \(\tilde{\bm{u}}_{l}^{\star}\).
9. Construct and solve equation [\[eq:ppe2\]](#eq:ppe2){reference-type="eqref" reference="eq:ppe2"} to obtain \(\alpha^{n+1}\tilde{p}^{n+1}\), and hence \(\tilde{p}^{n+1}\). The system [\[eq:ppe2\]](#eq:ppe2){reference-type="eqref" reference="eq:ppe2"} is solved using a BiCGStab algorithm with Jacobi preconditioning, with Neumann boundary conditions on solid surfaces, and homogeneous Dirichlet boundary conditions on free-surfaces as in .
10. Evaluate pressure gradient, and project \(\tilde{\bm{u}}_{l}^{\star}\) onto \(\tilde{\bm{u}}_{l}^{n+1}\) using [\[eq:fs2\]](#eq:fs2){reference-type="eqref" reference="eq:fs2"}.
11. Advect particles to final positions \(\bm{r}^{n+1}=\bm{r}^{n}+\delta{t}\left(\tilde{\bm{u}}_{l}^{n}+\tilde{\bm{u}}_{l}^{n+1}+2\bm{u}_{ps}\right)/2\).
The value of \(\delta{t}\) is set adaptively according to criteria for the Courant condition and viscous diffusion, as in . \[\delta{t}=0.2\min\left(\frac{h}{\displaystyle\max_{\mathcal{P}}\left(\lvert\tilde{\bm{u}}_{0}\rvert\right)},Re{h}^{2}\right),\] in which \(\displaystyle\max_{\mathcal{P}}\) is the maximum value over the set of all particles \(\mathcal{P}\). Although our numerical scheme is capable of capturing acoustic waves in the liquid, this does not impose an additional (stability related) constraint on the time-step, as the acoustic physics are treated implicitly in the fractional step approach. The acoustic part of the system is unconditionally stable, although for larger time-steps, the acoustic waves are subject to greater numerical dissipation. There is a trade-off here, between computational costs, and the degree to which acoustic information is of interest. Regardless of the application, there is a benefit of the present approach over a perfectly incompressible approach. Whilst both methods result in smooth pressure fields with no spurious noise (as is commonly experienced in explicit, weakly compressible SPH), the additional terms in [\[eq:ppe2\]](#eq:ppe2){reference-type="eqref" reference="eq:ppe2"} accounting for compressibilty increase the diagonal dominance of the linear system which represents the discrete form of this equation. This increased diagonal dominance renders the system more amenable to solution using iterative methods, which will converge more quickly. In the present application we are not interested in capturing the acoustic waves, and do not impose an additional time-step constraint proportional to \(hMa\). Despite this, for \(Ma=0.05\), we still obtain a reduction in the necessary number of iterations to solve [\[eq:ppe2\]](#eq:ppe2){reference-type="eqref" reference="eq:ppe2"} by a factor of typically around \(5\). For our SPH framework, the solution of [\[eq:ppe2\]](#eq:ppe2){reference-type="eqref" reference="eq:ppe2"} is the most expensive aspect of the simulation, so this presents a significant computational saving. With this in mind, we consider \(Ma\) in this framework to be a numerical parameter (rather than a physical one), as in weakly compressible SPH schemes. The value of \(Ma=0.05\) is consistent with the artificial sound speeds widely used in weakly compressible SPH, ensures the density fluctuations are well below \(1\%\) , and is used throughout this work.
## Bubble entrainment, breakup, and free-surface interaction {#sec:be}
### Entrainment
Our intention is to simulate flows where bubbles are entrained at the free-surface. We use an entrainment model, similar in ethos to that of and , in which a fraction of the turbulent kinetic energy of the liquid is assumed to be converted into surface energy as bubbles are created (or entrained) at the free-surface. The energy available for bubble creation at SPH particle \(i\) in a given time step is \[E_{bc,i}=C_{\varepsilon}\beta\alpha_{i}\varepsilon_{i}V_{i}\delta{t}\label{eq:ebc},\] where \(C_{\varepsilon}=0.01\) is a constant set empirically. The surface energy of a bubble of radius \(a_{b}\) is \[E_{se}=\frac{4\pi{a}_{b}^{2}}{We}\label{eq:bse}.\] The closure models used to evaluate the forces on each bubble are based on the assumption spherical non-interacting bubbles . When the concentration of bubbles is large (and hence \(\alpha\) is small), these assumptions cease to be valid. Therefore, we impose an additional constraint on bubble entrainment, such that \(1-\alpha\lesssim1/3\), by denoting the volume available for bubble entrainment as \[V_{bc,i}=\frac{1}{W\left(0\right)}\left(\frac{3}{2}-\frac{1}{\alpha_{i}}\right)\label{eq:vbc},\] where \(W\left(0\right)\) is the maximum value of the SPH kernel. \(V_{bc,i}\) is the volume of bubbles which must be entrained at SPH particle \(i\) to result in \(\alpha_{i}=2/3\). Following , we only entrain bubbles at the free-surface, and only when \(\varepsilon\) exceeds a certain threshold. Where others treated the bubbles as a continuum, we treat them discretely, and hence our algorithm differs somewhat, despite the principles of energy balancing being the same. Where modelled the bubbles through a set of bubble groups, each with a characteristic size, we are able to model individual bubbles, and hence obtain a continuous bubble size distribution. To achieve this, our entrainment algorithm is as follows. At every time-step, for every free-surface particle \(i\in\mathcal{P}_{FS}\) for which \(\varepsilon_{i}>0.2\) and \(\alpha>2/3\):
1. Evaluate the available energy \(E_{bc,i}\) and volume \(V_{bc,i}\) for bubble creation, according to [\[eq:ebc\]](#eq:ebc){reference-type="eqref" reference="eq:ebc"} and [\[eq:vbc\]](#eq:vbc){reference-type="eqref" reference="eq:vbc"}. Initialise counters for new bubbles, and new potential bubbles: \(n_{nb}=0\) and \(n_{npb}=0\).
2. Generate a potential bubble radius \(a_{b,p}\in\left[We^{-3/5}/40,\delta{r}\right]\) with uniform probability distribution. Increment the new potential bubble counter: \(n_{npb}=n_{npb}+1\).
3. Evaluate the surface energy \(E_{se,p}\) of the potential bubble via [\[eq:bse\]](#eq:bse){reference-type="eqref" reference="eq:bse"}. If \(E_{bc,i}\ge{E}_{se,p}\), there is enough energy. If \(E_{bc,i}<E_{se,p}\) and \(E_{bc,i}/E_{se,p}>\zeta_{E}\), where \(\zeta_{E}\in\left[0,1\right]\) is a uniformly distributed random number, then there is not enough energy for this potential bubble.
4. Evaluate the volume of the potential bubble \(V_{b,p}=4\pi{a}_{b,p}^{3}/3\). If \(V_{bc,i}\ge{V}_{b,p}\), there is enough volume available to create this bubble. If \(V_{bc,i}<V_{b,p}\) and \(V_{bc,i}/V_{b,p}>\zeta_{V}\), where \(\zeta_{V}\in\left[0,1\right]\) is a uniformly distributed random number, then there is not enough volume available for this potential bubble.
5. If the checks in steps 3 and 4 were passed, then make a new bubble \(j\) with radius \(a_{b,j}=a_{b,p}\), position \(\bm{r}_{b,j}=\bm{r}_{i}+\bm{\zeta}_{r}\delta{r}\), where \(\bm{\zeta}_{r}\) is a random vector with elements in \(\left[0,1\right]\), and \(\bm{u}_{b,j}=\tilde{\bm{u}}_{l,i}\). Denote the time of bubble creation as \(T_{b,j}\). Increment the new bubble counter: \(n_{nb}=n_{nb}+1\).
6. If the checks in steps 3 and 4 were passed, reduce the available energy and volume by
::: {.subequations}
\[\begin{aligned}
E_{bc,i}&=E_{bc,i}-E_{se,p}\\ V_{bc,i}&=V_{bc,i}-V_{b,p}. \end{aligned}\]
:::
7. Check whether to continue entraining bubbles. If all the inequalities
::: {.subequations}
\[\begin{aligned}
E_{bc,i}&>0\\ V_{bc,i}&>0\\ n_{nb}&<10\\ n_{npb}&<n_{nb}+5\end{aligned}\]
:::
are satisfied, return to step 2. Otherwise, move on to the next SPH particle.
The addition of the stochastic processes in steps 3 and 4 to allow some bubbles for which \(E_{se,p}>E_{bc,i}\) and \(V_{b,p}>V_{bc,i}\), based on the remainders, prevents the creation of large bubbles from being overly suppressed. The maximum potential bubble size \(a_{b,p}\le\delta{r}\) is set to ensure that the bubbles are smaller than the implicit LES filter scale \(\widetilde{\Delta}\). The minimum potential bubble size \(a_{b,p}\ge{We}^{-3/5}/40\) is equal to \(0.05a_{H}\), where \(a_{H}\) is the Hinze scale bubble radius evaluated a priori. This limit is imposed to prevent the generation of a very large number of very small bubbles, which would impose a significant computational cost on the simulation, but which play little role in the overall dynamics of the flow. In future, we consider that the present framework could be extended to treat bubbles with \(a_{b}\ll{a}_{H}\) as a continuum, as is done for all bubble sizes in "Eulerian-Eulerian" schemes.
### Breakup
As in work by and , we use a stochastic breakup model based on energy balance considerations. Our model is similar to , and is based on the imbalance between the surface restoring pressure and the stress on the bubble surface due to the motion of the liquid. For a bubble of size \(a_{b}\), the surface restoring pressure is \(6/\left(2a_{b}We\right)\). The stress exerted on the bubble by the turbulent motion of the liquid is \(\frac{1}{2}{c}_{def}\left(2\varepsilon{a}_{b}\right)^{2/3}\), where \(c_{def}=8.2\) as given in . The net deforming stress on the bubble is given by the difference, from which we obtain an expression for a characteristic deformation velocity \[u_{def}=\text{sgn}\left(c_{def}\left(2\varepsilon{a}_{b}\right)^{\frac{2}{3}}-\frac{6}{{a}_{b}{We}}\right)\sqrt{\left\lvert{c}_{def}\left(2\varepsilon{a}_{b}\right)^{\frac{2}{3}}-\frac{6}{{a}_{b}{We}}\right\rvert}\] In works such as and which model the dispersed phase through a continuous bubble number density field, a characteristic timescale of breakup is often given by \[T_{bu}=\frac{2a_{b}}{u_{def}}=\frac{2a_{b}}{\sqrt{c_{def}\left(2\varepsilon{a}_{b}\right)^{\frac{2}{3}}-\frac{6}{{a}_{b}{We}}}},\] assuming \(u_{def}>0\). In this work, we treat each bubble indivudually, and hence we can construct a model which attempts to account for the residence time of the bubble, and the fact that deformation and breakup occur over a finite time . For every bubble we numerical integrate the deformation velocity to obtain a "deformation distance" \[L_{def}=\max\left\{\displaystyle\int_{t_{bc,i}}^{t}u_{def}\left(\tau\right)d\tau,0\right\},\label{eq:defdist}\] which is a measure of how deformed the bubble is. Here \(t_{bc,i}\) is the time of creation of bubble \(i\). Since we allow \(u_{def}<0\), [\[eq:defdist\]](#eq:defdist){reference-type="eqref" reference="eq:defdist"} accounts for both deformation of the bubble due to turbulence, and relaxation of the bubble back towards a spherical shape. We assume that the bubble breaks once the deformation distance exceeds the bubble diameter: when \(L_{def}>2a_{b}\). We assume all breakup events are binary; that is, a parent bubble splits into two child bubbles. Once a bubble has been marked for a breakup event, we set \(V=V_{b,child}/V_{b,parent}\in\left[0,1\right]\) randomly with the probability distribution given by , where \[P\left(V\right)\propto\begin{cases}V^{-2/3}\left(1-V\right)^{-2/3}\left(V^{2/9}-\Lambda^{5/3}\right)\left[\left(1-V\right)^{2/9}-\Lambda^{5/3}\right]&V\in\left[V_{min},V_{max}\right]\\0&\text{otherwise}\end{cases}\label{eq:pdf}\] Here \(V_{min}\) is set to limit the size distributions to regions where the expression for \(P\left(V\right)\) is positive, with an additional hard limit of \(V_{min}\ge10^{-3}\), to ensure that breakup events do not occur in a capillary-action dominated regime. \(P\left(V\right)\) is even about \(V=1/2\), and \(V_{max}=1-V_{min}\). To normalise \(P\left(V\right)\), and evaluate the cumulative density function (required for generating a random number with distribution \(P\left(V\right)\)), we integrate [\[eq:pdf\]](#eq:pdf){reference-type="eqref" reference="eq:pdf"} numerically. \(\Lambda\) represents a critical bubble radius (relative to the parent bubble), below which the confining stresses due to surface tension exceed the turbulent stresses-it is the largest bubble which will not break due to the turbulent flow (at a given instant in time and space). For each breaking event, we evaluate \(\Lambda\) as \[\Lambda=\left(\frac{12}{c_{def}We}\right)^{3/5}\bar{\varepsilon}^{-2/5}\left(2a_{b,p}\right)^{-1},\] where \[\bar{\varepsilon}=\frac{1}{t-t_{bc,i}}\displaystyle\int_{t_{bc,i}}^{t}\varepsilon\left(\tau\right){d}\tau,\] is the mean dissipation rate of the bubble lifetime, with \(\varepsilon\) the instantaneous dissipation rate in the liquid at bubble \(i\). In the above model, a critical value of \(\Lambda_{crit}=2^{-6/45}\approx0.912\) exists, and the model is not valid for \(\Lambda>\Lambda_{crit}\) (\(P\left(V\right)<0\) \(\forall{V}\)). This effectively imposes an additional limit on the smallest bubble which can break, which is not necessarily consistent with the breakup criteria obtained by integration of \(L_{def}\). Even if the evaluation of \(L_{def}\) from [\[eq:defdist\]](#eq:defdist){reference-type="eqref" reference="eq:defdist"} suggests a bubble should break, if it has a corresponding \(\Lambda>\Lambda_{crit}\) it does not, as the child bubble sizes are undefined in [\[eq:pdf\]](#eq:pdf){reference-type="eqref" reference="eq:pdf"}. In practice, this situation rarely occurs, and has neglible impact on the overall bubble population dynamics. Figure [\[fig:breakupspecmodels\]](#fig:breakupspecmodels){reference-type="ref" reference="fig:breakupspecmodels"} shows \(P\left(V\right)\) for various values of \(\Lambda\). We see that for small \(\Lambda\) the child size distribution is largely flat, but with peaks at very small and very large bubbles (more apparent in the inset). As \(\Lambda\) increases, these peaks reduce, and for \(\Lambda=0.4\) the distribution is nearly flat. For larger \(\Lambda\), the possible range of child sizes decreases, with an increasing dominance of equal-sized breakup events. For \(\Lambda=\Lambda_{crit}\), all breakup events result in \(V=1/2\). Whilst binary break-up models have been utilised in a range of works , recent work has suggested that the binary breakup mechanism presents some limitation. DNS of individual bubbles breaking in isotropic turbulence show that even for a binary breakup event, a bubble filament joining breaking bubbles forms, and this filament ruptures into one or many small bubbles due to capillary effects. This mechanism provides a non-local element to the bubble size cascade, and has been suggested as responsible for the majority of sub-Hinze scale bubble formation . A breakup model constructed on the framework above, but which includes the formation of multiple small bubbles by capillary action, is an active area of development for us. However, at this stage we prefer to adopt a simple binary breakup model, and focus on the free-surface-bubble interactions.
### Free-surface interactions
In the vicinity of free-surfaces, the assumption that the fluid around the bubble is uniform and infinite does not hold, and the closure models for the forces in [\[eq:bforces\]](#eq:bforces){reference-type="eqref" reference="eq:bforces"} are not valid. The dynamics of the interactions between bubbles and free-surfaces are complex, even for individual bubbles in stationary reservoirs, and can include a direct rise to surface bursting, oscillatory bouncing behaviour , and long-term persistance of the bubble at the free-surface. Despite this, most mesh-based Eulerian-Lagrangian schemes simply represent the free-surface via a rigid free slip condition (e.g. ), or through a volume-of-fluid (VOF) approach to resolve the gas phase above the free-surface (e.g. ). In the latter approach, both the deceleration of bubbles on approach to the free-surface and the free-surface deformation are captured, although this occurs solely through modification of the relative densities in the closure models used to evaluate the forces on the bubbles. These closure models still assume a uniform liquid field surrounding the bubbles, and the complex and small scale physics involved are not included. Detailed modelling of free-surface-bubble interactions are beyond the scope of this work. However, some mechanism to describe the behaviour of the interaction is necessary, to prevent bubbles from simply rising due to buoyancy and leaving the domain. For each SPH particle, a surface normal vector \(\bm{n}\) is evaluated (and smoothed), as in . The surface normal vectors are then interpolated to each bubble position through [\[eq:l2b\]](#eq:l2b){reference-type="eqref" reference="eq:l2b"} to obtain \(\bm{n}_{b}\). For each bubble we construct a parameter \(\psi_{fs}\) which identifies when a bubble is in proximity to the free-surface as \[\psi_{fs}=\frac{\left\lvert\bm{n}_{b}\right\rvert}{\left\lvert\bm{n}_{0}\right\rvert{V}_{lb}},\] where \(\left\lvert\bm{n}_{0}\right\rvert\) is the magnitude of the surface normal vector at a plane surface, and \(V_{lb}\) is the SPH volume interpolated to the bubble location (i.e. [\[eq:l2b\]](#eq:l2b){reference-type="eqref" reference="eq:l2b"} with \(\phi=1\)). For bubbles far from the free-surface, \(\psi_{fs}\approx0\). For bubbles within the support radius of free-surface SPH particles, \(\psi_{fs}\) smoothly increases, to approximately \(1\) when bubbles are located at the free-surface. Denoting \(\hat{\bm{n}}_{b}=\bm{n}_{b}/\left\lvert\bm{n}_{b}\right\rvert\), the relative normal velocity between a bubble and the free-surface is \(\bm{u}_{rel}\cdot\hat{\bm{n}}_{b}\). Bubbles with \(\psi_{fs}>0.1\) and \(t-t_{bc}>T_{c}\) are flagged to interact with the free-surface. Here \(T_{c}=\delta{r}/\left\lvert\bm{u}_{rel}\cdot\hat{\bm{n}}_{b}\right\rvert\) is a threshold age (which varies as a bubble evolves), designed to prevent bubbles from destroyed at the moment of entrainment. When a bubble is marked for free-surface interaction, it is given an expected merge time \(t_{m}=t+T_{c}\), at which it will be located at the free surface. For bubbles interacting with the free-surface, [\[eq:mom_b\]](#eq:mom_b){reference-type="eqref" reference="eq:mom_b"} is modified as \[V_{b,i}\frac{d\bm{u}_{b,i}}{dt}=\bm{F}_{d}+\bm{F}_{l}+\bm{F}_{vm}+\bm{F}_{g}+\bm{F}_{fs},\label{eq:mom_b_fs}\] where the free-surface interaction force \(\bm{F}_{fs}\) is \[\begin{gathered}
\bm{F}_{fs}=\frac{1}{2}\left(1+\text{erf}\left(2\log\left(5\psi_{fs}\right)\right)\right)\times\\\frac{\bm{n}_{b}}{\left\lvert\bm{n}_{b}\right\rvert}\left\{\left\lvert\bm{u}_{rel}\cdot\hat{\bm{n}}_{b}\right\rvert\left(1+C_{vm}\beta\right)\frac{V_{b}}{\max\left(t_{m}-t,\delta{t}\right)}-\left[\bm{F}_{b}+\bm{F}_{d}+\bm{F}_{l}+\beta{C}_{vm}V_{b}\frac{d\bm{u}_{l}}{dt}\right]\cdot\frac{\bm{n}_{b}}{\left\lvert\bm{n}_{b}\right\rvert}\right\}.\end{gathered}\] The first term \(\frac{1}{2}\left(1+\text{erf}\left(2\log\left(5\psi_{fs}\right)\right)\right)\) varies smoothly from \(0\) when \(\psi_{fs}\) is small, to \(1\) for large \(\psi_{fs}\), and is approximately \(1\) for \(\psi_{fs}>0.4\). This term ensures the free-surface interaction force is switched on smoothly as a bubble approaches the free-surface. The free-surface interaction force decelerates the bubble in the direction normal to the free-surface, such that the bubble velocity approaches the free-surface velocity over the timescale \(T_{c}\) (or \(\delta{t}\), whichever is greater). The momentum exchange is modified to include \(\bm{F}_{fs}\) as \[\bm{M}=\left(-\bm{F}_{d}-\bm{F}_{l}-\bm{F}_{vm}-\bm{F}_{fs}\right)/\beta.\] We additionally impose a special treatment for bubbles interacting with spray. Where an individual SPH particle \(i\) becomes separated from the bulk of the fluid (such that \(\mathcal{P}_{i}\) contains only \(i\)), it is subjected only to a gravitational body force, and all other terms in [\[eq:mom_le\]](#eq:mom_le){reference-type="eqref" reference="eq:mom_le"} are set to zero. If a bubble has fewer than \(20\) (in three dimensions) SPH particle neighbours, *and* all of those SPH particle neighbours are identified as free-surface particles, then the bubble is discounted, and assumed to have burst. This provides increased stability in regions of violent spray, such as around breaking waves, and prevents bubbles from becoming completely isolated from the SPH simulation. The effect of the free-surface interaction model is demonstrated in Figure [\[fig:rise\]](#fig:rise){reference-type="ref" reference="fig:rise"}. A tank of still water is simulated, with an individual bubble rising due to buoyancy. We vary \(a_{b}\), and taking the characteristic velocity scale as the terminal velocity of the bubble \(u_{t}\), the bubble Reynolds number varies as \(Re=10^{5}u_{t}a_{b}\), and the bubble-scale Weber number is \(We_{b}=1.4\times{10}^{4}a_{b}u_{t}^{2}\). Figure [\[fig:rise\]](#fig:rise){reference-type="ref" reference="fig:rise"} shows the bubble trajectory (left) and rise velocity (right) for several values of \(We_{b}\). In both plots, a time shift has been applied so that the coordinate on the abscissa corresponds to the time since the bubble-surface interaction begins. We see that for larger \(We_{b}\), the bubble is decelerated more quickly. For all \(We_{b}\in\left[0.01,4.06\right]\), the bubble comes to rest on the free-surface. For larger bubble Weber numbers \(We_{b}\) (corresponding to larger \(a_{b}\)), the system becomes less stable after the bubble reaches the surface, and this is due to the settling over a finite time of the SPH particles to accommodate the bubble volume. In all cases, though, the final position of the bubble remains within a distance \(\delta{r}\) of the free-surface location. The final feature of the free-surface interaction model is the persistence time. Once a bubble has reached the free-surface, it remains subject to [\[eq:mom_b\_fs\]](#eq:mom_b_fs){reference-type="eqref" reference="eq:mom_b_fs"} for a persistence time \(T_{p}\). If the motion of the liquid phase is such that the bubble (now moving with the liquid) moves away from the free-surface (i.e., if \(\psi_{fs}<0.1\)), then the bubble is assumed to no longer interact with the free-surface, and its motion is again governed by [\[eq:mom_b\]](#eq:mom_b){reference-type="eqref" reference="eq:mom_b"}. In reality, when bubbles reach a free-surface a thin lubrication layer forms, and the liquid in this layer drains until the layer ruptures (see e.g. for a description of the mechanism). This process is complex, and is strongly influenced by the local geometry and the salinity (or other contaminants and surfactants) and gradients thereof. We cannot seek to accurately capture this process in our model, and instead we seek an order of magnitude estimate for \(T_{p}\) which has a physical basis. We use the (here made non-dimensional) expression of , where \[T_{p}=\frac{We^{3/4}}{Re}Fr^{1/2}Sc{a}_{b}^{1/2},\label{eq:tp}\] in which \(Sc\) is the Schmidt number, taken as \(Sc=700\) for sea water at \(20^{\circ}C\) . Equation [\[eq:tp\]](#eq:tp){reference-type="eqref" reference="eq:tp"} is based on a mechanistic model for film drainage, acounting for molecular diffusion and curvature-pressure induced flow, and provides a rough scaling of the persistence time, with a level of approximation which is consistent with the present framework: a more detailed physical model is beyond the scope of this work. Bubbles which remain in proximity to the free-surface beyond \(t_{m}\) are assumed to burst (and are removed from the simulation) at \(t=t_{m}+T_{p}\).
# Bubble plumes {#sec:plume}
We first use our model to simulate a buoyant bubble plume in a tank. Our simulation is configured to match the experimental set up of as follows. The domain is a unit cube of liquid, with no-slip conditions at the lateral and lower boundaries, and a free-surface boundary at the upper surface of the liquid. The origin of our coordinate system is at the centre of the base of the tank, with the unit vector in the \(z\) direction pointing upwards. A bubble sparger is simulated, centred at \(\left(x,y,z\right)=\left(0,0,0.09\right)\), and radius \(r_{0}=0.02\). The governing parameters are \(\beta=1000/1.2\), \(Re=10^{6}\), \(We=1.4\times{10}^{4}\), and \(Fr=1/\sqrt{9.81}\). The (dimensionless) volumetric flow rate of the bubble sparger is \(\dot{Q}=1.67\times{10}^{-5}\). With this dimensionless scaling, the Hinze scale is estimated at \(a_{H}=We^{-3/5}/2=1.62\times{10}^{-3}\). Initially, we simulate the release of uniform bubbles with \(a_{b}=10^{-3}\), as in . The left panel of Figure [\[fig:fraga_profile\]](#fig:fraga_profile){reference-type="ref" reference="fig:fraga_profile"} shows the evolution of the mean dissipation rate in the liquid for several resolutions. For all resolutions there are differences in the dissipation rate, although for \(\delta{r}<1/100\), the dissipation rates approximately converge for \(t>6\). We note that for the isotropic turbulence with \(Re=10^{6}\) investigated in Appendix [\[les\]](#les){reference-type="ref" reference="les"}, a resolution of \(\delta{r}=1/100\) is sufficient to yield accurate dissipation rates compared with the high-order reference data in . For the remainder of this section we set \(\delta{r}=1/100\). The right panel of Figure [\[fig:fraga_profile\]](#fig:fraga_profile){reference-type="ref" reference="fig:fraga_profile"} shows the variation of the mean vertical velocity in the liquid with radial position, at different depths. Here the vertical velocity is scaled by the (local) velocity at the plume centre \(w_{c}\left(z\right)\), and the radial position is scaled by the local plume width \(b_{v}\), taken as \(w\left(r=b_{v}\left(z\right),z\right)=e^{-1}w_{c}\left(z\right)\). The blue symbols indicate the results of our SPH framework just above (\(z=0.555\), circles) and below (\(z=0.453\), triangles) the centre of the tank. The numerical (red circles) and experimental (red crosses) data from are also shown, as is a Gaussian distribution (dashed black line). We see a good agreement at both vertical locations with both the Gaussian profile and the results of . For larger \(r/b_{v}>1.5\) our results deviate from the Gaussian profile, but match those of , tending towards a non-zero \(w/w_{c}\). As discussed in , this deviation from the Gaussian is due to the confined plume in our simulations. We now simulate a plume in the same configuration, but with a polydisperse bubble size distribution. The initial bubble size distribution is a truncated Gaussian with mean \(a_{b}=10^{-3}\), standard deviation \(10^{-3}\), and a minimum radius cut-off at \(a_{b}=10^{-4}\). Figure [\[fig:plume_a\_v_urel\]](#fig:plume_a_v_urel){reference-type="ref" reference="fig:plume_a_v_urel"} plots the magnitude of the relative velocity \(\lvert\bm{u}_{b}-\tilde{\bm{u}}_{l}\rvert\) against the bubble size near the centre of the tank, at \(z=0.555\). Each data point corresponds to a separate bubble, and the data are collected for \(t\in\left[10,20\right]\). We see a clear, almost linear trend, with larger bubbles rising faster relative to the liquid than smaller ones. There is also an obvious discrepancy between the results with (black symbols) and without (red symbols) the Langevin model for sub-resolution fluctuations. The Langevin model increases the variation of the relative velocity (without changing the mean). This observation is consistent with Figure [\[fig:plume_traces\]](#fig:plume_traces){reference-type="ref" reference="fig:plume_traces"}, which shows the traces of a random sample of bubbles as they rise through the plume. The positions of bubbles are projected onto polar coordinates \(\left(r/r_{0},z\right)\), and are coloured by the bubble radius. The left panel shows bubble traces where we set \(\bm{u}_{l}^{\prime}=\bm{0}\), whilst for the right panel, we calculate \(\bm{u}_{l}^{\prime}\) using the Langevin model. Without the Langevin model, there is a clear trend for larger bubbles to migrate away from the centre of the plume, and smaller bubbles to rise more vertically. This self-organising phenomenon has been previously observed in numerical simulations , and experimentally . The phenomenon can be explained by the linear dependence of the lift force on the relative velocity: as bubbles rise through the plume, the relative velocity is close to orthogonal to the gradient of the velocity in the liquid, and as larger bubbles have a larger relative velocity, they experience an increase lift force, resulting in greater lateral migration. When the Langevin model is included, the bubbles still self-organise, but there is increased lateral migration of small bubbles due to the sub-resolution fluctuations. This self-organisation is clearly depicted in Figure [\[fig:plume_r\_v_a\]](#fig:plume_r_v_a){reference-type="ref" reference="fig:plume_r_v_a"}, which plots the normalised bubble radius \(a_{b}/a_{H}\) against radial position \(r/r_{0}\) at several depths. At \(z=0.1\) (frame a)), just above the base of the plume, the bubble sizes are evenly distributed across the radius of the plume. Further up the plume, the bubbles migrate radially outwards, and the plume gets wider. There is a clear trend, both with (black symbols) and without (red symbols) the Langevin sub-resolution model, for larger bubbles to migrate further outwards, as can be seen in frames b) to d) of Figure [\[fig:plume_r\_v_a\]](#fig:plume_r_v_a){reference-type="ref" reference="fig:plume_r_v_a"}. When the Langevin model is included, this trend remains, although as observed in [\[fig:plume_traces\]](#fig:plume_traces){reference-type="ref" reference="fig:plume_traces"}, the lateral migration of smaller bubbles is increased-the greater spread in radial position of small bubbles is visible in frames b) to d). This behaviour is expected, as small scale fluctuations have a greater effect on small bubbles than larger bubbles through the increased relative importance of the drag force. Returning to Figure [\[fig:plume_traces\]](#fig:plume_traces){reference-type="ref" reference="fig:plume_traces"}, both with and without the Langevin model, the effect of the free-surface interaction model is clear: as bubbles approach the free-surface, they decelerate and move with the liquid, which takes them radially outwards. This behaviour is in qualitative agreement with numerical simulations , although in both cases, these works relied on resolution of the motion of the gas flow above the free-surface to capture the bubble-free-surface interactions.
# Breaking waves {#sec:bw}
We now use our numerical framework to simulate bubble entrainment in a breaking wave. We consider a periodic \(3^{rd}\)-order Stokes wave, which has been extensively studied in the literature with multi-phase mesh-based schemes . We take the wavelength as the integral length-scale, and the Froude speed as the characterstic velocity scale. The domain is a unit cube, periodic in the lateral directions, with a free-slip condition at the lower boundary. We set \(Re=4\times{10}^{4}\), \(Fr=1\), \(We=1.98\times{10}^{4}\) and \(\beta=1000/1.2\). These parameters correspond to the configuration of (although we note here that with our scaling, wave overturning occurs shortly after \(t=1\), as in and ), and gives a Bond number of \(Bo=We\left(\beta-1\right)/\left(4\pi^{2}\beta\right)=500\). The location of the initial free-surface is given by \[\eta\left(x\right)=\frac{1}{2\pi}\left\{\chi\cos\left(2\pi{x}\right)+\frac{1}{2}\chi^{2}\cos\left(4\pi{x}\right)+\frac{3}{8}\chi^{3}\cos\left(6\pi{x}\right)\right\},\] where \(\chi\) is the initial wave steepness. The velocity within the liquid is given by
and \(w\left(t=0\right)=0\). In many of the results that follow, we refer to the time since impact, \(t-t_{im}\), where \(t_{im}\) is the time at which the wave first breaks, indicated by a sharp increase in the (spatial) mean value of \(\varepsilon\), and confirmed visually. An estimate of the Hinze scale is \(a_{H}=We^{-3/5}/2\approx1.36\times{10}^{-3}\), and throughout the following, we report bubble sizes as relative to the Hinze scale. First, we perform simulations with \(\chi=0.55\) in the absence of bubbles, to ensure our SPH model provides a converged solution. Figure [\[fig:wave_conv\]](#fig:wave_conv){reference-type="ref" reference="fig:wave_conv"} shows variation of the mean kinetic energy for the single-phase wave with time. We see that the evolution of the energy is approximately converged for resolutions of \(\delta{r}\le1/300\). Our single phase SPH model does not contain surface tension, corresponding to the large Weber number limit. Accordingly, there is no physical limit imposed on the minimum droplet sizes expected during the breaking process. However, we note that the numerics of the SPH algorithm-specifically the particle shifting technique-introduces a surface tension-like effect into the simulation. Although we are unable to accurately quantify this effect, we note that droplets consisting of individual SPH particles are formed during the simulation at all resolutions, and hence the effective Weber number of the single phase SPH simulation is governed by resolution, for the range of resolutions explored. For the purposes of the present study, this degree of convergence is adequate, and except where otherwise specified, we set \(\delta{r}=1/300\) in the following.
## Bubble size distributions and the effect of resolution
We keep \(\chi=0.55\) and now include bubbles in our simulation. Figure [\[fig:wave_img1\]](#fig:wave_img1){reference-type="ref" reference="fig:wave_img1"} shows the wave at several instants during the first (dimensionless) time unit after breaking, from the side (left) and beneath (right). A movie showing the same views of the wave during the complete breaking process is available in the supplementary material. As the plunging breaker impacts on the surface below, bubbles are entrained in the region of impact (panels a) and f)). At the early times after impact (panels a-c) and f-h)) the flow is largely two-dimensional, with little variation in the transverse direction. Later, as the (resolved) barrel continues to roll, three-dimensional structure forms under the wave (panels d,e) and i,j)). In our simulation, air entrainment occurs through several mechanisms. Firstly, bubbles are entrained on impact of the plunging breaker. Secondly, bubbles are entrained at the surface ahead of the wave due to the impact of spray forward from the plunging breaker; this effect is particularly clear in panel i) of Figure [\[fig:wave_img1\]](#fig:wave_img1){reference-type="ref" reference="fig:wave_img1"}. At approximately \(t-t_{im}=0.8\) (panels e) and j) in Figure [\[fig:wave_img1\]](#fig:wave_img1){reference-type="ref" reference="fig:wave_img1"}), the barrel collapses, and results in further entrainment. We note here a limitation of our approach. As our model for the continuum is single phase (we do not resolve the gas above the free-surface), although we predict the entrainment of a large barrel-shaped bubble, the mass of this entrained air is partially lost when this bubble collapses. This limitation is common to all single-phase SPH models, and also to the model of . Our model does predict bubble entrainment as the barrel collapses, due to the large values of \(\varepsilon\) as the free-surfaces come together. Combining the present work with a multi-phase SPH scheme, such that large bubbles are resolved, and bubbles with \(a_{b}<\delta{r}\) are modelled, is an active area of research for us, but we reserve such a model for a future work. The left panel of Figure [\[fig:wave_pop1\]](#fig:wave_pop1){reference-type="ref" reference="fig:wave_pop1"} shows the evolution of the total volume of entrained bubbles (normalised by the total liquid volume), for several resolutions. There is roughly linear growth volume of air entrainment (the dashed black trendline has slope of \(1\)), as found in the adaptive-mesh simulations of . The right panel of Figure [\[fig:wave_pop1\]](#fig:wave_pop1){reference-type="ref" reference="fig:wave_pop1"} shows the evolution of the ratio \(N_{H-}/N_{H+}\) for several resolutions, where \(N_{H-}\) is the total number of sub-Hinze scale bubbles, and \(N_{H+}\) is the number of bubbles larger than the Hinze scale. For each resolution \(N_{H-}/N_{H+}\) is roughly constant, though it varies between resolutions. Our framework only admits bubbles smaller than the SPH resolution, and so for smaller \(\delta{r}\), the maximum bubble size is smaller. This reduction in the available range of super-Hinze scale bubble sizes reduces the number of super-Hinze scale bubbles. Given our entrainment model is based on considerations of energy and volume balance (which are roughly invariant with \(\delta{r}\)), the total volume entrained is approximately the same for all three values of \(\delta{r}\) in the right panel of Figure [\[fig:wave_pop1\]](#fig:wave_pop1){reference-type="ref" reference="fig:wave_pop1"}. Hence, we see a corresponding increase in sub-Hinze scale bubbles as \(\delta{r}\) is decreased. The reduction in entrained volume at approximately \(t-t_{im}=0.2\) for all resolutions (visible in the left panel of Figure [\[fig:wave_pop1\]](#fig:wave_pop1){reference-type="ref" reference="fig:wave_pop1"}) corresponds to the period shown in panels b-c) and g-h) of Figure [\[fig:wave_img1\]](#fig:wave_img1){reference-type="ref" reference="fig:wave_img1"}, when a portion of the bubbles entrained during the initial impact are ejected in the spray. A similar pattern is visible in the total bubble population evolution in the results of . Figure [\[fig:wave_bsd1\]](#fig:wave_bsd1){reference-type="ref" reference="fig:wave_bsd1"} shows the bubble size distribution \(P\left(a_{b}/a_{H}\right)\) for various resolutions at two times after impact. The left panel corresponds to \(t-t_{im}=0.09\), and panels b) and g) in Figure [\[fig:wave_img1\]](#fig:wave_img1){reference-type="ref" reference="fig:wave_img1"}, whilst the right panel corresponds to \(t-t_{im}=0.84\), and panels e) and j) in Figure [\[fig:wave_img1\]](#fig:wave_img1){reference-type="ref" reference="fig:wave_img1"}. In both panels of Figure [\[fig:wave_bsd1\]](#fig:wave_bsd1){reference-type="ref" reference="fig:wave_bsd1"}, the Deane and Stokes spectrum is plotted (dashed black lines), with slope \(-3/2\) for sub-Hinze bubbles, and \(-10/3\) for super-Hinze bubbles. Our model yields bubble size distributions which match the expected Deane and Stokes spectrum, over more than an order of magnitude of bubble radii. This match includes accurately predicting the change in slope of the bubble size distribution about the Hinze scale. This result is significant, given the simplicity of our entrainment and breakup models, which are based on simple energy and volume balance arguments. The shift in bubble size distribution as \(\delta{r}\) is reduced is clear in Figure [\[fig:wave_bsd1\]](#fig:wave_bsd1){reference-type="ref" reference="fig:wave_bsd1"}. For smaller \(\delta{r}\), \(P\left(a_{b}/a_{H}\right)\) drops below the \(-10/3\) slope and decays to zero at a smaller \(\delta{r}\), whilst the extent to which the sub-Hinze \(-2/3\) slope is predicted increases. The match between the expected and simulated bubble size distribution is consistent across both time instants, although \(P\left(a_{b}/a_{H}\right)\) shows more noise at early times, as the bubble population provides a smaller sample then. We again mention a limitation of our framework, and of models of this type. Our entrainment model is relatively simple, and as such the local (in time and space) entrainment bubble size spectrum depends only on \(\varepsilon\) and \(\alpha\), and not on the mechanism of bubble entrainment. The three separate entrainment processes discussed earlier-by plunging breaker, spray impact, and barrel bubble collapse-in reality involve quite different mechanisms. Our model, however, cannot differentiate between them. This similarly applies to the entrainment models used in Eulerian-Eulerian work . It is this limitation which drives us to propose the coupling of a multi-phase SPH scheme, with the gas above the liquid surface resolved, to the discrete bubble model. This would enable the entrainment of large scale bubbles to be captured more accurately, with the computationally cheaper discrete bubble model used once large bubbles have broken to \(a_{b}\approx\delta{r}\), and for breakup, entrainment and tracking of smaller bubbles. Figure [\[fig:wave_late\]](#fig:wave_late){reference-type="ref" reference="fig:wave_late"} shows the wave at several times after \(t-t_{im}=1\). At \(t-t_{im}\approx{2}\), a large cloud of bubbles is entrained through a reverse breaking wave, visible in panel a) of Figure [\[fig:wave_late\]](#fig:wave_late){reference-type="ref" reference="fig:wave_late"}. Our simulations predict the obliquely descending eddies observed by , and modelled by and . The obliquely descending eddies drag a cloud of bubbles downwards away from the surface, which persists for several time units, and is clearly visible in panels b) and c). The observations of bubble distribution between breaking waves predicted by our model, both at early (Figure [\[fig:wave_img1\]](#fig:wave_img1){reference-type="ref" reference="fig:wave_img1"}) and late (Figure [\[fig:wave_late\]](#fig:wave_late){reference-type="ref" reference="fig:wave_late"}) times, is in good qualitative agreement with the Eulerian-Eulerian model of , the detailed simulations of and the experimental observations of .
## Influence of wave steepness \(\chi\)
Finally we change the wave steepness \(\chi\), to cover both breaking and non-breaking waves. Figure [\[fig:slopeimg\]](#fig:slopeimg){reference-type="ref" reference="fig:slopeimg"} shows the wave profiles (coloured by dissipation rate), and bubbles (coloured by \(a_{b}/a_{H}\)), for several wave steepnesses \(\chi\in\left[0.3,0.5\right]\). The wave with \(\chi=0.3\) (panel a)) doesn't break. For \(\chi=0.33\) and \(\chi=0.35\) the wave breaks by overspilling, whilst as \(\chi\) is increased further, there is a transition to the plunging breaker studied in the previous section. As a limiting test, we observe that for the case where the wave does not break, no bubbles are entrained, and that for all cases where the wave (visibly) breaks, bubbles are entrained. Figure [\[fig:wave_slopevol\]](#fig:wave_slopevol){reference-type="ref" reference="fig:wave_slopevol"} shows the time evolution of the total entrained volume (normalised by the total liquid volume), for various wave steepnesses \(\chi\in\left[0.33,0.55\right]\). For \(\chi=0.33\) and \(\chi=0.35\), the total entrained volume remains small, although the wave breaking and entrainment process has a long duration. As \(\chi\) increases and the plunging breaker regime is approached, the total volume of air entrained increases significantly, whilst the peak moves earlier, to approximately \(1.5\) time units after breaking. As in Figure [\[fig:wave_pop1\]](#fig:wave_pop1){reference-type="ref" reference="fig:wave_pop1"}, the linear growth of the entrained volume is clearly visible in the plunging breaker regime.
# Conclusions {#sec:conc}
In recent years the potential of Smoothed Particle Hydrodynamics (SPH) for simulations of free-surface flows has been widely demonstrated, but limitations in adaptivity remain in comparison to adaptive-mesh methods, preventing the use of SPH high-fidelity simulations of bubbly flows. Approaches which resolve a continuous liquid phase, and include bubbles as discrete Lagrangian particles are established in the mesh-based community, but have not previously been developed in a mesh-free framework, despite the benefits of mesh-free approaches for free-surface flows. In this work we have presented a numerical framework for Large Eddy Simulations (LES) of bubbly, free-surface flows. The framework employs a continuum-discrete structure, where we use SPH for the LES of the liquid phase, whilst treating bubbles as discrete Lagrangian particles, which interact with the liquid via exchanges of volume and momentum. We introduce a Langevin model for the sub-resolution velocity fluctuations, and several closure models for bubble break-up, entrainment, and free-surface interaction. The modification of a projection method to admit a small degree of compressibility provides a significant reduction in computational costs, whilst preserving smooth pressure fields obtained with incompressible SPH. Exchanges between bubbles and liquid are implemented by SPH interpolation, and the additional construction of neighbour lists required to enable this is straightforward in an SPH framework. Individual bubbles are able to be tracked over a lifetime, including motion due to turbulent structures below the resolution of the SPH scheme. Hence, models for bubble breakup (or in future, deformation and oscillation), which happen over a finite time, may be constructed in integral formulations, rather than as probabilistic events based on an instantaneous flow state. We have demonstrated the ability of our method to simulate bubble plumes and breaking waves, with quantitative agreement with previous numerical and experimental data in terms of mean flow statistics, bubble size distributions, and bubble population evolution. Despite the inclusion of bubble entrainment and breakup through simplified energy-balance models, the numerical scheme is capable of accurately predicting the Hinze scale, and the multi-slope bubble size distribution present in breaking waves, alongside the bubble population growth rate. The bubble distributions between breaking waves generated by our model compare qualitatively well with experimental data, including the generation of bubble clouds dragged downwards by obliquely descending eddies. Our investigations have highlighted a limitation of models of this type, which is that bubble entrainment models based on turbulence dissipation rates alone cannot account for the different physical mechanisms of entrainment in different flow types. Further developments which are planned include the extension of the SPH model to multi-phase flows, to yield a framework in which large-scale bubbles are resolved, with small-scale bubbles modelled as in the present scheme. As bubbles influence the forces and loads on structures due to wave impacts, and given the strengths of SPH in computationally affordable free-surface flow simulations, the work herein offers the potential for improvements in the accuracy of predictive modelling for wave-structure interactions. | {'timestamp': '2022-06-17T02:16:47', 'yymm': '2206', 'arxiv_id': '2206.01641', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01641'} |
null | null |
# Introduction
Hawking observed that a black hole (BH) radiates all kinds of particles and the radiated spectrum of particles are absolutely thermal and the spectrum of emission probability for different types of particles has been studied. The phenomenon of tunneling is established on the physical action of particles theory which gives BHs radiation. The law of energy conservation and momentum conservation has been adapted to this phenomenon. The particular temperature value at which vector particles are emitted from the BH horizon is called boson radiation. The Hawking radiation for scalar and fermion particles from different spacetimes have been investigated. The tunneling probability of charged scalars and fermions outgoing from the event horizon from these BHs have been studied. The temperature and entropy in the cosmological horizon of Schwarzschild de-Sitter spacetime have been investigated in. The charged particles radiated through tunneling from 4-dimensional as well as 5-dimensional BHs and Vaidya BH have been investigated, however, the Hawking temperatures and tunneling rates which depend on the properties of space-time have been observed. The noncommutative acoustic BHs of quantum-corrected finite entropy have been discussed and the corrections in terms of electric charge are conserved. The energy and mass of outgoing fermions and Kerr and Kerr-Newman BHs have been investigated and the tunneling particles would be stopped at some particular range and it is also concluded that the tunneling rate of radiated particles and Hawking temperature for both BHs which is depends on different parameters. The Hawking radiation spectrum under the influence of quantum gravity has been analyzed from different types of BH in. The energy spectrum and the Hawking temperature contribute holomorphic and anti-holomorphic functions, GUP and conformal gravity effects on tunneling radiation from various BHs have been observed, which are consistent with original results. The BHs tunneling radiation with the assumption of GUP, unitary theory , loop quantum theory and coordinate system have been examined and found that which depends on GUP, unitary theory, string and loop quantum theory but independent on coordinate system. The tunneling particles from the BH, BTZ like BH and black ring have been examined and it contains the hypothesis; BHs and black rings have both unstable and stable properties. Many authors computed the tunneling radiation of boson and fermion particles to get the Hawking temperature for different wormholes and BHs. The researcher computed the quantum gravity effect on tunneling radiation and it is a phenomenon for different types of BHs. The Newmananis algorithmic rule to analyze the Hayward BH solution with rotation parameter has been computed in.The temperature for regular Hayward BH metric with rotation parameter has been derived with the help of semi-classical phenomenon. An extension for temperature of 3 and 4 dimensional BH metric as well as higher dimensional BHs metric under the influences of quantum gravity have been analyzed in. The Rastall rotating BH with surrounding quintessence and dust parameter has been discussed in and showed that the Rastall theory metric solution is different from the standard metric solution. The charged bosons particles tunneling with spin equals \(1\) that is \(Z\) and \(W_{\pm}\) play a significant role for the standard model of electro-weak interaction so that the radiation of charged boson particles should be importance in the observation of tunneling radiation. Many different approaches have been analyzed for the investigation of the tunneling spectrum through the horizons of BHs. Visser analyzed that the Rastall gravity is utterly equivalent to standard Einstein general relativity which is obtained the artificially splitting the physical conserved stress-energy tensor into two non-conserved pieces. This paper aims to study the quantum gravity effects on BH thermodynamics. This paper is organized as follows. In the Sec. (**II**) deals the introduction of Rastall gravity Kerr-Newman BH surrounded by perfect fluid matter. We analyze the tunneling radiation of BH by assuming the Lagrangian equation in the scenario of GUP and calculated the temperature. The graphical representation of temperature in terms of rotation and quantum gravity are discussed in Sec. (**III**). The last section deals with our results.
# Kerr-Newman Black Hole in Rastall Gravity Theory
By applying Newman-Janis algorithmic rule to study the rotating Rastall BH space-time, which analyzes by mass (\(M\)), rotation parameter(\(a\)), surrounding fluid structure parameter (\(\sigma\)), state parameter of surrounding fluid \(u\) and Rastall coupling parameter \(\alpha\). The Rastall BH metric in the Boyer-Lindquist coordinates from can be expressed as \[ds^{2}=-Adt^{2}+Bdr^{2}+Cd\theta^{2} +Dd\phi^{2}+2Fdtd\phi,\label{J}\] where \(A\), \(B\), \(C\), \(D\) and \(E\) are given by the following equations: \[\begin{aligned}
A&=&1-\frac{2Mr+\sigma r^{\gamma}}{\Sigma},~~~ B=\frac{\Sigma}{\lambda_r},~~~C=\Sigma,\nonumber\\ D&=&sin^{2}\theta \left(r^{2}+a^{2}+\frac{(2Mr+\sigma r^{\gamma})a^{2}sin^{2}\theta}{\Sigma}\right),\nonumber\\ F&=&\frac{-a sin^{2}\theta\left(2Mr+\sigma r^{\gamma}\right)}{\Sigma}.\nonumber \end{aligned}\] Here, these \(\lambda_r\), \(\Sigma\) and \(\gamma\) parameters are defined as \[\begin{aligned}
\lambda_r&=&r^{2}-2Mr+a^{2}-\sigma r^{\gamma},\nonumber\\ \Sigma &=&a^{2}cos^{2}\theta+a^{2},\nonumber\\ \gamma&=&{\frac{1-3u}{1-3\alpha(1+u)}}\nonumber \end{aligned}\] It is important to mention that, if \(a=0\) and \(-1<u<\frac{-1}{3}\), the metric ([\[J\]](#J){reference-type="ref" reference="J"}) constitutes the Kerr BH surrounded by quintessence and if \(\sigma\) and \(a\) both are zero, we get Schwarzschild BH solution in Ref.. In order to study the tunneling radiation of charged particles via the BH horizon, we shall analyze electromagnetic effects of the Lagrangian gravity equation. The electromagnetic-charged fields, which describe the particle's motion in the Lagrangian charged equation \[\frac{1}{\sqrt{-g}}\partial_{\mu}(\sqrt{-g}\Psi^{\nu\mu})+\frac{i}{h}e A_{\mu}\Psi^{\nu\mu}+ \frac{m^{2}}{h^{2}}\Psi^{\nu}+ \hbar^{2}\beta\partial_{0}\partial_{0}\partial_{0}(g^{00}\Psi^{0\nu})-\hbar^{2}\beta\partial_{i}\partial_{i}\partial_{i}(g^{ii}\Psi^{i\nu})=0\label{L},\] where \(g\), \(\Psi^{\mu\nu}\) and \(m\) are the matrix determinant of coefficients, particle mass and anti-symmetric tensor, respectively. \[\Psi_{\nu\mu}=(1-\hbar^{2}\beta\partial^{2}_{\nu})\partial{\nu} \Psi_{\mu}-(1-\hbar^{2}\beta\partial^{2}_{\mu})\partial{\mu}\Psi_{\nu} +(1-\hbar^{2}\beta\partial^{2}_{\nu})\frac{i}{h}eA_{\nu}\Psi_{\mu}-(1-\hbar^{2}\beta\partial^{2}_{\mu})\frac{i}{h}eA_{\mu}\Psi_{\nu}.\nonumber\] Here, \(\nabla_{\mu}\), \(e\) and \(A_\mu\) are the covariant derivative, particle charged and BH potential, respectively. So the tunneling of positive and negative particles is similar but opposite sign \((W_{+}=-W_{-})\). For explanation, here we shall study the positive particles tunneling and this tunneling can be changed to negative particles tunneling. The values of \(\Psi^{\mu}\) and \(\Psi^{\nu\mu}\) are given by \[\begin{aligned}
\Psi^{\nu}&=&\Psi_{\mu}g^{\mu\nu},\nonumber\\ \Psi^{0}&=&\frac{D\Psi_{0}-F\Psi_{3}}{AD+F^{2}},~~~ \Psi^{1}=\frac{\Psi_{1}}{B},~~~ \Psi^{2}=\frac{\Psi_{2}}{C},~~~ \Psi^{3}=\frac{A \Psi_{3}-F \Psi_{0}}{AD+F^{2}},\nonumber\\ \Psi^{pq}&=&\Psi_{\mu\nu} g^{p\mu}g^{q\nu}\nonumber\\ \Psi^{01}&=&\frac{D\Psi_{01}-F\Psi_{31}}{B(AD+F^{2})},~~~ \Psi^{02}=\frac{D\Psi_{02}-F\Psi_{32}}{C(AD+F^{2})},~~~ \Psi^{03}=\frac{-\Psi_{03}}{AD+F^{2}},~~~ \Psi^{12}=\frac{\Psi_{12}}{BC},\nonumber\\ \Psi^{13}&=&\frac{A\Psi_{13}-F\Psi_{10}}{B(AD+F^{2})},~~~\Psi^{23}=\frac{A\Psi_{23}-F\Psi_{20}}{C(AD+F^{2})}.\nonumber \end{aligned}\] Applying the WKB approximation, \[\psi_{\nu}=k_{\nu}\exp[\frac{i}{\hbar}N_{0}(t,r,\theta,\phi)+ \Sigma \hbar^{n}N_{n}(t,r,\theta,\phi)],\] The term \(\hbar\) is assumed only for the lowest (\(1^{st}\)) order and the higher (higher order\(>1\)) dictate contributions are neglected in the Lagrangian gravity Eq. ([\[L\]](#L){reference-type="ref" reference="L"}), we get set of field equations. The variables separation technique in, we can take \[N_{0}=-E_{0}t+W(r)+V(\theta)+J\phi,\] where \(E_{0}=E-\omega\Omega\), \(\omega\) and \(E\) represent particle's angular momentum and energy respectively. The set of field equations, we can get a matrix field equation \[K(k_{0},k_{1},k_{2},k_{3})^{T}=0,\] which implies the \"\(K\)\" is label as \(4\times4\) matrix, whose elements are given as follows: \[\begin{aligned}
K_{00}&=&\frac{D}{B}\dot{W}^{2}W_{1}-\frac{D}{C}J^{2} J_{1}-\dot{V}^{2}V_{1}-eA_{3}[2\dot{V}+2\beta\dot{V}^{3}+eA_{3}+\beta eA_{3}\dot{V}^{2}]-m^{2}D,\nonumber\\ K_{01}&=&-\frac{D}{B}\dot{W}E_{0}[1+\beta E_{0}^{2}-eA_{0}E_{0}- \beta eA_{0}E_{0}]-\frac{F}{B}\dot{W}[\dot{V}+\beta\dot{V}^{3}+eA_{3}+\beta eA_{3}\dot{V}^{2}],\nonumber\\ K_{02}&=&-\frac{D}{C}J[E_{0}-\beta E_{0}^{^{3}} +eA_{0}+\beta eA_{0}E_{0}]-\frac{F}{C}J [\dot{V}+\beta\dot{V}^{3}+eA_{3}+\beta\dot{V}^{2}],\nonumber\\ K_{03}&=&\frac{F}{B}\dot{W}^{2}W_{1}+\frac{F}{C}J^{2}J_{1}-E_{0}E_{1}\dot{V} +eA_{0}\dot{V}E_{1}-eA_{3}E_{0} E_{1}+e^{2}A_{0}A_{3}E_{1}+m^{2}F,\nonumber\\ K_{10}&=&-D\dot{W}E_{0}W_{1}-F\dot{V}\dot{W}W_{1}+DeA_{0}\dot{W}W_{1}-FeA_{3}\dot{W}W_{1},\nonumber\\ K_{11}&=&-DE_{0}^{2}E_{1}-DeA_{0}E_{0}E_{1}-F\dot{V}E_{0}V_{1}-FeA_{3} E_{0}V_{1}-\frac{F_{0}}{C}J^{2}J_{1}-A\dot{V}^{2} V_{1}\nonumber\\&&-A_{0}eA_{3}\dot{V}V_{1}-F\dot{V}E_{0}E_{1} +FeA_{0}\dot{V}[2+\beta E_{0}^{2}+\beta\dot{V}^{2}] +DeA_{0}E_{0}E_{1}-De^{2}A^{2}_{0}E_{1} \nonumber\\&&+Fe^{2}A_{0}A_{3}V_{1}-AeA_{3}\dot{V}V_{1}-FeA_{3}E_{0}E_{1}+Fe^{2}A_{0}A_{3}E_{1}-m^{2}F_{0}\nonumber \end{aligned}\] \[\begin{aligned}
K_{12}&=&\frac{F_{0}}{C}[\dot{W}J+\beta \dot{W}^{3} J],\nonumber\\ K_{13}&=&F\dot{W}E_{0} W_{1}+A\dot{W}\dot{V}W_{1} +AeA_{3}[E_{0}^{2}+\beta\dot{W}^{3}],\nonumber\\ K_{20}&=&-DE_{0}J_{1}-FJ\dot{V}J_{1} +DeA_{0}JJ_{1}-FeA_{3}JJ_{1},\nonumber\\ K_{21}&=&\frac{F_{0}}{B}\dot{W}JJ_{1},\nonumber\\ K_{22}&=&-DE_{0}^{2}E_{1} +DeA_{0}E_{0}E_{1}-F\dot{V}E_{0}V_{1}-FeA_{3}E_{0}J_{1}-\frac{F_{0}}{B}[\dot{W}^{2}-\beta \dot{W}^{4}]\nonumber\\&&-\dot{V}^{2}A V_{1}-AeA_{3}\dot{V}V_{1}-F\dot{V}E_{0}E_{1}+ FeA_{0}\dot{V}[1+E_{1}+\beta\dot{V}^{2}]\nonumber\\&&+DeA_{0}E_{0}E_{1}-De^{2}A^{2}_{0}E_{1}+Fe^{2}A_{0}A_{3}V_{1}-AeA_{3}\dot{V}V_{1}-FeA_{3}E_{0}E_{1} \nonumber\\&&+Fe^{2}A_{0}A_{3}E_{1}-Ae^{2}A^{2}_{3} V_{1}-m^{2}F_{0},\nonumber\\ K_{23}&=&FJE_{0}J_{1}+AJ\dot{V}J_{1} +AeA_{3}JJ_{1}-FeA_{0}JJ_{1},\nonumber\\ K_{30}&=&-\dot{V}E_{0}V_{1}-eA_{3}E_{0}V_{1} +\frac{F}{B}\dot{W}^{2}W_{1} +\frac{F}{C}J^{2} J_{1}+eA_{0}\dot{V}V_{1} +e^{2}A_{0}A_{3}V_{1}+m^{2}F,\nonumber\\ K_{31}&=&\frac{A}{B}\dot{V}\dot{W}V_{1}+ \frac{A}{B}eA_{3}\dot{W}V_{1}+ \frac{F}{B}\dot{W}[E_{0}+\beta E_{0}^{3}-eA_{0}-eA_{0} \beta E_{0}^{2}],\nonumber\\ K_{32}&=&\frac{A}{C}J\dot{V}V_{1}+ \frac{A}{C}eA_{3}JV_{1}+\frac{F}{C}J[E_{0} +\beta E_{0}^{3}+eA_{0}+ eA_{0}\beta E_{0}^{2}],\nonumber\\ K_{33}&=&-E_{0}^{2}E_{1} +eA_{0}E_{0}E_{1}-\frac{A}{B}\dot{W}^{2}W_{1}-\frac{A}{C}J^{2}J_{1}+eA_{0} \nonumber\\&&[E_{0}+\beta E_{0}^{3}-eA_{0}-\beta eA_{0}]-m^{2}A,\nonumber \end{aligned}\] where \(\dot{W}=\frac{\partial W}{\partial r}\), \(\dot{V}=\frac{\partial V}{\partial\theta}\), \(E_{1}=1+\beta E_{0}^{2}\), \(J_{1}=1+\beta J^{2}\), \(W_{1}=1+\beta\dot{W}^{2}\), \(V_{1}=1+\beta\dot{V}^{2}\) and \(F_{0}=AD+F^{2}\) and the imaginary parts of \(imW_{-}\) and \(imW_{+}\) yield \[imW_{+}=-imW_{-}=\pm \int\sqrt{\frac{(E_{0}-eA_0)^{2} +X_1[1+\beta\frac{X_2}{X_1}]}{-AB^{-1}}}dr\label{R123}\] Here, \(W_{-}\) and \(W_{+}\) represent the solution of absorbing and radiating charged boson particles action respectively and \(X_1=m^{2}A+AC^{-1}J^{2}\) and \(X_2=E_{0}^{4}-eA_{0}E_{0}^{3}+\frac{A}{B}\dot{W}^{4}+\frac{A}{C}J^{4}-eA_{0}E_{0}^{3}+e^{2}A^{2}_{0}\). In the series of Taylor's, we are expanding the functions \(A(r)\) and \(B(r)\) near the horizon, we get \[A(r_{+})\approx \acute{A}(r)(r-r_{+}),~~B(r_{+})\approx \acute{B}(r)(r-r_{+})\label{w2}\] Since applying Eqs. ([\[R123\]](#R123){reference-type="ref" reference="R123"}) and ([\[w2\]](#w2){reference-type="ref" reference="w2"}) and integrate the around the pole, we get \[imW_{+}=+\pi\iota\frac{(E_{0}-eA_0)^{2}}{2\kappa(r_{+})}(1+\Xi\beta),\label{w1}\] where \(\Xi\) is arbitrary parameter and \(\kappa(r_+)\) is a surface gravity of BH. The surface gravity of BH defines as; \[\kappa(r_{+})=\frac{2M(M-r_{+})}{(a^{2}+a^{2}cos^{2}\theta)^{2}}\] Here, the tunneling probability \(\Gamma\) forbidden trajectories of the classically of the \(s\)-waves coming from inside to outside from the BH horizon. Applying the WKB approximation, by the terms of classical action \(S_{0}\) of charged boson particles tunneling across the BH horizon as trajectories up to leading order is, \[\begin{aligned}
\Gamma &=&\frac{\Gamma_{emission}}{\Gamma_{absorption}} =\frac{\exp(-2imW_{+}-2imV)}{\exp(-2imW_{-}-2imV)} =e^{-4imW_{+}}\nonumber\\&=&\exp\left[-2\pi \frac{(E-\omega\Omega-eA_{0})}{\kappa(r_{+})}(1+\Xi\beta)\right]. \end{aligned}\] Thus, for computing the Hawking temperature and we expand the action in terms of particles energy \(E\). Hence, the Hawking temperature is calculated at linear order given by \[T\cong \frac{M(r_{+}-M)}{\pi(a^{2}+a^{2}cos^{2}\theta)^{2}}[1-\Xi\beta].\] The charged boson massive particles case is similar as for massless case for Rastall gravity Kerr-Newman BH surround by perfect fluid matter from the above Hawking temperature. Moreover, the temperature for spin-up particles are similar \((-W_{+}=W_{-})\) as for spin-down case with the change \(r_{+}\) into \(r_{-}\). We observe that both spin-down and spin-up charged boson particles are radiate with like rate the temperature is assumed for this case.
# Graphical Temperature Analysis
In this section, we analyze graphical behavior of thermodynamical quantities in Rastall gravity Kerr-Newman BH surrounded by perfect fluid matter. For this purpose, we take a parameter \((\Xi=1)\) and observe the effects of the thermodynamical quantities (quantum gravity and rotation parameter) on BH radiation. The BH remains unstable if temperature increases or decreases. From Fig. 1, the Hawking temperature is decreasing due to the more quantum gravity. The stability of the BH depends on quantum gravity and also shows that BH remains stable unless the quantum gravity parameter is zero \((\beta=0)\). Also, temperature decreases in the approximation range \((0<r_{+}<5)\).\
If the quantum gravity is assumed to be non-zero, then Hawking temperature attains the similar value in figure 2. The Hawking temperature is decrease to rotation parameter in the values 0.1-0.3. The gravity parameter is fixed and varying the rotation parameter then we concluded that rotation parameter is inversely proportional to Hawking temperature and stability of BH. The stability of BH remains constant in very small range, i.e., approximation \(0<r_{+}<0.5\) and then temperature decreases with a wide variation. The rotation parameter is inversely proportional to the temperature of BH by keeping the quantum gravity constant in Fig. 2 for 2D. The temperature is behaving constant when \(r_{+}\) is very small, but at a certain value of rotation parameter, it attain a constantly decrease value in Fig. 2 for 2D. Hence our BH is stable when \(r_{+}\) is very small range otherwise is unstable. We concluded that BH radiation depends on BH mass, rotation parameter and gravity parameter but keeping constant \(cos^{2}\theta=1\).\
# Conclusions
In this paper, we have observed radiation spectrum by considering Hawking temperature for Rastall gravity Kerr-Newman BH is surrounded by perfect fluid matter. For this purpose, we have utilized the WKB approximation to the Hamilton-Jacobi ansatz for massive bosons charged spin-1 particles. The calculation yields, the Hawking temperature reliable with BH universality. In our investigation, we have changed the Lagrangian field equation in curved space time by effect of incorporate GUP quantum gravity. We have computed the tunneling probability at event horizon as well as the equating Hawking temperature. We have observed temperature in detail the effect of quantum gravity and rotation parameters in graphically. The paper have studied graphical investigation of the result and the BH will emit away all types of particles as the black-body radiation and the BH lose completely its information. In adjustment to break this problem, the after-effect rotation and gravity to be modified. We assume into account the law of conservation of momentum and energy. In our observation, the back reaction of the radiated boson particles of the self-gravitational interaction and BH geometry are reasonably ignored, the calculated Hawking temperatures depend on quantum gravity and the leading terms of metric. The modified form of temperature depends on BH mass, rotation parameters and quantum gravity parameter. When the gravity parameter \((\beta=0)\) influences are ignored, we have obtained the temperature of Rastall gravity Kerr-Newman BH. So, the quantum gravity effect minimum then stability of BH are increased. The quantum gravity and rotation are present in the BH, which shows that the instability of BH. The BH is completely stable when rotation and quantum gravity do not appear.\
**Acknowledgments**\
The work of KB was supported in part by the JSPS KAKENHI Grant Number JP21K03547.
# Appendix {#appendix .unnumbered}
In the WKB approximation the term \(\hbar\) is considered just for the \(1^{st}\) order and the higher dictate contributions are neglected in the Lagrangian Eq. ([\[L\]](#L){reference-type="ref" reference="L"}), we get following field equations; \[\begin{aligned}
&&\frac{D}{B}[k_{1}(\partial_{0}N_{0})(\partial_{1}N_{0}) +\beta k_{1}(\partial_{0}N_{0})^{3}(\partial_{1}N_{0}) +k_{0}(\partial_{1}N_{0})^{2}-\beta k_{0}(\partial_{1}N_{0})^{4}+eA_{0}k_{1}(\partial_{0}N_{0})^{2}(\partial_{1}N_{0})+ \beta eA_{0}k_{1}(\partial_{0}N_{0})^{2}(\partial_{1}N_{0})]\nonumber\\ &&-\frac{F}{B} [k_{1}(\partial_{1}N_{0})(\partial_{3}N_{0}) +\beta k_{1}(\partial_{1}N_{0})(\partial_{3}N_{0})^{3}-k_{3}(\partial_{1}N_{0})^{2}-\beta k_{3}(\partial_{1}N_{0})^{4} +eA_{3}k_{1}(\partial_{1}N_{0}) +\beta eA_{3}k_{1}(\partial_{1}N_{0})(\partial_{3}N_{0})^{2}]\nonumber\\ &&+\frac{D}{C}[k_{2}(\partial_{0}N_{0})(\partial_{2}N_{0}) +\beta k_{2}(\partial_{0}N_{0})^{3}(\partial_{2}N_{0})-k_{o}(\partial_{2}N_{0})^{2}-\beta k_{0}(\partial_{2}N_{0})^{4} +eA_{0}k_{2}(\partial_{2}N_{0})+\beta eA_{0} k_{2}(\partial_{0}N_{0})^{2}(\partial_{2}N_{0})]\nonumber\\ &&-\frac{F}{C}[k_{2}(\partial_{2}N_{0})(\partial_{3}N_{0})+\beta k_{2} (\partial_{2}N_{0})(\partial_{3}N_{0})^{3}-k_{3} (\partial_{2}N_{0})^{2}-\beta k_{3}(\partial_{2}N_{0})^{4} +eA_{3}k_{2}(\partial_{2}N_{0})+\beta eA_{3}k_{2} (\partial_{2}N_{0})(\partial_{3}N_{0})^{2}] \nonumber\\ &&+[k_{3}(\partial_{0}N_{0})+\beta k_{3}(\partial_{0}N_{0})^{3} (\partial_{3}N_{0})-k_{0}(\partial_{3}N_{0})^{2}-\beta k_{0} (\partial_{3}N_{0})^{4}+eA_{0}k_{3}(\partial_{3}N_{0}) +\beta eA_{0}k_{3}(\partial_{0}N_{0})^{2}(\partial_{3}N_{0})-eA_{3}k_{0}(\partial_{3}N_{0})\nonumber\\ &&-\beta eA_{3}k_{0} (\partial_{3}N_{0})^{3}]+eA_{3}[k_{3} (\partial_{0}N_{0})+\beta k_{3}(\partial_{0}N_{0})^{3}-k_{0}(\partial_{3}N_{0})-\beta k_{0}(\partial_{3}N_{0})^{3} +eA_{0}k_{3}+\beta eA_{0}k_{3}(\partial_{0}N_{0})^{2}-eA_{3}k_{0}\nonumber\\ &&-\beta eA_{3}k_{0}(\partial_{3}N_{0})^{2}]=0,\\ &&D[k_{0}(\partial_{0}N_{0})(\partial_{1}N_{0})+ \beta k_{0}(\partial_{0}N_{0})(\partial_{1}N_{0})^{3}-k_{1}(\partial_{0}N_{0})^{2}-\beta k_{1}(\partial_{0}N_{0})^{4}-eA_{0}k_{1}(\partial_{0}N_{0})-\beta eA_{0}k_{1}(\partial_{0}N_{0})^{3}]-F [k_{3}(\partial_{0}N_{0})\nonumber\\ &&(\partial_{1}N_{0})+\beta k_{3}(\partial_{0}N_{0})(\partial_{1}N_{0})^{3}-k_{1}(\partial_{0}N_{0})(\partial_{3}N_{0})-\beta k_{1} (\partial_{0}N_{0})(\partial_{3}N_{0})^{3}-eA_{3}k_{1}(\partial_{0}N_{0})-\beta eA_{3}k_{1}(\partial_{0}N_{0})(\partial_{3}N_{0})^{2}] \nonumber\\ &&+(AD-F^{2})C^{-1}[k_{2}(\partial_{1}N_{0})(\partial_{2}N_{0}) +\beta k_{2}(\partial_{1}N_{0})^{3}(\partial_{2}N_{0})-k_{1}(\partial_{2}N_{0})^{2}-\beta k_{1}(\partial_{2}N_{0})^{4}] +A[k_{3}(\partial_{1}N_{0})(\partial_{3}N_{0})+\beta k_{3}\nonumber\\ &&(\partial_{1}N_{0})^{3} (\partial_{3}N_{0})-k_{1}(\partial_{3}N_{0})^{2}-\beta k_{1}(\partial_{3}N_{0})^{4}-eA_{3}k_{1}(\partial_{3}N_{0})-\beta eA_{3}k_{1}(\partial_{3}N_{0})^{3}]-F[k_{0}(\partial_{1}N_{0})(\partial_{3}N_{0})+\beta k_{0} (\partial_{1}N_{0})^{3}\nonumber\\ &&(\partial_{3}N_{0})-k_{1}(\partial_{0}N_{0}) (\partial_{3}N_{0})-\beta k_{1}(\partial_{0}N_{0})^{3}(\partial_{3}N_{0})-eA_{0}k_{1}(\partial_{3}N_{0})-\beta eA_{0}k_{1}(\partial_{0}N_{0})^{2}(\partial_{3}N_{0})] +D eA_{0}[k_{0}(\partial_{1}N_{0})+\beta k_{0}\nonumber\\ &&(\partial_{1}N_{0})^{3}-k_{1}(\partial_{0}N_{0})-\beta k_{1}(\partial_{0}N_{0})^{3}-eA_{0}k_{1}-\beta eA_{0}k_{1}(\partial_{0}N_{0})^{2}]-F eA_{0}[k_{3}(\partial_{1}N_{0})+\beta k_{3}(\partial_{1}N_{0})^{3}-k_{1}(\partial_{3}N_{0})-\beta k_{1}\nonumber\\ &&(\partial_{3}N_{0})^{3}-eA_{3}k_{1}-\beta eA_{3}k_{1}(\partial_{3}N_{0})^{2}]+eAA_{3}[k_{3}(\partial_{0}N_{0})^{2}+\beta k_{3} (\partial_{1}N_{0})^{3}-k_{1}(\partial_{3}N_{0})-\beta k_{1}(\partial_{3}N_{0})^{3}-eA_{3}k_{1}-\beta eA_{3}\nonumber\\ &&k_{1}(\partial_{3}N_{0})^{2}]- eFA_{3}[k_{0}(\partial_{1}N_{0})+\beta k_{0}(\partial_{1}N_{0})^{3}- k_{1}(\partial_{0}N_{0})-\beta k_{1}(\partial_{0}N_{0})^{3}-eA_{0}k_{1}-\beta eA_{0}k_{1} (\partial_{0}N_{0})^{2}]-m^{2}k_{1}\nonumber\\ &&(AD-F^{2})=0\\ &&D[k_{0}(\partial_{0}N_{0})(\partial_{2}N_{0})+\beta k_{0} (\partial_{0}N_{0})(\partial_{2}N_{0})^{3}-k_{2}(\partial_{0}N_{0})^{2}-\beta k_{2}(\partial_{0}N_{0})^{4}-eA_{0}k_{2}(\partial_{0}N_{0})-\beta eA_{0}k_{2}(\partial_{0}N_{0})^{3}]-F [k_{3}(\partial_{0}N_{0})\nonumber\\ &&(\partial_{2}N_{0})+\beta k_{3}(\partial_{0}N_{0})(\partial_{2}N_{0})^{3}-k_{2}(\partial_{0}N_{0}) (\partial_{3}N_{0})-\beta k_{2}(\partial_{0}N_{0})(\partial_{3}N_{0})^{3}-eA_{3}k_{2}(\partial_{0}N_{0})-\beta eA_{3}k_{2}(\partial_{0}N_{0})(\partial_{3}N_{0})^{2}]\nonumber\\ &&+(AD-F^{2})B^{-1}[k_{1}(\partial_{1}N_{0})(\partial_{2}N_{0}) +\beta k_{1}(\partial_{1}N_{0})(\partial_{2}N_{0})^{3}-k_{2}(\partial_{1}N_{0})^{2}-\beta k_{2}(\partial_{1}N_{0})^{4}] +A[k_{3}(\partial_{2}N_{0})(\partial_{3}N_{0})+\beta k_{3}\nonumber\\ &&(\partial_{2}N_{0})^{3} (\partial_{3}N_{0})-k_{2}(\partial_{3}N_{0})^{2}-\beta k_{3}(\partial_{3}N_{0})^{4}-eA_{3}k_{2} (\partial_{3}N_{0})-\beta eA_{3}k_{2}(\partial_{3}N_{0})^{3}]-F[k_{0}(\partial_{2}N_{0})(\partial_{3}N_{0})+\beta k_{0} (\partial_{2}N_{0})^{3}\nonumber\\ &&(\partial_{3}N_{0})-eA_{0}k_{2}(\partial_{3}N_{0})-\beta eA_{0}k_{2}(\partial_{0}N_{0})^{2}(\partial_{3}N_{0})-k_{2}(\partial_{0}N_{0})(\partial_{3}N_{0})-\beta k_{2} (\partial_{0}N_{0})^{3}(\partial_{3}N_{0})] +DeA_{0}[k_{0}(\partial_{2}N_{0})+\beta k_{0}\nonumber\\ &&(\partial_{2}N_{0})^{3}-k_{2}(\partial_{0}N_{0})-\beta k_{2}(\partial_{0}N_{0})^{3}-eA_{0}k_{2}-\beta eA_{0}k_{2}(\partial_{0}N_{0})^{2}]-FeA_{0}[k_{3}(\partial_{2}N_{0}) +\beta k_{3}(\partial_{2}N_{0})^{3}-k_{2}(\partial_{3}N_{0})-\beta k_{2}\nonumber\\ &&(\partial_{3}N_{0})^{3}-eA_{3}k_{2}-\beta eA_{3}k_{2}(\partial_{3}N_{0})^{2}] +eAA_{3}[k_{3}(\partial_{2}N_{0})+\beta k_{3} (\partial_{2}N_{0})^{3}-k_{2}(\partial_{3}N_{0})-\beta k_{2}(\partial_{3}N_{0})^{3}\nonumber\\ &&-eA_{3}k_{2}-\beta eA_{3}k_{2}(\partial_{3}N_{0})^{2}]-eFA_{3} [k_{0}(\partial_{2}N_{0})+\beta k_{0}(\partial_{2}N_{0})^{3}- k_{2}(\partial_{0}N_{0})-\beta k_{2}(\partial_{0}N_{0})^{3}-eA_{0}k_{2}-\beta eA_{0}k_{2} (\partial_{0}N_{0})^{2}]\nonumber\\ &&-m^{2}k_{2}(AD-F^{2})=0\\ &&[k_{0}(\partial_{0}N_{0})(\partial_{3}N_{0})+\beta k_{0} (\partial_{0}N_{0})(\partial_{3}N_{0})^{3} +k_{3}(\partial_{0}N_{0})^{2}-\beta k_{3}(\partial_{0}N_{0})^{4} +eA_{3}k_{0}(\partial_{0}N_{0})+\beta eA_{3}k_{0} (\partial_{0}N_{0})(\partial_{3}N_{0})^{2}\nonumber\\ &&-eA_{0}k_{3}(\partial_{0}N_{0})-\beta eA_{0}k_{3} (\partial_{0}N_{0})^{3}] +\frac{A}{B} [k_{1}(\partial_{1}N_{0})(\partial_{3}N_{0})+\beta k_{1} (\partial_{1}N_{0})(\partial_{3}N_{0})^{3}-k_{3}(\partial_{1}N_{0})^{2}-\beta k_{3}(\partial_{1}N_{0})^{4} +\nonumber\\ &&eA_{3}k_{1}(\partial_{1}N_{0})+\beta eA_{3}k_{1}(\partial_{1}N_{0})(\partial_{3}N_{0})^{2}]-\frac{F}{B}[k_{1}(\partial_{0}N_{0})(\partial_{1}N_{0})-\beta k_{1}(\partial_{0}N_{0})^{3}(\partial_{1}N_{0})-k_{o}(\partial_{1}N_{0})^{2}-\beta k_{0}(\partial_{1}N_{0})^{4} \nonumber\\ &&+eA_{0}k_{1}(\partial_{1}N_{0})+\beta eA_{0} k_{1}(\partial_{0}N_{0})^{2}(\partial_{1}N_{0})] +\frac{A}{C}[k_{2}(\partial_{2}N_{0})(\partial_{3}N_{0})+\beta k_{2} (\partial_{2}N_{0})(\partial_{3}N_{0})^{3}-k_{3}(\partial_{2}N_{0})^{2}- \beta k_{3}(\partial_{2}N_{0})^{4}\nonumber\\ &&+eA_{3}k_{2}(\partial_{2}N_{0})+\beta eA_{3}k_{2} (\partial_{2}N_{0})(\partial_{3}N_{0})^{2}]-\frac{F}{C}[k_{2}(\partial_{0}N_{0})(\partial_{2}N_{0})+ \beta k_{2}(\partial_{0}N_{0})^{2}(\partial_{2}N_{0})-k_{0}(\partial_{2}N_{0})^{2}-\beta k_{0}(\partial_{2}N_{0})^{4} \nonumber\\ &&+eA_{0}k_{2}(\partial_{2}N_{0}) +\beta eA_{0}k_{2}(\partial_{0}N_{0})^{2}(\partial_{2}N_{0})] +eA_{0}[k_{0} (\partial_{3}N_{0})+\beta k_{0}(\partial_{3}N_{0})^{3}-k_{3}(\partial_{0}N_{0})-\beta k_{3}(\partial_{0}N_{0})^{3} +eA_{3}k_{0}+\nonumber\\ &&\beta eA_{3}k_{0}(\partial_{3}N_{0})^{2}-eA_{0}k_{3}-\beta eA_{0}k_{3}(\partial_{0}N_{0})^{2}]+m^{2} [Fk_{0}-Ak_{3}]=0. \end{aligned}\] | {'timestamp': '2022-06-06T02:13:25', 'yymm': '2206', 'arxiv_id': '2206.01546', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01546'} |
null | null |
# Introduction {#sec: intro}
Mean-field games (MFG) study the behavior of a large number of rational agents in an non-cooperative game. It has wide applications in various fields, such as economics, engineering as well as machine learning and reinforcement learning. Recently, mean field control problems have been extended into chemistry, biology, pandemic control, traffic flow models and social dynamics. An important task in mean-field games is to study the flow of all the agents in the state space and to understand behavior of mean-field Nash equilibrium. Conventional studies of MFG focus on choice of the state space as a Euclidean flat domain, for instance, \([0,1]^d\) with periodic boundary conditions. Besides research on Euclidean flat domains, there are existing work focusing MFGs on graphs or graphon state spaces . However, such spaces may not be adequate to reflect the metric structure of state spaces in many applications. For instance, the problems of population flows or resource distributions on the Earth are actually defined on a sphere. In machine learning, the manifold hypothesis is commonly used , since many real world data sets are actually samples from low-dimensional manifolds in a high dimensional ambient space. Therefore, it is quite natural and necessary to explore mean-field game/control problems on manifolds. In this work, we would like to generalize the concepts of finite horizon mean-field games and mean-field Nash Equilibrium from Euclidean spaces to manifolds, and propose a numerical method to compute the Nash Equilibrium. In this study, we consider a game with infinitely many indistinguishable agents on a compact and smooth manifold \(\mathcal{M}\) within the time interval \([0,1]\). At any time \(t\in[0,1]\), each agent is in a certain state \(\mathbf{x}\in\mathcal{M}\) and the state of all agents forms a distribution \(\rho(\cdot,t)\in\mathcal{P}(\mathcal{M})\). For each agent at \(t\), given its current state \(\mathbf{x}\) and the anticipation of future state distribution \(\rho(\cdot,s),s\in[t,1]\), the game is to optimize a control \(\mathbf{v}(\mathbf{x}(s),s)\) to guide its future trajectory \(\mathbf{x}(s),s\in(t,1]\) in order to minimize a cost \(J^{\rho}(\mathbf{x},t,\mathbf{v})\). Therefore the optimal control \(\mathbf{v}\) depends on the state distribution \(\rho\). Although the state change of any single agent does not change \(\rho(\cdot,t)\), when all the agents take the same control, the state distribution \(\rho\) changes accordingly. Thus the optimal control \(\mathbf{v}\) and the state distribution \(\rho\) are interdependent, and the Nash Equilibrium, the special pair of \((\mathbf{v},\rho)\), is an especially interesting topic in mean-field game. In the conventional Euclidean setup, it has been shown that the mean-field Nash Equilibrium is the solution of a forward-backward PDE system . We generalize this result to MFG on manifolds. Meanwhile, for a potential mean-field game on a Euclidean domain , its optimality condition is exactly the forward-backward PDE system under mild conditions. Thus, the Nash Equilibrium can be obtained by searching for the stationary point of the optimization problem. In this work, we show that the equivalence between the PDE formulation and variational formulation of mean-field game still holds on manifolds. It is worth mentioning that studies dynamic optimal transport, a special form of potential mean-field games, on manifolds. In this work, we consider more general forms of mean-field games on manifolds and we are interested in both the PDE and variational formulations. There are different approaches to numerically solve mean-field games on Euclidean domains, such as finite difference methods, monotone flows, optimization algorithms and neural networks. We refer readers to the surveys for more details of the numerical methods on Euclidean domains. In our manifold setting, we focus on the variational formulation to numerically compute the Nash Equilibrium. With the help of triangular mesh and computational geometry strategies, we approximate the manifold, probability space and vector field space and formulate the discrete optimization problem. Once the discretization is provided, most of existing optimization-based algorithms can be adapted to solve the proposed discretization problem. In this work, we specifically use an optimization-based algorithm proposed in since it is flexible and efficient. This algorithm is adapted from the proximal gradient descent method considered in. **Contributions:** We summarize our contributions as follows:
1. We generalize the concept of mean-field games to manifolds and derive the corresponding geometric PDE formulation of the Nash Equilibrium.
2. We show the equivalence of the PDE formulation and variational formulation of mean-field games on manifolds.
3. We propose a numerical method for solving the variational problem based on a proximal gradient descent method. Comprehensive experiments demonstrate the effectiveness of the proposed method.
**Organization:** Our paper is organized as follows. In section [2](#sec: mfg mf){reference-type="ref" reference="sec: mfg mf"}, we derive the PDE formulation of mean-field Nash Equilibrium on manifolds. We also show that the PDE system is actually the optimality condition of an optimization problem, the potential mean-field game, on the manifold. We discretize the potential MFG problems in space and time domain in section [3](#sec: disct){reference-type="ref" reference="sec: disct"}, and adapt a proximal gradient method to solve the discrete counterparts in section [4](#sec: alg){reference-type="ref" reference="sec: alg"}. In section [5](#sec: num res){reference-type="ref" reference="sec: num res"}, we provide numerical experiments that solve potential mean-field games with local or non-local interaction costs on different manifolds.
# Mean-field games on manifolds {#sec: mfg mf}
In this section, we generalize the concepts of finite horizon mean-field games (MFGs) and their variational forms from conventional Euclidean domains to smooth and compact Riemannian manifolds.
## Mean-field games on manifold {#subsec: mfg mf}
Let's begin with some notations for convenience. We consider MFG on \((\mathcal{M},g)\), a \(d\)-dimensional smooth compact manifold with a Riemannian metric \(g\). As a natural extension of MFG on Euclidean domains, controls at the state \(\mathbf{x}\in\mathcal{M}\) are defined as elements in \(\mathrm{T}_{\vx}\mfd\), the tangent space of \(\mathcal{M}\) at \(\mathbf{x}\in\mathcal{M}\). We further denote \(\mathrm{T}\mfd =\{(\mathbf{x},\mathbf{p})~|~ \mathbf{p}\in\mathrm{T}_{\vx}\mfd \}\) for the tangent bundle of \(\mathcal{M}\); use \(\Gamma(\mathrm{T}\mfd)\) for the set of continuous vector fields on \(\mathcal{M}\); and write \(\mathcal{P}(\mathcal{M})\) for all probability density on \((\mathcal{M},g)\) under the volume measurement induced by the metric \(g\). To derive a first-order MFG system on \(\mathcal{M}\), we consider a finite horizon game on the time interval \([0,1]\) with the state space \(\mathcal{M}\). More specifically, we assume that there is a continuum number of agents, and each agent takes a state \(\mathbf{x}\in\mathcal{M}\) at any time \(t\in[0,1]\). We write the state density of all the agents along \(t\in[0,1]\) as \(\rho\in C([0,1];\mathcal{P}(\mathcal{M}))\); and assume that the impact of any single agent to \(\rho\) is negligible. Since all the agents have the same goal in a mean-field game, it is sufficient to take a representative agent as an example. Suppose that an agent is in state \(\mathbf{x}\) at time \(t\), the agent aims at choosing a control \(\mathbf{v}\in C((t,1];\Gamma(\mathrm{T}\mfd))\) to guide the trajectory \[\mathrm{d}\mathbf{x}(t) = \mathbf{v}(\mathbf{x},t)\mathrm{d} t. \label{eq: agent dynamic}\] in order to minimize the cost \[J^{\rho}(\mathbf{x},t,\mathbf{v}):=\int_t^1 \left[ L(\mathbf{x}(s),\mathbf{v}(\mathbf{x}(s),s) + F(\mathbf{x}(s),\rho(\cdot,s)) \right] \mathrm{d} s +F_{T}(\mathbf{x}(1),\rho(\cdot,1)). \label{eq: J func}\] Here \(L:\mathrm{T}\mfd\to[0,+\infty)\) is the dynamic cost, \(F:\mathcal{M}\times \mathcal{P}(\mathcal{M})\to[0,+\infty)\) is the interaction cost, \(\rho(\cdot,s)\in \mathcal{P}(\mathcal{M})\) is the density of all agents at time \(s\), and \(F_{T}:\mathcal{M}\times\mathcal{P}(\mathcal{M})\to[0,+\infty)\) is the terminal cost. Note that the control \(\mathbf{v}\) and the state distribution \(\rho\) are involved interactively. The optimal control \(\mathbf{v}^{\rho}:=\argmin_{\mathbf{v}}J^{\rho}(\mathbf{x},t,\mathbf{v})\) generally depends on the evolution of the state distribution \(\rho\). Meanwhile, with given initial distribution \(\rho(\cdot,0):=\rho_0\in\mathcal{P}(\mathcal{M})\), the distribution of agents \(\rho\) is determined by the control \(\mathbf{v}\in C([0,1];\Gamma(\mathrm{T}\mfd))\) through equation [\[eq: agent dynamic\]](#eq: agent dynamic){reference-type="eqref" reference="eq: agent dynamic"}. The mean-field game problem is especially interested in a special pair of them, the Nash Equilibrium, which is the same as the conventional Euclidean case,
With the definition, if \((\mathbf{v},\rho)\) is a Nash Equilibrium of a MFG on \((\mathcal{M}, g)\), then the optimality condition ensures that \(\mathbf{v}\) is the optimal control for given state distribution \(\rho\), and the consistency requires that \(\mathbf{v}\) lead to the state distribution \(\rho\). According to, in Euclidean space, a Nash Equilibrium can be described by a PDE system, which includes a backward Hamiltonian-Jacobi-Bellman (HJB) equation by the optimality condition and a forward continuity equation by the consistency condition. In the rest, we will establish a similar PDE description of a Nash Equilibirum on manifolds. Similar as the Euclidean case, let the value function \(\phi\) be the cost with the optimal control, \[\phi^\rho(\mathbf{x},t) := \inf_{\mathbf{v}\in C([0,1];\Gamma(\mathrm{T}\mfd))}J^\rho (\mathbf{x},t,\mathbf{v}). \label{eq: value func def}\] and \(H\) be the manifold Hamiltonian defined on the tangent bundle of \(\mathcal{M}\) \[H:\mathrm{T}\mfd\to\mathbb{R},\quad H(\mathbf{x},\mathbf{q}):=\sup_{\mathbf{p}\in\mathrm{T}_{\vx}\mfd} \left\{-\langle\mathbf{q},\mathbf{p}\rangle_{g(\mathbf{x})}-L(\mathbf{x},\mathbf{p}) \right\}. \label{eq: control func hamiltonian def}\] we have the following theorem.
Before proving the theorem, we give several remarks to explain notations.
Next, we prove theorem [\[thm:NE_PDE\]](#thm:NE_PDE){reference-type="ref" reference="thm:NE_PDE"}.
At the end of this part, we present some common examples.
## Potential mean-field games on manifold {#subsec: vmfg mf}
According to, when the state space is Euclidean, with mild conditions, the local minimizer of an optimization problem and the corresponding dual variable is a weak solution to the MFG PDE system. In this part, we establish the parallel results on manifolds. We formulate the potential MFG on manifold and show that the necessary optimality condition of this variational problem is exactly the PDE system [\[eq: mfmfg pde\]](#eq: mfmfg pde){reference-type="eqref" reference="eq: mfmfg pde"} under certain conditions.
With the above theorem, the forward-backward system [\[eq: mfmfg pde\]](#eq: mfmfg pde){reference-type="eqref" reference="eq: mfmfg pde"} can be solved by searching for the local minimizer of variational problem [\[eq: mfmfg opt\]](#eq: mfmfg opt){reference-type="eqref" reference="eq: mfmfg opt"}. In this study, we majorly focus on the variational problem [\[eq: mfmfg opt\]](#eq: mfmfg opt){reference-type="eqref" reference="eq: mfmfg opt"}. In the rest part of this section, we present some examples of potential mean-field games as well as their corresponding PDE systems.
# Discretization on Manifolds {#sec: disct}
The optimization problem [\[eq: mfmfg opt\]](#eq: mfmfg opt){reference-type="eqref" reference="eq: mfmfg opt"} is defined in an infinite-dimension space and in general is solved approximately using appropriate discretization. Different from conventional MFG problems in Euclidean space, we need to approximate the ground manifold as well as functions and vector fields on the manifold. In this section, we focus on two-dimensional manifolds and discuss the discrete counterpart of [\[eq: mfmfg opt\]](#eq: mfmfg opt){reference-type="eqref" reference="eq: mfmfg opt"}. We first approximate the manifold with a triangular mesh. This leads a semi-dicrete version of [\[eq: mfmfg opt\]](#eq: mfmfg opt){reference-type="eqref" reference="eq: mfmfg opt"} and its associated KKT system. After that, we derive a fully discretized version for our numerical implementation by equally splitting the time interval.
## Space discretization {#subsec: vmfg trg}
We follow a conventional approach to approximate a two-dimensional manifold \(\mathcal{M}\subset\mathbb{R}^3\) by a triangular mesh \(\widetilde{\mfd}:=\{V,T\}\) . Here \(V=\{{\vtc_i}\in\mathbb{R}^3\}_{i=1}^{h}\) represents the set of vertices, \(T=\{{\trg_j}\}_{j=1}^{s}\) is the set of triangles where each triangle has three vertices in \(V\). Figure [\[fig: trimesh illu\]](#fig: trimesh illu){reference-type="ref" reference="fig: trimesh illu"} shows several triangular meshes used in our numerical experiments later. For convenience, we also abuse our notation \(\widetilde{\mfd}\) for the piecewise linear approximation of \(\mathcal{M}\) obtained from the the given triangular mesh. For a real-valued function \(\psi:\mathcal{M}\to\mathbb{R}\), we approximate it with a piece-wise linear function \(\Psi:\widetilde{\mfd}\to\mathbb{R}\), where on vertices \(\Psi({\vtc_i}):=\psi({\vtc_i})\) and on each triangle \(\Psi(\mathbf{x})\) is linear. In this way, any piece-wise linear function on \(\widetilde{\mfd}\) is fully represented by its values on vertices. With a slight abuse of notations, we denote \(\Psi:=\left(\begin{matrix} \Psi(V_1)\\ \vdots \\ \Psi(V_{h}) \end{matrix}\right)\) as a vector in \(\mathbb{R}^{h}\). By piece-wise linearity, the gradient of \(\Psi\) is a piece-wise constant vector field. For consistency, we let \(\mathrm{T}\disctmfd:=\sqcup_{j=1}^{s}\spn{{\trg_j}}\) mimic the tangent bundle, \(\Gamma(\mathrm{T}\disctmfd):=\left\{U:T\to\mathrm{T}\disctmfd,U({\trg_j}) = \left(\begin{matrix}U^1({\trg_j})\\U^2({\trg_j})\\U^3({\trg_j})\end{matrix}\right)\in\spn{{\trg_j}}\right\}\) denote the set of piece-wise constant vector field. Similar to the function discretization, we use the matrix \(U=\left(\begin{matrix}U^1,U^2,U^3\end{matrix}\right)=\left(\begin{matrix} (U(T_1))^\top\\ \vdots \\ (U(V_{s}))^\top \end{matrix}\right)\in\mathbb{R}^{s\times3}\) to fully describe the vector field. Given any \(\Psi:\widetilde{\mfd}\to\mathbb{R}\), its gradient \(\nabla_{\disctmfd}\Psi\in\Gamma(\mathrm{T}\disctmfd)\) can be written as \(\nabla_{\disctmfd}\Psi=\left(\begin{matrix}G^1\Psi&G^2\Psi&G^3\Psi\end{matrix}\right)\), where \(G^1,G^2,G^3\in\mathbb{R}^{s\timesh}\) provides a discretization of \(\nabla_\mathcal{M}\). To see this, take a triangle \(T_1\) with vertices \(V_1,V_2,V_3\) as an example (see the left image in Figure [\[fig: disct illu\]](#fig: disct illu){reference-type="ref" reference="fig: disct illu"}). We first parameterise \(T_1\) by \[V(\] Thus the induced metric on \(T_1\) is the constant matrix \[g= \left(\begin{matrix} \langle V_2-V_1,V_2-V_1\rangle & \langle V_2-V_1,V_3-V_1\rangle\\ \langle V_3-V_1,V_2-V_1\rangle & \langle V_3-V_1,V_3-V_1\rangle \end{matrix}\right)\] Because \(\Psi\) restricted on \(T_1\) is linear, we have \[\Psi(V(\] The definition of gradient gives us \[\begin{aligned} \left(\begin{matrix} \langle V_2-V_1,(\nabla_{\disctmfd}\Psi)(T_1) \rangle\\ \langle V_3-V_1,(\nabla_{\disctmfd}\Psi)(T_1) \rangle \end{matrix}\right) &= \left(\begin{matrix} \Psi(V_2)-\Psi(V_1)\\ \Psi(V_3)-\Psi(V_1) \end{matrix}\right) = \left(\begin{matrix}-1 & 1 & 0\\-1 & 0 & 1 \end{matrix}\right) \left(\begin{matrix} \Psi(V_1)\\\Psi(V_2)\\\Psi(V_3)\end{matrix}\right). \end{aligned} \label{eq: gd op 1}\] Since \((\nabla_{\disctmfd}\Psi)(T_1)\in\spn{T_1}\), it is clear that the gradient of \(\Psi\) on \(T_1\) has the decomposition \((\nabla_{\disctmfd}\Psi)(T_1) = \left(\begin{matrix} V_2-V_1,&V_3-V_1 \end{matrix}\right) \left(\begin{matrix} \mu^1\\ \mu^2 \end{matrix}\right)\) and therefore we have \[\left(\begin{matrix} \langle V_2-V_1,(\nabla_{\disctmfd}\Psi)(T_1) \rangle\\ \langle V_3-V_1,(\nabla_{\disctmfd}\Psi)(T_1) \rangle \end{matrix}\right) = g\left(\begin{matrix} \mu^1\\ \mu^2 \end{matrix}\right) \label{eq: gd op 2}\] Because the triangle is non-degenerative, [\[eq: gd op 1\]](#eq: gd op 1){reference-type="eqref" reference="eq: gd op 1"} and [\[eq: gd op 2\]](#eq: gd op 2){reference-type="eqref" reference="eq: gd op 2"} together solves \(\mu^1,\mu^2\) \[\begin{aligned} \left(\begin{matrix} \mu^1\\ \mu^2 \end{matrix}\right) = g^{-1} \left(\begin{matrix}-1 & 1 & 0\\-1 & 0 & 1 \end{matrix}\right) \left(\begin{matrix} \Psi(V_1)\\\Psi(V_2)\\\Psi(V_3)\end{matrix}\right). \end{aligned}\] And this implies \[\begin{aligned} (\nabla_{\disctmfd}\Psi)(T_1) = \left(\begin{matrix} V_2-V_1,&V_3-V_1 \end{matrix}\right) g^{-1} \left(\begin{matrix}-1 & 1 & 0\\-1 & 0 & 1 \end{matrix}\right) \left(\begin{matrix} \Psi(V_1)\\\Psi(V_2)\\\Psi(V_3)\end{matrix}\right). \end{aligned}\] Assigning \(G^d(T_1,{\vtc_i}) = 0\) for \({\vtc_i}\in V\backslash\{V_1,V_2,V_3\},d=1,2,3\) and \[\begin{aligned} \left(\begin{matrix} G^1(T_1,V_1) & G^1(T_1,V_2) & G^1(T_1,V_3) \\ G^2(T_1,V_1) & G^2(T_1,V_2) & G^2(T_1,V_3) \\ G^3(T_1,V_1) & G^3(T_1,V_2) & G^3(T_1,V_3) \end{matrix}\right) = \left(\begin{matrix} V_2-V_1,&V_3-V_1 \end{matrix}\right) g^{-1} \left(\begin{matrix}-1 & 1 & 0\\-1 & 0 & 1 \end{matrix}\right) \end{aligned}\] assures \((\nabla_{\disctmfd}\Psi)(T_1) = \left(\begin{matrix}(G^1\Psi)(T_1)&(G^2\Psi)(T_1)&(G^3\Psi)(T_1) \end{matrix}\right)\). Following the same approach to define \(G^d({\trg_j},{\vtc_i})\) on \({\trg_j}\inT\), we have \(\nabla_{\disctmfd}\Psi=\left(\begin{matrix}G^1\Psi&G^2\Psi&G^3\Psi\end{matrix}\right)\). Next, we define discretization of the divergence operator based on its adjoint relation to the gradient operator. Consider the following discretization of surface area and inner product. Let \(\area_{\trgj}\) be the area of triangle \({\trg_j}\) and \(\area_{\vtci}:=\frac{1}{3}\sum_{j:\vtci\in\trgj}\area_{\trgj}\) be the area of the barycentric dual cell of \({\vtc_i}\) (Figure [\[fig: disct illu\]](#fig: disct illu){reference-type="ref" reference="fig: disct illu"} right), and denote \(\area_{\vtc}:=\diag(A_{V_1},\cdots,A_{V_{h}})\in\mathbb{R}^{h\times h}\) and \(\area_{\trg}:=\diag(A_{T_1},\cdots,A_{T_{s}})\in\mathbb{R}^{s\times s}\) be the mass matrices of vertices and of triangles. We then define the inner products of vector fields as \(\langle U_1,U_2 \rangle_{T}:=\trace(U_1^\top\area_{\trg} U_2)=\sum_{d=1}^3 (U_1^d)^\top\area_{\trg} U_2^d\) and of functions as \(\langle\Psi_1,\Psi_2\rangle_{V}:=\Psi_1^\top\area_{\vtc}\Psi_2\). To preserve the adjoint relation between negative gradient and divergence under above inner products, i.e. \(\langle-\nabla_{\disctmfd}\Psi,U\rangle_T = \langle \Psi,\nabla_{\disctmfd}\cdot U\rangle_V\), we assign \(\nabla_{\disctmfd}\cdot U:=-\sum_{d=1}^3\area_{\vtc}^{-1}(G^d)^\top\area_{\trg} U^d\in\mathbb{R}^{h}\). With the notations of area, we write the set of probability density functions on \(\widetilde{\mfd}\) as \(\mathcal{P}(\widetilde{\mfd}):=\{P\in\mathbb{R}_+^{h}:\area_{\vtc}P = 1\}\). If the the initial density is given as \(P_0\in\mathcal{P}(\widetilde{\mfd})\), the semi-discrete constraint is then \[\begin{aligned} & \frac{\mathrm{d}}{\mathrm{d} t}P({\vtc_i},t) + \left(\nabla_{\disctmfd}\cdot M\right)({\vtc_i},t) = 0,P(\cdot,0)=P_0,\\ & P\in C([0,1];\mathcal{P}(V)), M\in C([0,1];\Gamma(\mathrm{T}\disctmfd)). \end{aligned}\] For the objective function, let \(W:\mathbb{R}^{h}\to\mathbb{R}^{s}: \Psi\mapsto\overline{\Psi},\quad\overline{\Psi}({\trg_j}) = w(\{\rho({\vtc_i}):{\vtc_i}\in{\trg_j}\})\) average the density values on each triangle. Below, we list some typical choices of \(w:=w\left(\left\{\rho(V_1),\rho(V_2),\rho(V_3)\right\}\right)\).
- Arithmetic mean: \[\omega:=\frac{1}{3}\sum_i\rho({\vtc_i});\]
- Geometric mean: \[w:=\left(\prod_i\rho({\vtc_i})\right)^{\frac{1}{3}};\]
- Harmonic mean: \[w:=3\left(\sum_i\frac{1}{\rho({\vtc_i})}\right)^{-1}.\]
We remark that these choices of average functions are useful in defining the related discrete mean-field variational problems. They connect with the gradient flow studies of Markov processes on discrete states. See related studies in. For simplicity, we select the arithmetic mean (i) in this work. We evaluate the dynamic cost on triangles \(\widetilde{\costfuncctl}:\mathrm{T}\disctmfd\to\mathbb{R}\), and the interaction and terminal cost on vertices \(\widetilde{\Costevo}:\mathcal{P}(\widetilde{\mfd})\to\mathbb{R}\), \(\widetilde{\Costterm}:\mathcal{P}(\widetilde{\mfd})\to\mathbb{R}\). With suitable choice of triangular mesh and discrete cost functions, the continuous cost is approximated by \[\begin{aligned} \widetilde{\Cost}(P,M):=& \int_0^1\sum_{j=1}^{s} \area_{\trgj} \overline{\Rho}({\trg_j},t)\widetilde{\costfuncctl}\left({\trg_j},\frac{M({\trg_j},t)}{\overline{\Rho}({\trg_j},t)}\right)\mathrm{d} t +\int_0^1 \widetilde{\Costevo}(P(\cdot,t))\mathrm{d} t +\widetilde{\Costterm}(P(\cdot,1)).\\ \end{aligned}\] where \(\overline{\Rho}(\cdot,t)=W(P(\cdot,t))\). The semi-discrete formulation of [\[eq: mfmfg opt\]](#eq: mfmfg opt){reference-type="eqref" reference="eq: mfmfg opt"} on triangular mesh \(\mathcal{M}\) is then \[\begin{aligned} \min_{P,M} \quad& \widetilde{\Cost}(P,M)\\ \text{ subject to } &\quad \frac{\mathrm{d}}{\mathrm{d} t}P({\vtc_i},t) + \left(\nabla_{\disctmfd}\cdot M\right)({\vtc_i},t) = 0,P(\cdot,0)=P_0,\\ &\quad P\in C([0,1];\mathcal{P}(\widetilde{\mfd})), M\in C([0,1];\Gamma(\mathrm{T}\disctmfd)). \end{aligned} \label{eq: mfmfg opt semi-disct}\]
Before discretizing the time interval, we discuss the semi-discrete formulations of examples [\[eg: vmfg local\]](#eg: vmfg local){reference-type="ref" reference="eg: vmfg local"} and [\[eg: vmfg nonlocal\]](#eg: vmfg nonlocal){reference-type="ref" reference="eg: vmfg nonlocal"}.
## Time discretization {#subsec: disct time}
To numerically solve [\[eq: mfmfg opt semi-disct\]](#eq: mfmfg opt semi-disct){reference-type="eqref" reference="eq: mfmfg opt semi-disct"}, we fully discretize the problem by dividing the time interval \([0,1]\) into \(n\) segments and let \({t_k}=\frac{k}{n}\). Now we consider the density on central time steps \(P=\{P(\cdot,{t_k})\}_{k=1,\cdots,n}\in(\mathcal{P}(\widetilde{\mfd}))^n\) and the flux on staggered time steps \(M=\{M(\cdot,{t_{k-\frac{1}{2}}})\}_{k=1,\cdots,n}\in(\Gamma(\mathrm{T}\disctmfd))^n\). Let the time differential operator be \[(\DtP)(\cdot,{t_{k-\frac{1}{2}}}):=\begin{cases} \frac{1}{1/n}(P(\cdot,{t_k})-P(\cdot,t_{k-1})),& \quad k=2,\cdots,n,\\ \frac{1}{1/n}(P(\cdot,t_1)-P_0(\cdot)),&\quad k=1. \end{cases}\] Then the discrete constraint set \(\widetilde{\cstr}(P_0)\) is \[\widetilde{\cstr}(P_0):=\left\{ \begin{aligned} (P,M): &~(\DtP)({\vtc_i},{t_{k-\frac{1}{2}}}) + (\nabla_{\disctmfd}\cdot M)({\vtc_i},{t_{k-\frac{1}{2}}}) = 0,\forall{\vtc_i}\inV,k=1,\cdots,n\\ & \quadP\in\mathcal{P}(\widetilde{\mfd}))^n, M\in (\Gamma(\mathrm{T}\disctmfd))^n \end{aligned}\right\}.\] Additionally, letting \[\begin{aligned} \overline{\Rho}(\cdot,{t_{k-\frac{1}{2}}})&:=\halfW(P(\cdot,{t_k})) + \halfW(P(\cdot,t_{k-1})),\quad k=1,\cdots,n\\ \widetilde{\Cost}(P,M)&:=\frac{1}{n}\sum_{k=1}^{n}\sum_{j=1}^{s} \area_{\trgj} \overline{\Rho}({\trg_j},{t_{k-\frac{1}{2}}})L\left({\trg_j},\frac{M({\trg_j},{t_{k-\frac{1}{2}}})}{\overline{\Rho}({\trg_j},{t_{k-\frac{1}{2}}})}\right) + \frac{1}{n}\sum_{k=1}^{n-1}\widetilde{\Costevo}(P(\cdot,{t_k})) + \widetilde{\Costterm}(P(\cdot,{t_n})). \end{aligned}\] we formulate the discrete optimization problem as \[\begin{aligned} &\min_{P,M} \widetilde{\Cost}(P,M) + \chi_{\widetilde{\cstr}(\rho_0)}(P,M).\\ \end{aligned} \label{eq: mfmfg opt disct}\] Here \(\chi\) is the indicator function \(\chi_\mathcal{C}(\mathbf{x})=\left\{\begin{array}{cc} 0, & \mathbf{x} \in\mathcal{C}\\ +\infty, & \mathbf{x}\not\in\mathcal{C} \end{array}\right.\) of a convex set \(\mathcal{C}\). In the next section, we focus on solving the optimization problem [\[eq: mfmfg opt disct\]](#eq: mfmfg opt disct){reference-type="eqref" reference="eq: mfmfg opt disct"}.
# Algorithm for Solving Variational MFGs on Triangular Meshes {#sec: alg}
In this section, we adapt the fast algorithm proposed in to solve the discretized potential mean-field game [\[eq: mfmfg opt disct\]](#eq: mfmfg opt disct){reference-type="eqref" reference="eq: mfmfg opt disct"}. This algorithm is based on a proximal gradient method (PGD) . To solve [\[eq: mfmfg opt disct\]](#eq: mfmfg opt disct){reference-type="eqref" reference="eq: mfmfg opt disct"}, we conduct gradient descent on the smooth component \(\widetilde{\Cost}\) of the objective function and proximal descent on the non-smooth component \(\chi_{\widetilde{\cstr}(P_0)}\). The gradient descent step is trivially \[\left(P^{(l+\frac{1}{2})},M^{(l+\frac{1}{2})}\right) = \left(P^{(l)},M^{(l)}\right) -\eta^{(l)}\nabla_{P,M}\widetilde{\Cost} \left(P^{(l)},M^{(l)}\right)\] with stepsize \(\eta^{(l)}\). The proximal descent is exactly the projection to \(\widetilde{\cstr}(P_0)\). Let the inner product in discrete spaces be \[\begin{aligned} &\Psi_1,\Psi_2\in(\mathcal{P}(\widetilde{\mfd}))^n, &\langle \Psi_1,\Psi_2\rangle_{V,t} := \frac{1}{n}\sum_{k=1}^n \langle \Psi_1(\cdot,{t_k}),\Psi_2(\cdot,{t_k}) \rangle_{V},\\ &U_1,U_2\in(\Gamma(\mathrm{T}\disctmfd))^n, &\langle U_1,U_2\rangle_{T,t} := \frac{1}{n}\sum_{k=1}^{n}\langle U_1(\cdot,{t_{k-\frac{1}{2}}}),U_2(\cdot,{t_{k-\frac{1}{2}}}) \rangle_{T}, \end{aligned}\] Then \[(P^{{(l+1)}},M^{{(l+1)}})=\proj_{\widetilde{\cstr}(P_0)}(P^{{(l+\frac{1}{2})}},M^{{(l+\frac{1}{2})}}):=\argmin_{(P,M)\in\widetilde{\cstr}(P_0)} \frac{1}{2}\left\|P-P^{{(l+\frac{1}{2})}}\right\|^2_{V,t} + \frac{1}{2}\left\|M-M^{{(l+\frac{1}{2})}}\right\|^2_{T,t} \label{eq: mfmfg disct proj}\] To solve the optimization problem [\[eq: mfmfg disct proj\]](#eq: mfmfg disct proj){reference-type="eqref" reference="eq: mfmfg disct proj"}, we introduce a dual variable \(\Psi=\{\Psi(\cdot,{t_{k-\frac{1}{2}}})\}_{k=1,\cdots,n}\in(\mathcal{P}(\widetilde{\mfd}))^n\) on vertices and the staggered time steps. The Lagrangian is therefore \[\begin{aligned} \mathcal{A}(P,M,\Psi):=&\frac{1}{2}\left\|P-P^{{(l+\frac{1}{2})}}\right\|^2_{V,t} +\frac{1}{2}\left\|M-M^{{(l+\frac{1}{2})}}\right\|^2_{T,t} \\ &+\frac{1}{n}\sum_{k=1}^{n}\langle\Psi(\cdot,{t_{k-\frac{1}{2}}}), (\DtP)(\cdot,{t_{k-\frac{1}{2}}})+\nabla_{\disctmfd}\cdot M(\cdot,{t_{k-\frac{1}{2}}}) \rangle_{V}. \end{aligned}\] If we define \(\widetilde{\partial}_t^*\) as \[(\widetilde{\partial}_t^*\Psi)(\cdot,{t_k}):= \begin{cases} \frac{1}{1/n}(\Psi(\cdot,{t_{k-\frac{1}{2}}})-\Psi(\cdot,{t_{k+\frac{1}{2}}})), &k=1,2,\cdots,n-1,\\ \frac{1}{1/n}\Psi(\cdot,t_{n-\frac{1}{2}}),& k=n, \end{cases}\] then the Lagrangian is also \[\begin{aligned} \mathcal{A}(P,M,\Psi):=&\frac{1}{2}\left\|P-P^{{(l+\frac{1}{2})}}\right\|^2_{V,t} +\frac{1}{2}\left\|M-M^{{(l+\frac{1}{2})}}\right\|^2_{T,t} \\ &+\frac{1}{n}\sum_{k=1}^{n}\langle (\widetilde{\partial}_t^*\Psi)(\cdot,{t_k}),P(\cdot,{t_k}) \rangle_{V}-\langle \Psi(\cdot,t_{\frac{1}{2}}),P_0 \rangle_{V}\\ &+\frac{1}{n}\sum_{k=1}^{n}\langle-\nabla_{\disctmfd}\Psi(\cdot,{t_{k-\frac{1}{2}}}),M(\cdot,{t_{k-\frac{1}{2}}}) \rangle_{T}. \end{aligned}\] Thus the saddle point \((P,M,\Psi)\) satisfies the linear system \[(\DtP)(\cdot,{t_{k-\frac{1}{2}}})+\nabla_{\disctmfd}\cdot M(\cdot,{t_{k-\frac{1}{2}}})=\mathbf{0},\quad k=1,\cdots,n.\] and \[\left\{\begin{aligned} P(\cdot,{t_k}) &= P^{{(l+\frac{1}{2})}}(\cdot,{t_k})-(\widetilde{\partial}_t^*\Psi)(\cdot,{t_k}),\quad k=1,\cdots,n\\ M(\cdot,{t_{k-\frac{1}{2}}}) &= M^{{(l+\frac{1}{2})}}(\cdot,{t_{k-\frac{1}{2}}}) + \nabla_{\disctmfd}\Psi(\cdot,{t_{k-\frac{1}{2}}}),\quad k=1,\cdots,n, \end{aligned}\right.\] Note that \(\widetilde{\partial}_t\) is a full rank operator, for \(k=1,\cdots,n\), \(\Psi(\cdot,{t_{k-\frac{1}{2}}})\) is the unique solution to \[(\widetilde{\partial}_t\Dt^*\Psi)(\cdot,{t_{k-\frac{1}{2}}})-\nabla_{\disctmfd}\cdot\nabla_{\disctmfd} \Psi(\cdot,{t_{k-\frac{1}{2}}}) = (\DtP^{{(l+\frac{1}{2})}})(\cdot,{t_{k-\frac{1}{2}}}) + \nabla_{\disctmfd}\cdot M^{{(l+\frac{1}{2})}}(\cdot,{t_{k-\frac{1}{2}}}). \label{eq: mfmfg phi equ}\] Since this linear solver is invariant to the data and the iteration number, in practice, we precompute it to save cost in the main iteration. We summarize our algorithm in Alg. [\[alg: pgd for mfmfg\]](#alg: pgd for mfmfg){reference-type="ref" reference="alg: pgd for mfmfg"}.
[\[alg: pgd for mfmfg\]]{#alg: pgd for mfmfg label="alg: pgd for mfmfg"}
[\[tab: homer cost compare\]]{#tab: homer cost compare label="tab: homer cost compare"}
## MFGs with non-local interactions
In this part, we show some mean-field games with non-local interaction cost.
#### The unit Sphere
In this example, we work on the triangular mesh of the unit sphere in three-dimensional space. The initial density \(P_0\) and desired terminal density \(P_1\) are spherical Gaussian. Again, we use the terminal cost \(\widetilde{\Costterm}(P(\cdot,t)) = 0.5\sum_{i=1}^{h}\area_{\vtci}P({\vtc_i},1)\log\left(\frac{P({\vtc_i},1)}{P_1({\vtc_i})}\right).\) We then compute the game with vanilla interaction cost \(\widetilde{\Costevo}_v(P(\cdot,t)) =0\) and non-local \(\widetilde{\Costevo}_n(P(\cdot,t)) =25 \sum_{i=1}^{h}\sum_{i'}\area_{\vtci}P({\vtc_i})\widetilde{\kernel}(V_i,V_{i'})A_{V_{i'}}P(V_{i'})\). The kernel is defined as \[\widetilde{\kernel}({\vtc_i},V_{i'}) = \exp\left(-(\arccos{{\vtc_i}^\topV_{i'}})^2/\sigma^2\right).\] Here \(\sigma=0.1\) and \({\vtc_i}^\topV_{i'}\) is the inner product of the two vectors in Euclidean space and \(\arccos{{\vtc_i}^\topV_{i'}}\) is the geodesic distance between \({\vtc_i}\) and \(V_{i'}\) on sphere. We use the ground truth geodesic distance for simplicity. One can also compute the shortest path on the mesh and store it when pre-processing the manifold. We conduct the quantitative and snapshot comparison in Table [2](#tab: sphere cost compare){reference-type="ref" reference="tab: sphere cost compare"} and in Figure [\[fig: sphere compare\]](#fig: sphere compare){reference-type="ref" reference="fig: sphere compare"}. The table shows our algorithm effectively leverage the dynamic, interaction and terminal cost when taking the non-local cost \(\widetilde{\Costevo}_n\). And the comparison in Figure [\[fig: sphere compare\]](#fig: sphere compare){reference-type="ref" reference="fig: sphere compare"} clearly illustrates that the non-local cost \(\widetilde{\Costevo}_n\) encourages the dispersion of mass.
[\[tab: sphere cost compare\]]{#tab: sphere cost compare label="tab: sphere cost compare"}
#### Kitten
In the last example, we work with the kitten surface (Figure [\[fig: kitten compare\]](#fig: kitten compare){reference-type="ref" reference="fig: kitten compare"}). Let the initial density \(P_0\) to concentrate on the paws and the desired terminal density on ears. We take the terminal cost \(\widetilde{\Costterm}(P(\cdot,1)) = \sum_{i=1}^{h}\area_{\vtci}\left(P({\vtc_i},1)-P_1({\vtc_i})\right)^2\) to push the mass moving from bottom to top. We also compare the non-local interaction cost \(\widetilde{\Costevo}_n(P(\cdot,t)) =\frac{1}{2} \sum_{i=1}^{h}\sum_{i'}\area_{\vtci}P({\vtc_i},t)\widetilde{\kernel}(V_i,V_{i'})A_{V_{i'}}P(V_{i'},t)\) with the vanilla \(\widetilde{\Costevo}_v(P(\cdot,t))=0\) in Figure [\[fig: kitten compare\]](#fig: kitten compare){reference-type="ref" reference="fig: kitten compare"}. The kernel is chosen as a weighted Laplacian matrix on the triangular mesh \[\widetilde{\kernel}({\vtc_i},V_{i'})=\frac{1}{\area_{\vtci}}\frac{1}{A_{V_{i'}}} \sum_{d=1}^3\sum_{j=1}^{s}\area_{\trgj}G^d({\trg_j},{\vtc_i})G^d({\trg_j},V_{i'}).\] With this choice, the interaction cost is exactly \[\widetilde{\Costevo}_n(P(\cdot,t))=\frac{1}{2}\sum_{j=1}^{s}\area_{\trgj}\left\|(\nabla_{\disctmfd}P)({\trg_j},{t_k})\right\|_2^2,\] and approximates \(\frac{1}{2}\int_{\mathcal{M}}\left\|\nabla_{\mathcal{M}}\rho(\mathbf{x},t)\right\|_{g(\mathbf{x})}^2\dd_\mfd \vx\). To reduce this cost, the density at each time step \(P(\cdot,{t_k})\) tends to be smooth on the space domain. The quantitative result in Table [3](#tab: kitten cost compare){reference-type="ref" reference="tab: kitten cost compare"} shows the value \(\frac{1}{n}\sum_{k=1}^{n-1}\widetilde{\Costevo}_n(P(\cdot,{t_k}))\) is reduced by adding \(\widetilde{\Costevo}_n\) to the objective function. And the comparisons of densities and colorbars in Figure [\[fig: kitten compare\]](#fig: kitten compare){reference-type="ref" reference="fig: kitten compare"} show that with \(\widetilde{\Costevo}_n\) in objective function, at each time step, the density distributes more uniformly on the manifold.
[\[tab: kitten cost compare\]]{#tab: kitten cost compare label="tab: kitten cost compare"}
## Computation time and accuracy {#subsec: comptime}
At the end of this numerical section, we report the computation time and accuracy of above experiments in Table [\[tab: time and acc\]](#tab: time and acc){reference-type="ref" reference="tab: time and acc"}. As the computational complexity of our algorithm depends on the number of vertices \(h\) and number of triangles \(s\) on the mesh, we also include \(h,s\) in the table. As we mentioned in section [4](#sec: alg){reference-type="ref" reference="sec: alg"}, the proximal descent step requires solving a linear system [\[eq: mfmfg phi equ\]](#eq: mfmfg phi equ){reference-type="eqref" reference="eq: mfmfg phi equ"}. And since the linear solver is invariant to iteration numbers, we precompute it to reduce the total cost in the main iteration. The times reported in Table [\[tab: time and acc\]](#tab: time and acc){reference-type="ref" reference="tab: time and acc"} include both the precomputation and the main iteration. To show that our numerical result is close to the local minimizer of the fully discretized problem [\[eq: mfmfg opt disct\]](#eq: mfmfg opt disct){reference-type="eqref" reference="eq: mfmfg opt disct"}, we report the KKT residue in Table [\[tab: time and acc\]](#tab: time and acc){reference-type="ref" reference="tab: time and acc"}. To compute the KKT residue for a given output \((P,M)\), we first solve for \(\Psi=\{\Psi(\cdot,{t_{k-\frac{1}{2}}})\}_{k=1,\cdots,n}\in(\mathcal{P}(\widetilde{\mfd}))^n\) such that \[\widetilde{\partial}_t\Dt^*\Psi-\nabla_{\disctmfd}\cdot\nabla_{\disctmfd} \Psi = \widetilde{\partial}_t(\area_{\vtc}^{-1}\partial_{P}\widetilde{\Cost}(P,M)) + \nabla_{\disctmfd}\cdot (\area_{\trg}^{-1}\partial_M\widetilde{\Cost}(P,M)).\] Then let \[\left\{ \begin{aligned} &E_{P}({\vtc_i},{t_k}):=\min\left\{\frac{1}{\area_{\vtci}}\nabla_{P({\vtc_i},{t_k})}\widetilde{\Cost}(P,M)-(\widetilde{\partial}_t^*\Psi)({\vtc_i},{t_k}),P({\vtc_i},{t_k})\right\},\forall{\vtc_i}\inV,k=1,\cdots,n\\ &E_{M}({\trg_j},{t_k}):=\frac{1}{\area_{\trgj}}\nabla_{M({\trg_j},{t_{k-\frac{1}{2}}})}\widetilde{\Cost}(P,M)+(\nabla_{\disctmfd}\Psi)({\trg_j},{t_{k-\frac{1}{2}}})=0,\forall{\trg_j}\inT,k=1,\cdots,n\\ &E_c({\vtc_i},{t_{k-\frac{1}{2}}}):=(\DtP)({\vtc_i},{t_{k-\frac{1}{2}}}) + (\nabla_{\disctmfd}\cdot M)({\vtc_i},{t_{k-\frac{1}{2}}}) ,\forall{\vtc_i}\inV,k=1,\cdots,n. \end{aligned}\right.\] The KKT residue is defined as \(\min\{\|E_{P}\|_{V,t},\|E_M\|_{T,t},\|E_c\|_{V,t}\}\). And \(P,M\) is the local minimizer of [\[eq: mfmfg opt disct\]](#eq: mfmfg opt disct){reference-type="eqref" reference="eq: mfmfg opt disct"}, if and only if the KKT residue is 0.
[\[tab: time and acc\]]{#tab: time and acc label="tab: time and acc"}
# Conclusion {#sec: conclu}
In this work, we generalize mean-field games from Euclidean space to manifolds, design an optimization based algorithm to solve variational mean field games and conduct numerical experiments on various manifolds with triangular mesh representation. We first propose both the PDE formulation and the variational formulation of the Nash Equilibrium of a mean-field game. We also establish their equivalence on manifolds. To solve the potential mean-field games on manifolds, we use triangular meshes, piece-wise linear functions and piece-wise constant vector fields for discretization. Then we apply proximal gradient method to solve the corresponding discrete optimization problems. We conduct comprehensive numerical experiments to demonstrate flexibility of the model on handling different MFG problems on various manifolds. | {'timestamp': '2022-06-06T02:16:03', 'yymm': '2206', 'arxiv_id': '2206.01622', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01622'} |
null | null |
# Introduction
The study of Exceptional Points (EPs) is at the focus of intense experimental and theoretical research . EPs are spectral singularities of non-Hermitian operators where both eigenvalues and eigenvectors coalesce , thus characterizing the dynamics of open quantum systems. Such a singularity is mathematically associated with the nontrivial topological structure of the eigenvalue manifold, and the EP corresponds to a branching point of the solution of the characteristic polynomial of the corresponding non-Hermitian operator. Beyond their theoretical and mathematical interest, EP's interest among physicists sparked from the discovery of parity-time (\(\cal PT\)) symmetry-breaking, leading to the characterization of \(\cal PT\) non-Hermitian Hamiltonians (NHHs) and to the study of phase transitions in finite-dimensional systems . Many experiments confirmed and demonstrated the unique properties of EPs and their influence on system dynamics, e.g., unidirectional invisibility , lasers with and enhanced-mode selectivity , low-power nonreciprocal light transmission, thresholdless phonon lasers, enhanced light-matter interactions, loss-induced lasing. The nontrivial properties of the EPs have also been studied and analyzed in electronics , optomechanics, acoustics, plasmonics , and metamaterials . At an EP, systems are also known for exhibiting nontrivial topological and localization properties, particularly in 1D and higher-dimensional lattice architectures . For extensive reviews see, e.g., Refs. and references therein. Many of the previously-cited works dealt with "semiclassical" configurations, where the equation of motion of a strong coherent field can be mapped onto effective Schrödinger equations, leading to the appearance of Hamiltonian EPs (HEPs) . When fully taking into account *quantum dissipative processes*, it is necessary to include the action of quantum jumps (Langevin noise), often significantly changing the dynamics of a quantum system. To circumvent such a problem, and witness the EPs of a quantum NHH, two experimental strategies have been recently realized: dilation of a non-Hermitian Hamiltonian in a larger Hermitian space and post-selection. In the later case, EPs can be seen as one of the manifestations of the non-trivial dynamics of quantum systems under post-selection. Indeed, under quite general hypotheses, the dynamics of open quantum systems can be described by a Lindblad master equation that, in turn, can be separated into the action of an effective NHH, periodically interrupted by abrupt events called quantum jumps. According to this quantum trajectory picture, the Lindblad master equation is simply the description of the average dynamics of a system continuously probed by a set of detectors (modelling the environment), each one associated with a jump operator. By postselecting those trajectories where no quantum jump occurs (no detector clicks), the spectral properties of the NHH can be investigated. Within this postselected approach, the presence of EPs admits thus a simple and fascinating explanation: the very information gained by the fact that a quantum jump did not occur induces a nonunitary state update that, in turn, introduces the non-Hermiticity necessary to witness the wanted EP. The formulation of quantum jumps, however, is not unique, and several quantum trajectory equations can be associated with the same Lindblad master equation. Even if, usually, the quantum jumps and the NHH are represented in a "standard" form--the jump operators are chosen to be orthonormal and traceless--there exist a whole class of transformations, changing the effective Hamiltonian and the jump operators, which recover the same Lindblad master equation. However, the dynamics at a single trajectory level can drastically change according to the form of the effective Hamiltonian and of the jump operators stemming from these transformations. The different forms of quantum trajectories, associated with a given Lindblad master equation, admit a clear physical interpretation and are called unraveling: they correspond to different ways to collect the information leaking from the system into the environment. Although the unconditional evolution (averaged overmthe detectors' output) is unchanged, these different monitorings modify the way the system behaves along single runs of an experiment, which are conditioned on a given sequence of detector outputs. This striking effect was experimentally demonstrated in, e.g., Ref. . Since monitoring the occurence of quantum jumps is the key ingredient for postselection, and the way the jumps operators act modifies the non-Hermitian Hamiltonian, a natural question is what effects can be witnessed with these different unravelings of the same dissipative dynamics. In this article, we demonstrate that different NHHs associated with the *same Lindlbad* master equation display *completely different* spectral properties. In particular, by modifying the form of the quantum jumps (i.e., the way quantum information is collected) and postselecting those trajectories where no quantum jump occurred, we can induce an EP or modify its properties. The article is structured as follows. In Sec. [2](#Sec:Invariances){reference-type="ref" reference="Sec:Invariances"} we introduce the Lindblad master equation and its invariances, leading to the different quantum trajectories associated with the same dynamics, as well as their postselection. We then introduce our first example in Sec. [3](#Sec:ExampleI){reference-type="ref" reference="Sec:ExampleI"}, where we consider a two-level system (qubit) with gain and losses that does not display any EPs using the "standard" representation of the quantum jumps. However, tuning the previously-introduced canonical transformations of the Lindblad master equation one can induce an EP. We then show in Sec. [4](#Sec:ExampleII){reference-type="ref" reference="Sec:ExampleII"} that it is possible to induce an EP in a driven Kerr resonator with a similar procedure, but just in the presence of photon loss. Finally, in Sec. [5](#Sec:ExampleIII){reference-type="ref" reference="Sec:ExampleIII"}, we show that the canonical transformation can also be used to tune the properties of an EPs. We present our conclusions in Sec. [6](#Sec:Conclusions){reference-type="ref" reference="Sec:Conclusions"}.
# Lindblad invariances and quantum trajectories {#Sec:Invariances}
The state of an open quantum system is captured by its density matrix \(\hat{\rho}(t)\). If the system interacts with a Markovian (memoryless) environment, and within the Born approximation, \(\hat{\rho}(t)\) evolves under the Lindblad master equation, which reads \[\frac{\partial \hat{\rho}(t)}{\partial t} = \mathcal{L} \hat{\rho}(t) =-i \left[\hat{H}, \hat{\rho}(t) \right] + \sum_{\mu} \gamma_\mu \mathcal{D} \left[\hat{J}_\mu\right] \hat{\rho}(t).\label{Eq.:Lindblad}\] In this description, \(\hat{H}\) is an Hermitian operator describing the coherent evolution of the system, while \(\hat{J}_\mu\) are the jump operators describing the dissipation induced by the environment via the action of the dissipators, which reads \[\mathcal{D} \left[\hat{J}_\mu\right] \hat{\rho}(t) = \hat{J}_\mu \hat{\rho}(t) \hat{J}_\mu^\dagger-\frac{\hat{J}_\mu^\dagger \hat{J}_\mu \hat{\rho}(t) + \hat{\rho}(t) \hat{J}_\mu^\dagger \hat{J}_\mu}{2}.\]
## Quantum trajectories {#Subsec:Postselection}
From a theoretical point of view, the Lindblad master equation is a particular form of a quantum map. On a general ground, any quantum map can be rewritten in terms of its Kraus operators, and \[\hat{\rho}(t+dt) = \sum_{\nu} \hat{K}_{\nu} \hat{\rho}(t) \hat{K}_{\nu}^\dagger,\label{Eq:Kraus}\] where the condition \[\label{Eq:condition_Kraus} \sum_{\nu} \hat{K}_{\nu}^\dagger \hat{K}_{\nu} = \hat{\mathds{1}}\] is required to ensure that the quantum map is CPTP (i.e., Completely Postitive and Trace Preserving). Given the form of [\[Eq.:Lindblad\]](#Eq.:Lindblad){reference-type="eqref" reference="Eq.:Lindblad"}, one can verify that a set of Kraus operators that recover the Lindblad master equation is \[\begin{split} \hat{K}_{0} &= \hat{\mathds{1}}-i \hat{H} dt-\sum_{\mu} \gamma_\mu \frac{\hat{J}_\mu^\dagger \hat{J}_\mu}{2} \, dt = \hat{\mathds{1}}-i \hat{H}_{\rm eff} dt, \\ \hat{K}_{\nu} &= \sqrt{\gamma_\mu dt} \hat{J}_\nu, \end{split} \label{Eq.:KrausOp}\] where the non-Hermitian Hamiltonian (NHH) \(\hat{H}_{\rm eff}\) is \[\label{Eq:NHH} \hat{H}_{\rm eff} = \hat{H}-i \sum_\mu \gamma_\mu \frac{\hat{J}_\mu^\dagger \hat{J}_\mu}{2}.\] Each Kraus operator can be associated with one of the possible outcomes of a measurement process, whose backaction on the system is associated with the \(\hat{K}_{\nu}\) operator. In this regard, the Lindblad master equation can be interpreted as the dynamics of a system upon the continuous action of several measurement instruments. Whenever the outcome \(\nu\neq 0\) is obtained at time \(t\), one of the detectors "clicks" and the system evolves under the action of \(\hat{K}_\nu\). If none of the detectors click, the system evolves according to \(\hat{K}_{0}\), as the absence of clicks still conveys some information about the system state, yielding [\[Eq:Kraus\]](#Eq:Kraus){reference-type="eqref" reference="Eq:Kraus"} when the average over the measurement outcomes is taken. From this interpretation it is natural to introduce quantum trajectories. By keeping track of the measurement outcomes (the sequence of quantum jumps), we can exactly reconstruct the state of a given system initialized in a pure state \(\ket{\Psi(t=0)}\) . If the \(\hat{K}_0\) acts, the time (unnormalized) evolution of the system is described by \[\partial_t \ket{\Psi(t)} =-i \hat{H}_{\rm eff} \ket{\Psi(t)},\] while, if the \(\nu\)th Kraus operator acts, the system evolves as \[\ket{\Psi(t+dt)} \propto \hat{J}_\nu \ket{\Psi(t)}.\] If we consider the dynamics where no quantum jumps occurs, the system evolves solely under the action of the NHH \(\hat{H}_{\rm eff}\). Thus, a postselection on those trajectories with no quantum jumps reveals the spectral properties of the NHH \(\hat{H}_{\rm eff}\) in [\[Eq:NHH\]](#Eq:NHH){reference-type="eqref" reference="Eq:NHH"}. In particular, we can define the eigenvectors \(\ket{\Psi_j}\) and the associated "eigenenergies" \(E_j\) such that \[\hat{H}_{\rm eff} \ket{\Psi_j} = E_j \ket{\Psi_j}.\] An EP of the NHH is then defined as a point where \(E_j = E_k\) and \(\ket{\Psi_j} = \ket{\Psi_k}\). Higher-order degeneracies can take place, and for this reason one calls the order of an EP the number of coalescing eigenvectors (e.g., \(\ket{\Psi_j} = \ket{\Psi_k}=\ket{\Psi_l}\) is an EP of order 3). As it has been experimentally shown in Ref. , this postselection procedure allows studying the emergence of the *Hamiltonian EPs*, i.e., the degeneracy of the NHH, by reconstructing the evolution of the quantum system. In principle, one needs perfect detectors that detect the jumps with unitary efficiency and exhibit no dark counts. In the presence of finite-efficiency detectors, however, one can still observe the effects of the EP of NHH by analyzing the associated hybrid-Liouvillian. Notice that, just as for any quantum dynamics, in order to experimentally assess the properties of a system evolving at an EP, it is not sufficient to just know that no jump occurred. Indeed, the simple measurement of the jump operators does not not yield all the information about the system state, allowing to completely characterize the properties of the system. For instance, in a qubit system, determining the presence of an EP amounts to (i) postselect the trajectory where no jump happened, and (ii) perform a measurement of the system at a given time \(t\); (iii) Repeat the measurement for different runs (having initialized the system in the same state); (iv) repeat the same procedure for several different times \(t\). This procedure highlights the presence of an anomalous dynamics associated with an EP.
## Lindbladian invariances, measurements, and new trajectories
Jump operators are usually chosen to be traceless (\(\operatorname{Tr}[\hat{J}_{\mu}] = 0\)) and orthonormal (\(\operatorname{Tr}[\hat{J}_{\mu} \hat{J}_{\nu}] \propto \delta_{\mu \nu}\)). Such a choice, although mathematically convenient, should not be privileged from a physical point of view. Indeed, the set of jump operators, associated with a given Lindblad equation, is not uniquely determined. Consider, for instance, the affine transformations \[\label{Eq:Lindblad_invariance_2} \left\lbrace \begin{split} \hat{J}_{\mu}' &= \hat{J}_{\mu} + \beta_\mu \hat{\mathds{1}}, \\ \hat{H}' &= \hat{H}-\sum_{\mu} \frac{i \gamma_\mu}{2} \left( \beta_\mu^* \hat{J}_\mu-\beta_\mu \hat{J}_\mu^\dagger \right). \end{split} \right.\] Although [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"} modifies both the Hamiltonian and the jump operators, it does not change the Lindblad master equation result. Indeed, one can easy show that, the Lindblad master equation stemming from \(\hat{H}'\) and \(\hat{J}_{\mu}'\) is the same as [\[Eq.:Lindblad\]](#Eq.:Lindblad){reference-type="eqref" reference="Eq.:Lindblad"}. However, for \(\beta\neq 0\), the effective Hamiltonian changes as \[\label{Eq:NHH_beta} \begin{split} \hat{H}_{\rm eff}(\beta) &= \hat{H}-\sum_\mu \frac{i \gamma_\mu}{2} \left( \beta_{\mu}^* \hat{J}_\mu-\beta_{\mu} \hat{J}_\mu^\dagger \right) \\ & \quad-\sum_\mu\frac{i \gamma_\mu}{2} \left( \hat{J}_\mu + \beta_{\mu} \hat{\mathds{1}} \right)^\dagger \left( \hat{J}_\mu + \beta_{\mu} \hat{\mathds{1}} \right). \end{split}\] In quantum optics, the quantum trajectory stemming from the orthonormal set \(\hat J_\nu \equiv \hat{a}(\omega)\) \[where \(\hat{a}(\omega)\) is the annihilation operator of the mode at frequency \(\omega\)\] describes a situation in which the photons emitted by the system are instantaneously detected by a photon-counter, while the set of jumps \(\hat J'_\nu\), defined by [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"}, is relevant in the case of a homodyne detection setup. In the latter case, the emitted photons are mixed with a coherent field on a beam-splitter before being detected. For usual homodyne detection, one chooses \(\vert\beta\vert^2\gg 1\), while here we will consider finite values of \(\beta\) (associated with a weaker coherent field). For this reason we call such a trajectory resulting from [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"} and from [\[Eq:NHH_beta\]](#Eq:NHH_beta){reference-type="eqref" reference="Eq:NHH_beta"} a \(\beta\)-*dyne* unraveling (see also Appendices [\[App:Homodyne\]](#App:Homodyne){reference-type="ref" reference="App:Homodyne"} and [\[App:BetaDyneModel\]](#App:BetaDyneModel){reference-type="ref" reference="App:BetaDyneModel"} for a theoretical model of this detection setup, its physical interpretation, and the derivation of the associated dynamics). It is rather remarkable that the affine transformation in [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"} produces an ambiguity in the realization of the system's quantum trajectories. For instance, consider again a bosonic system characterized, this time, by just one annihilation operator \(\hat J_\nu = \hat{a}\) (i.e., the environment induces dissipation in the form of particle losses described by \(\mathcal{D} [\hat{a}]\) *in the average dynamics*). The effect of the betadyne unraveling at a *single trajectory* level is that (i) the action of the jumps operator \(\hat J'_\nu\) do not eject one full photon from the system and (ii) when the jump \(\hat J'_\nu\) does not occur the bosonic field within the cavity gets displaced by a coherent amplitude \(\gamma_\mu \beta /2\), according to [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"}. What are then the effects of this invariance on the spectral properties of the non-Hermitian Hamiltonian, and can they be detected experimentally? That is, can one practically implement a postselection procedure that, even if the average dynamics is purely incoherent, exhibits a significantly different dynamics at the NHH level? Below, we show that the transformation [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"} can profoundly change the structure of the NHH, even inducing the presence of EPs in otherwise non-degenerate Hamiltonians. Owing to the postselected measurement, we prepared the system at an EP, and the "standard" measurement protocol on the system, i.e., the points (i)-(iv) described in Sec. [2.1](#Subsec:Postselection){reference-type="ref" reference="Subsec:Postselection"}, can demonstrate the properties of the system. Even if in the following, we will focus on the \(\beta\)-dyne transformation induced by [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"}, for the sake of completeness let us notice that not all transformations on the Lindblad master equation modify the structure of the NHH. For instance, the unitary transformation \[\label{Eq:Lindblad_invariance_1} \sqrt{\gamma'_{\mu} }\hat{J}_{\mu}' = \sum_{\nu} R_{\mu, \nu} \sqrt{\gamma_{\nu} } \hat{J}_{\nu}\] leaves both the Lindlbad master equation and the NHH unchanged if \(R_{\mu, \nu}\) is a unitary matrix. Indeed, \[\begin{split} \hat H'_{\rm eff} &= \hat H-\sum\limits_\mu \frac{i \gamma'_\mu}{2} \hat J'^{\dagger}_\mu\hat J'_\mu \\ & =\hat H-\sum\limits_{\mu,\nu,\chi} R^{*}_{\mu,\nu}R_{\mu,\chi} \frac{i \sqrt{\gamma_\nu\gamma_\chi}}{2}\hat J^{\dagger}_\nu\hat J_\chi \\ & = \hat H-\frac{i}{2}\sum\limits_{\mu}\gamma_\mu\hat J^{\dagger}_\mu\hat J_\mu = \hat H_{\rm eff}, \end{split}\] because \(R_{\mu,\nu}\) is unitary, i.e., \(\sum\limits_\mu R^*_{\mu,\nu}R_{\mu,\chi}=\delta_{\nu,\chi}\). Thus, if we "mix" the leaking fields, all postselected dynamics result in the same NHH.
# Example I: inducing an EP by postselection of a \(\beta\)-dyne trajectory {#Sec:ExampleI}
As a starting point for our discussion, let us consider a two level system, whose most general NHH reads \[\left( \begin{array}{cc} a & b \\ c & d \\ \end{array} \right) = \left( \begin{array}{cc} \tilde{a} & b \\ c & 0 \\ \end{array} \right) +d \left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \\ \end{array} \right),\] where \(\tilde{a} = a-d\). The eigenvalues read \[\frac{2d + \tilde{a}\pm\sqrt{\tilde{a}^2+4 b c}}{2},\] and the (unnormalized) eigenvectors are \[\left\{ \tilde{a} \pm \sqrt{\tilde{a}^2+4 b c}, \, 2 c \right\}.\] Consequently, the condition to observe an EP in a two-level system can be recast as \[\tilde{a} = 2 i \sqrt{b c}.\]
## EP of a qubit with loss and gain
Consider a two level system whose Hamiltonian is \[\hat{H} = \frac{\omega}{2} \hat{\sigma}_z,\] and with jump operators \[\hat{J}_1 = \sqrt{\gamma}_-\hat{\sigma}_-,\quad \hat{J}_2 = \sqrt{\gamma}_+\hat{\sigma}_+,\] where \(\hat{\sigma}_z\) is the \(z\) Pauli matrix and \(\hat{\sigma}_{\pm}\) are the raising and lowering qubit operators, respectively. If we monitor the jump operator \(\hat{J}_1\), then the effective Hamiltonian of this system is \[\hat{H}_{\rm eff} = \frac{1}{2} \left( \begin{array}{cc} \omega-i \gamma_-& 0 \\ 0 &-\omega-i \gamma_+ \\ \end{array} \right).\] Consequently, this model can *never* display any exceptional point, because the eigenvalues are always distinct and separated. Let us now consider the \(\beta\)-dyne detection of both the jump operators with the same intensity \(\beta\), via the transformation \[\hat{J}_1' = \sqrt{\gamma}_-\left( \hat{\sigma}_-+ \beta \hat{\mathds{1}} \right),\quad \hat{J}_2' = \sqrt{\gamma}_+ \left( \hat{\sigma}_+ + \beta \hat{\mathds{1}} \right).\] The corresponding effective Hamiltonian is \[\begin{split}\label{Eq:Heff_beta_1} \hat{H}_{\rm eff}(\beta) =&-i \frac{|\beta|^2 (\gamma_-+ \gamma_+) }{2} \hat{\mathds{1}} \\ & + \frac{1}{2} \left( \begin{array}{cc} \omega-i \gamma_- &-2 i \beta ^* \gamma_+ \\-2 i \beta ^* \gamma_-&-\omega-i \gamma_+ \\ \end{array} \right). \end{split}\] The NHH \(\hat{H}_{\rm eff}\) has now the right structure to *display an EP*. In particular, the eigenvalues and eigenvectors read
The EP emerges when \[\beta =\pm i \frac{\gamma _--\gamma _+-2 i \omega}{4 \sqrt{\gamma _-\gamma _+}}.\] We show the effect of \(\beta\) in Fig. [\[fig:Fig1\]](#fig:Fig1){reference-type="ref" reference="fig:Fig1"}, where we plot the overlap between the two eigenvectors of \(\hat{H}_{\rm eff} (\beta)\). In particular, we notice that there is a whole region around the \(\beta_{\rm EP}\) value where the eigenvectors almost coalesce, showing the dramatic effect that the introduction of \(\beta\) can induce on the spectral properties of the NHH. In Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"}, instead, we show the emergence of the EP as a function of \(\gamma_+/\gamma_-\), having fixed the value of \(\beta\). Compared to the \(\beta=0\) case in Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"}(a), we remark that the eigenfrequencies now change both in real and imaginary parts as a function of \(\gamma_+\).
## Physical realization
Despite its relative algebraic simplicity, the previous examples require the simultaneous postselection of both the jumps occurring from \(\hat{\sigma}^-\) and \(\hat{\sigma}^+\). While the former implies a spontaneous emission that can be, in principle, achieved via a high-fidelity detector, the latter corresponds to spontaneous excitation through gaining mechanisms, and its detection can be remarkably more difficult. Such a proof-of-concept model can be realized, however, by using a three level system instead of a qubit. Consider a three-level system, whose undriven energy eigenstates \(\ket{g}\), \(\ket{e}\), and \(\ket{f}\) are coupled by a weak coherent drive resonant with the transition between states \(\ket{g}\) and \(\ket{f}\) (c.f. Fig. [\[fig:Setup1\]](#fig:Setup1){reference-type="ref" reference="fig:Setup1"}), according to the Hamiltonian \[\begin{split} \hat{H}(t) =& \ket{e}\bra{e} + \left( 2 \omega + \delta \omega \right) \ket{f}\bra{f} \\ & \, + \Omega \left( \ket{g}\bra{f}e^{i\omega_{ef}t} + \ket{f}\bra{g}e^{-i\omega_{ef}t} \right). \end{split}\] The spontaneous emission of photons induces the decay of level \(\ket{f}\) to \(\ket{e}\), and from level \(\ket{e}\) to the ground state \(\ket{g}\), via the Lindbladian: \[{\cal L}_1= \gamma_{eg} \mathcal{D}\left[\ket{g}\bra{e} \right]+ \gamma_{fe} \mathcal{D}\left[\ket{e}\bra{f} \right].\] If we assume that \(\gamma_{fe} \gg \gamma_{eg}, \Omega\), we can adiabatically eliminate the state \(\ket{f}\). The combined action of the driving and the dissipation results in a new Lindbladian \[{\cal L}_2 = \gamma_{eg} \mathcal{D}\left[\ket{g}\bra{e} \right]+ \gamma_\text{eff}\mathcal{D}\left[\ket{e}\bra{g} \right],\] which therefore implements the wanted model with \({\gamma_+ \equiv \gamma_\text{eff} = 4\Omega^2/\gamma_{fe}}\) and \(\gamma_-\equiv \gamma_{eg}\). Notice that a jump from state \(\ket{g}\) to \(\ket{e}\) is associated with the emission of a photon at the frequency \((\omega+\delta\omega)\), which can thus be detected with a photon counter. Such a jump can be distinguished from the one associated with \(\ket{g} \to \ket{e}\), because the latter leads to the emission of a photon at frequency \(\omega\).
# Example II: Inducing an EP in the driven Kerr resonator. {#Sec:ExampleII}
In this section, we present a protocol based on a driven-dissipative Kerr resonator (see ), which requires detection of only emitted photons (i.e., the detection of one jump operator) in order to reveal the emergence of EPs in the postselected effective Hamiltonians. The Hamiltonian of a Kerr resonator, in the frame rotating on the cavity frequency, is \[H =-\Delta \hat{a}^{\dagger} \hat{a} + U\hat a^{\dagger2}\hat a^2-i(\alpha \hat a^{\dagger}-\alpha^*\hat a),\] where \(\hat a\) (\(\hat a^{\dagger}\)) is an annihilation (creation) field operator, \(\Delta\) is the pump-to-cavity detuning, \(U\) is a Kerr nonlinearity, \(\alpha\) is the amplitude of a the coherent field which drives the resonator. We assume that the system is subject to one-photon loss events described by the dissipator \(\mathcal{D}[\hat{a}]\), occurring at a rate \(\gamma\). Let us now assume that the system is at resonance, \(\Delta=0\), and it is postselected in a \(\beta\)-dyne picure; thus, the NHHs \(\hat H_{\rm eff}(\beta)\) is coherently displaced by the amplitude \(\beta\), according to [\[Eq:NHH_beta\]](#Eq:NHH_beta){reference-type="eqref" reference="Eq:NHH_beta"}. Assuming the weak-driving limit, i.e., \(\alpha\ll\gamma, U\), and assuming also that \(\beta\ll\gamma\), one can write down the effective Hamiltonian in the two-photon limit (i.e., truncating the Fock space at two photons), resulting in a three-level system that reads \[\label{Hkerr2} H_{\rm eff} = \begin{pmatrix} 0 & i\alpha^*-i\gamma\beta^*& 0 \\ -i\alpha &-\frac{i}{2}\gamma & i\sqrt{2}(\alpha^*-\gamma\beta^*)\\ 0 &-i\sqrt{2}\alpha &-i\gamma+2U \end{pmatrix}-\frac{i\gamma|\beta|^2}{2}{\mathbb I}_3.\] If \(\beta=0\), no combination of parameters of the NHH results in an EP. In other words, whenever one is monitoring the system's environment and leaked photons are not "displaced" by a coherent field, then the corresponding postselected NHH \(\hat H_{\rm eff}(\beta=0)\) does not have any spectral singularity \[see, e.g., (a)\]. For \(\beta \neq 0\), instead, the condition for the NHH to have an EP of the second order reads: \[\begin{aligned}
\label{EPcond} &a^2-3b-9ab+27c+2a^3\nonumber \\ &-3\sqrt{3}\left[27c^2+(4a^3-18ab)c-a^2b^2+4b^3\right]^{\frac{2}{3}}=0, \end{aligned}\] where \[\begin{aligned}
a &=& \frac{3i\gamma}{2}-2U, \, b = 3\alpha\beta^*\gamma-3|\alpha|^2-\frac{1}{2}\gamma^2-iU\gamma, \\ c &=& (i\gamma-2U)\left(\gamma\beta^*\alpha-|\alpha|^2\right). \end{aligned}\] Having fixed the system parameters, these equations can be numerically solved to find the values of \(\beta\) resulting in an EP (see ). Note that only a second-order EP can be observed in the \(\hat H_{\rm eff}(\beta)\), since at most only two of the three eigenvalues of the NHH coincide. We conclude that, if one *postselects* those experiments where the detector *never* clicks, and for an appropriate choice of \(\beta\), satisfying the condition in [\[EPcond\]](#EPcond){reference-type="eqref" reference="EPcond"}, then the effective Hamiltonian in [\[Hkerr2\]](#Hkerr2){reference-type="eqref" reference="Hkerr2"} exhibits a second-order EP in the system \[see (b)\]. We stress that the full Lindbladian of the system does *not* have a Liouvillian EP, in the sense described in Ref. . Indeed, the Linbladian describes the dynamics of the operator averages and, as such, its averaging over various quantum trajectories eventually smears out EPs which emerge in certain effective NHHs. The presence of an EP in the NHH is thus an effect which can only really emerge at the single-trajectory level.
# Example III: shifting an EP {#Sec:ExampleIII}
The effect of \(\beta\) is not only to generate new EPs in a quantum system; adjusting \(\beta\) also allows one to *shift* the position and *change* the nature (i.e., the associated eigenvalues and eigenvectors) of an existing EP. Consider a resonantly driven two level system, whose Hamiltonian in the pump frame reads \[\hat{H} = \frac{\omega}{2} \hat{\sigma}_x.\] Let us assume that the system is affected by a dissipation channel \(\gamma_-\mathcal{D}[\hat\sigma_-]\). The effective non-Hermitian Hamitonian of such a system in the standard representation is given by \[\hat{H}_{\rm eff} = \frac{1}{2} \left( \begin{array}{cc} -i \gamma_-& \omega \\ \omega & 0 \\ \end{array} \right).\] which exhibits an exceptional point for \[\label{eq:excpoint_ex2} \omega = \frac{\gamma_-}{2}.\] We now assume that the emitted photons are detected via the \(\beta\)-dyne scheme. The associated non-Hermitian Hamiltonian is: \[\label{Eq:EP_betadyne} \begin{split} &\hat{H}_{\rm eff}(\beta) =-i \frac{|\beta|^2 \gamma_-}{2} \hat{\mathds{1}} \\ & \quad + \frac{1}{2} \left( \begin{array}{cc}-i\gamma_- & \omega \\ \omega-2 i\beta^*\gamma_-& 0 \\ \end{array} \right), \end{split}\] whose eigenvalues are
The exceptional point is now determined at \[\beta = i\frac{4\omega^2-\gamma_-^2}{8\gamma_-\omega}.\] For \(\beta = 0\) we retrieve [\[eq:excpoint_ex2\]](#eq:excpoint_ex2){reference-type="eqref" reference="eq:excpoint_ex2"}; for different values of \(\beta\), the EP position, eigenvalues, and eigenvectors change. We show this effect in Fig. [\[fig:Fig5\]](#fig:Fig5){reference-type="ref" reference="fig:Fig5"}. Notice that the shift of the EP comes at the expense of a decreased no-jump trajectory probability. A detailed computation of the postselection probability is reported in the Appendix [\[App:BetaDyneModel\]](#App:BetaDyneModel){reference-type="ref" reference="App:BetaDyneModel"}. A final comment is necessary. As we already discussed, the postselection procedure of the Lindblad master equation can be, in principle, associated with different types of unraveling. There are myriads of possible trajectories, including those with various values of \(\beta\), inducing different coherent displacements \[due to the Lindbladian invariance in [\[Eq:Lindblad_invariance_2\]](#Eq:Lindblad_invariance_2){reference-type="eqref" reference="Eq:Lindblad_invariance_2"}\]. This illustrates the fundamental difference between Liouvillian and Hamiltonian EPs. Liouvillian EPs exist at the level of the average dynamics of an open quantum system, and appear independently of the specific characteristics of the system-environment coupling. The EPs of a NHH, which are induced by post-selection, appear only for specific types of jump operators. Furthermore, the fact that EPs of an NHH can be "shifted" by different unraveling via the action of \(\beta\) is also a signature of the relative fragility of the EPs of postselected NHHs. Indeed, both the system's parameters and those of the detector, whose action determines the effect of the postselection procedure, must be finely tuned.
# Conclusion {#Sec:Conclusions}
By exploiting homodyne Lindbladian invariance, i.e., by displacing the emitted leaking photons with a laser field and keeping the whole Linbladian unchanged, we show that one can end up with different forms of quantum trajectories. Under postselection of no-jump trajectories, this scheme generates dynamics generated by a Non Hermitian Hamiltonian (NHH) whose spectral properties can be tuned via the parameters of the laser field used for the displacement. We illustrate the potential of this scheme on three examples based on simple quantum systems and realsistic detection setups, where Exceptionnal Points (EPs) can be generated or controlled only by changing the displacement of the leaking photons. This control on the EP however comes at the price of a more stringent postselection, decreasing the probability of occurrence of the no-jump trajectory. More generally, our approach exploits the mathematical invariance of the Lindblad equation to provide a whole toolbox to engineer EP properties, opening the perspectives of new implementations of EP and their predicted applications, e.g. in metrology or optimal energy transfer. | {'timestamp': '2022-06-06T02:16:30', 'yymm': '2206', 'arxiv_id': '2206.01639', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01639'} |
# Introduction
Ownership can be described, in very broad terms, as the set of exclusive rights over property, usually classified as tangible (e.g., real state, chattel) or intangible (e.g., intellectual property, digital objects). Such rights are universally recognized as fundamental , yet the very concept of ownership evolves over time, according to the mores of society and technological innovations of the place and moment. Ownership has in fact been intrinsically linked to information technology (in its broadest sense) from the very beginning of history, as attested by the small clay objects inscribed with proto-cuneiform dating back to 8000 BCE found in the Near East---called *tokens*---used to keep track of the number of animals owned . Millennia later, the printing press, another technological writing milestone, greatly eased the massive copying of books and also gave way to a new form of piracy of intellectual property . More recently, digital technology, which allows perfect copies of digital objects at near-zero cost, brought into question aspects of goods once considered to be central , such as rivalry (impossible simultaneous use) and scarcity (limited availability). Incidentally, such «fundamental shift from tactile to digital, physical to code, and hard to soft media» in ownership is also reflected by hypertext within the history of information technology, as described by . Moreover, present-day digital ecosystems---with most based on hypermedia---have heavily influenced our ownership experience. For instance, whereas traditional market models are based on ownership, the *sharing economy* is based on using and sharing products among each other, and is driven by consumer behavior (preference for convenience, low prices, sustainability), social networks and electronic markets, and mobile devices and electronic services . Another example is the push towards a *subscription model* of digital goods and services, such as news, music, video, videogames, and apps. Nonetheless, not all changes in digital ownership are welcome, and much debate is being made regarding the rights of buyers, creators, and distributors of digital objects, and the loss of expected privileges associated with ownership . Further, the increasing practice of license-only digital objects and the abuse of digital rights management (DRM) have given rise to harsh criticism against this "end of ownership" in the digital economy .
# A new kind of digital ownership
Discontent with diminishing digital property rights led many tech enthusiasts to seek alternative approaches. Arguably, one of the most interesting is the use of a non-fungible token (NFT) to certify the authenticity and ownership of tangible or intangible assets. An NFT is a unique identifier recorded using distributed ledger technology (DLT), which is based on decentralized immutable lists of records spread over a peer-to-peer network without the need of a trusted authority, with the most famous example being blockchain . NFTs are a specific form of *smart contracts*, versatile transaction protocols that automatically execute, control, and document an agreement without the need of a trusted intermediator . Amusingly enough, NFTs seem to carry on the ancient practice of using *tokens* to keep track of property, only this time via digital cryptographic hashes instead of inscriptions on clay. In the last couple of years, there has been a surge in public interest on NFTs due to the high-profile sale of certain digital assets , such as the sale for US\$5.4M of the initial implementation source code of the World Wide Web by Tim-Berners Lee.[^1] Users and buyers of NFTs are allured by the promise of ownership, traceability, and financial gain . And the high figures involved in much publicized sales have sparked a frenzy in the development of NFTs applications and ecosystems in different sectors, such as collectibles, sports, visual arts, and video games . Even "traditional" social media have begun integrating NFTs. Twitter now can display special NFT profile pictures with a hexagonal shape present in the Ethereum blockchain (the most common DLT for NFTs), and Reddit has established a collaboration with OpenSea (the largest NFT marketplace) to sell unique animated avatars, called CryptoSnoos. Many more ideas are emerging around NFTs and DLT beyond finance, such as decentralized autonomous organizations (DAOs), a novel manner to codify and automatically manage governance via smart contracts without a central leadership . Hence, due to the aforementioned commonalities with respect to information technology, I believe that using an hypertext approach to augment our capabilities on digital ownership based on DLT is a logical yet untapped opportunity.
# Envisioning hyperownership
Bitcoin, the first successful decentralized virtual currency, gave way to the idea of using DLT as a means to register property ownership. In particular, posited that property could be seen as a set of information describing who may do what, when, and with which resource: «property is the law of lists and ledgers». In this perspective, property does not consist in the *thing* itself, but in the packaging, tracking and transmission of data regarding ownership via information systems; thus, a property system has two key actions: *store* and *communicate* ownership information. However, DLT has mainly focused on improving the latter action from the point of view of machines, neglecting the human element. For example, in a survey of DLT applications from the perspective of human-computer interaction (HCI), state that these «are currently overwhelmingly driven by a mix of engineering, investment and crypto-anarchist visions», and they express the need for more research on the human challenges of DLT, such as design for trust, algorithmic governance, influence on societal and individual values, ease of use for end users, and implications of publicly sharing one's property data . further confirm the need of a human-centered approach given crucial DLT issues: high onboarding learnability; lack of design considerations for usability, efficiency, and accessibility; and design challenges in decentralized applications . In this respect, I believe there is an unexploited opportunity in using a hypertext approach to tackle these issues, for as hypertext has been historically touted as a means to augment human intellect , by allowing a person to approach complex problems, to fulfill their comprehension needs, and to derive solutions to problems . To seize such an opportunity, I think the view posited by of a «hypertext method of inquiry, as way of viewing arbitrary systems», is particularly apt. In this view, hypertext shifts its focus from a particular set of technologies and techniques, to the more general explicit association of information. *Hyperownership* is thus a concept to augment human capabilities on digital ownership, by synthesizing these two ideas: property is the law of lists and ledgers, and hypertext is a method of inquiry of arbitrary systems. Based on their form of structure from a hypertext perspective , NFTs are a first class structure, as by definition DLT records explicit relationships between resources and allows the association of metadata with a link for a given transaction. On top of these records, we could use a new representation to exploit the explicit ownership associations in new ways. For instance, we could use *temporal bipartite graphs* to represent the evolving relationships between the two main kinds of DLT entities: tokens (property) and their respective account (owner). Compared to their static and unipartite counterparts, temporal bipartite graphs model better the complexity and dynamism of real-world problems . In this *owner--property* temporal bipartite graph, the state of ownership in a given moment could be computed from a derived static graph, i.e., a *snapshot*, in which each token is associated with at most one owner and each owner is associated with zero or more tokens. A second ownership-based useful representation could be a *property--usage* temporal bipartite graph, in which we model distinct usages of property tokens in a given context (e.g., profile picture in social media, an audio or video reproduction). In the snapshots of this graph, each property could be associated to zero or more usages. Finally, a third temporal bipartite graph, *owner--usage* could be derived from the first two. Further styling could be done through automatic context-aware hypermedia sculpting to remove irrelevant links based on context or manual calligraphic linking by users . With such an approach, illustrated in Fig. [\[fig:diagram\]](#fig:diagram){reference-type="ref" reference="fig:diagram"}, we could augment human capabilities to inquiry on ownership by: using novel HCI techniques to visualize and interact with digital property; leveraging recent advances in big graph mining to discover interesting phenomena and gain insight into the dynamics of ownership; integrating existing knowledge graphs to enrich the semantics of the ownership network and viceversa; creating novel non-linear narratives regarding the history of digital objects and the preferences of owner accounts; easing the exploration and recommendation of potentially interesting digital objects to buy, use, or admire; and much more.
# Risks and challenges
Due to the uncertainties surrounding NFTs, there are significant risks in undertaking the realization of hyperownership. For instance, despite the promise and eagerness of improved usage and financial control and financial by digital content creators, many are divided on the merits of NFTs . Above all, crucial issues remain unanswered regarding the ownership of NFTs and the objects that they represent , such as a likely economic bubble, legal recognition, and the currently rising fraud and stealing in the community; as simply put by : «The NFT is a social contract that values property over material experience. That contract can be broken.» Nevertheless, I think we should look beyond these hectic times and focus on the innovative technology and techniques of NFTs and DLT in general. After all, the hypertext community has already gone trough a similar tumultuous technological period caused by hype and unwise investments, during the rise of the World Wide Web and subsequent *dot-com bubble burst* . Many technical and ethical challenges must also be resolved to realize hyperownership. For instance, at the moment we have competing DLT and NFT community-driven standards, resulting in low cross-platform compatibility, albeit many initiatives are in course to remedy this . In addition, DLT and related technologies present significant challenges in software engineering . For instance, many DLT applications store assets on the InterPlanetary File System (IPFS), a distributed peer-to-peer hypermedia protocol aimed to be resilient and persistent , which entails a different development approach. In fact, given this shift in paradigm, many people call Web3 a potential decentralized and token-based iteration of the Web , albeit this has been decried as a mere buzzword . Still, Web3 tenets might represent the next step in hypermedia infrastructure, which began as monolithic systems, then client-server systems, later as open hypermedia systems (OHS), and currently as component-based OHS . Hence, distributed and decentralized component-based OHS (using for instance IPFS for media storage) should be further explored. Another significant technical challenge concerns scalability, especially of a graph representation of DLT records which might manifest rapid exponential growth . In this regard, the community on big graph processing systems is working towards adapting new distributed workloads, standard models, and suitable performance metrics to model complex real-world phenomena, with a keen eye towards the possibility of creating a Big Graph Memex , inspired by Bush's original Memex concept. Hyperownership as described herein thus fits and builds upon such an ambitious vision. Last but not least, ethical issues are also significant for hyperownership. Privacy, a main driver of DLT , is far from perfect, and as some of us can imagine, not everyone is keen in facilitating access to the information regarding what we own. Such issues are also present in hypertext, given that despite its potential to empower users and democratize access to information, it is susceptible to surveillance and manipulation . Indeed, both sculptural and calligraphic hypertext styles could be used to restrict access or to manipulate the narrative regarding the ownership of certain assets by certain individuals or groups. In fact, it could argued that both hypertext and DLT influence and are influenced by political motives, as respectively demonstrated by the OHS and DAOs. Hence, particularly attention should be paid to these sensitive issues in order to avoid abuse, intentional or not.
# Conclusion
Once again, novel technology is changing the perception and experience of ownership, more so in an environment as *interwingled* as the digital economy. This change represents an opportunity for the development of novel hypertext ideas, such as hyperownership, to unravel and augment our relationship with digital property. Hence, I invite the hypertext community to look and take inspiration by the novel (yet ancient) *token economy*.
[^1]: <https://www.bbc.com/news/technology-57666335> | {'timestamp': '2022-06-06T02:14:29', 'yymm': '2206', 'arxiv_id': '2206.01566', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01566'} |
null | null |
null | null |
null | null |
## PH-preserving MOR techniques | {'timestamp': '2022-06-06T02:15:37', 'yymm': '2206', 'arxiv_id': '2206.01608', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01608'} |
# Introduction {#introduction .unnumbered}
Today's ocean numerical prediction skills relies on the availability of in-situ and remote ocean observations at the time of the predictions only. Because observations are scarce and discontinuous in time and space, numerical models are often unable to accurately model and predict real ocean dynamics, leading to a lack of fulfillment of a range of services that require reliable predictions at various temporal and spatial scales. To this end, it is crucial to have sustained and reliable observations of the parts of the ocean where the predictions are made. The process of constraining free numerical models with observations is known as data assimilation. The primary objective is to minimize the misfit of model states with the observations while respecting the rules of physics. Observations assimilated in ocean forecasting systems now include altimetry, ocean colour, surface velocities, sea ice and data from emerging platforms such as ocean gliders. Many systems now employ multi-model approaches or ensemble modeling techniques. A variety of data assimilation are used in operational oceanography and a detailed review of the data assimilation methods can be found in. The most common data assimilation method closely parallels the procedures employed in numerical weather prediction in which ocean state estimates are computed sequentially through time, and the resulting estimates updated when sufficient new observations become available. The caveat of this approach is that measurements are used only once, at the time of the prediction. The information contained in the history of the measurements and its role in the determinism of the prediction is, therefore, not accounted for. Consequently, historical measurement cannot be used in real-time forecasting systems. Because oceans play a key role in global issues such as climate change, food security, and human health, effort to create long term repositories with global quality assurance/quality control standards have received significant support over the last decade. These efforts have come to fruition in the Global Ocean Observing System (GOOS). GOOS has opened a field of opportunity for new collaborations across regions, communities, and technologies facilitating enhanced engagement in the global ocean observing enterprise to benefit all nations. In the United States, these efforts led to the creation of the U.S. Integrated Ocean Observing System (IOOS), which through its eleven Regional Associations implement regional observing systems covering all U.S. coasts and Great Lakes with activities spanning from head of tide to the U.S. exclusive economic zone. Through these efforts comprehensive ocean datasets, spanning several decades have become available to the research community and novel technologies rooted in big data science such as Artificial Intelligence (AI) have emerged. The research presented in this paper provides a novel approach rooted in AI to expand the usability of observations made before the time of the prediction. To demonstrate our method, we take advantage of a recent and comprehensive set of observations in the Gulf of Mexico (GoM), which has received over many decades, and continues to receive, a lot of attention by oceanographers. The GoM ocean dynamics is controlled by the pulsating Loop Current, which is the most energetic circulation feature of the basin. Predicting the Loop Current evolution is fundamental to almost all aspects of the GoM, including (a) anthropogenic and natural disaster response; (b) the prediction of short-term weather anomalies and, hurricane intensity and trajectories; (c) national security and safety; and (d) ecosystem services. The recent Deep Water Horizon oil spill in 2010 has bolstered a large amount of research toward improving our understanding and forecasting of the Loop Current System in hope of mitigating further environmental and ecological damage. Furthermore, because of the reinforcing interaction between hurricanes, the Loop Current and its eddies, long-term prediction of the Loop Current states is becoming more and more relevant. Thus, developing accurate and robust medium-term forecasting models of the Loop Current System and its eddy formation is imperative and relies on the availability of ocean observations. Improved ocean observing systems are expected to reduce the uncertainty of ocean/weather forecasting and to enhance the value of ocean/weather information throughout the Gulf region. The remainder of the paper is organized as follows. In Section 2 [\[sec:Sustained temporal correction model\]](#sec:Sustained temporal correction model){reference-type="ref" reference="sec:Sustained temporal correction model"} we provide an overview of the concept of data usability temporal expansion. Section 3 [\[sec:Dataset\]](#sec:Dataset){reference-type="ref" reference="sec:Dataset"} describes the numerical models fields and the observations used in this study. The usability temporal expansion concept is applied to both a free and data assimilated simulations. Section 4 [\[sec:Deep Learning\]](#sec:Deep Learning){reference-type="ref" reference="sec:Deep Learning"} introduces the deep learning model used in this study. Section 5 [\[sec:Methodology\]](#sec:Methodology){reference-type="ref" reference="sec:Methodology"} presents the implementation of the temporal expansion concepts, which results in the transformation of the ocean models' velocity fields beyond the observation period. Discussion and conclusion follow in Section 6
# Temporal Expansion Model Concept {#sec:Sustained temporal correction model .unnumbered}
The rationale behind this concept lies in the fact that ocean observations in the data assimilation process only serve once, at the time of the forecast, when the background field is created. They are no longer used in forecasts past the time of the observation except in three-dimensional variational data assimilation where all data collected within a short time window (days) are assimilated as though collected at a single time. Considering, observations are no longer relevant after a few days and become history. However, there is a history of the ocean dynamics captured in the observations that is not part of the constraint provided by the data assimilation concept. The dynamical history in the observations can be captured with deep learning models as shown by prediction exercises of the Loop Current dynamics, , ,. Deep learning allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction. These methods have dramatically improved the state-of-the-art in speech recognition, visual object recognition, object detection and many other domains such as drug discovery and genomics. Deep learning discovers intricate structure in large data sets by using the back-propagation algorithm to indicate how a machine should change its internal parameters that are used to compute the representation in each layer from the representation in the previous layer. Ocean numerical models are mathematical approximations of the laws of physics that are used to estimate ocean states given a set of boundary conditions and constraints, such as the atmospheric and the tidal forcing. When not constrained by data assimilation, the simulated states of the \"free-run\" are likely to be very different from the the real ocean state, although the main features of the ocean dynamics will be captured and resolved. These features include the mesoscale dynamics, the water masses, the eddy kinetic energy, among others, whose variability and magnitude is likely to be dissimilar to the observed ones, although existent in both natural and virtual systems. Therefore one can assume that both the simulated and real systems are on a parallel track although different in many ways. While Deep Learning has been mostly applied to a single data set for pattern recognition, we show in this study that it can be re-purposed to learn the differences between the simulated and real systems and their evolution over time and space. The acquisition of that knowledge between a given ocean numerical model simulation and a set of observations in the model domain over a time period concurrent to both, enables the prediction of the differences with a Deep Learning model. Therefore, the knowledge of the differences can be converted into a model field observations-based correction tool, called a Transform Model, just like a data assimilation method, with the advantage of predicting the correction of the model field beyond the period of observation. This methods also accounts for the common history of the model fields and measurements in the correction process. We will show that the improvements to the model fields, namely the reduction of errors is significant, despite the fact that no physical constraints were applied to the correction. This correction method evaluation would certainly benefit from a numerical model integration to enforce the model conservation constraint although they might be different from the ones set by the measurements. However, testing of the transformed field in a numerical model is beyond the scope of this study. The concept of the Transform model is depicted in Figure [\[fig:Temporal_model\]](#fig:Temporal_model){reference-type="ref" reference="fig:Temporal_model"}. It is based on the availability of long enough concurrent time series of model field and observations. Both are jointly used in a Deep Learning model that creates a transformation tool of the numerical model field based on the observations. The Transform Model can then be used to correct the numerical model field at any time in the future and in the past, outside of the period of measurements. It thus expands the usability of the observations beyond the measurement period, both before and after.
# Data sets {#sec:Dataset .unnumbered}
For the demonstration of our Transform Model we used three data sets that overlap in time, composed of two numerical models current velocity vector fields and one set of in-situ current measurements from an observing array in the GoM. Model's inherent dynamics may influence how effective the Transform Model, which will be shown in Section 5 The first numerical model data set was obtained from the HYbrid Coordinate Ocean Model (HYCOM) Consortium. The data was generated by the HYCOM \(+\) NCODA Gulf of Mexico \(1/25^{\circ}\) Reanalysis (GoMu \(0.04\)/expt\_\(50.1\)), which spans year 1993 to 2012 and has a horizontal resolution of \(4.4\)km. HYCOM \(+\) NCODA is a data-assimilative hybrid isopycnal-sigma-pressure (generalized) coordinate ocean model. The system uses the Navy Coupled Ocean Data Assimilation (NCODA) system, for data assimilation. NCODA uses the model forecast as a first guess in a three-dimensional variational scheme and assimilates available satellite altimeter observations (along track obtained via the NAVOCEANO Altimeter Data Fusion Center), satellite and in-situ Sea Surface Temperature (SST) as well as available in-situ vertical temperature and salinity profiles from XBTs, ARGO floats and moored buoys. The Modular Data Assimilation System (MODAS) is used for downward projection of surface information. The experiment includes tidal forcing. As noted here, no current vector flow information is used in the data assimilation process. The second model's current vector fields were obtained from a free-running Massachusetts Institute of Technology (MIT) generalized circulation model simulation of the GoM circulation (MITgcm-GoM), a z-level model. The MITgcm--GoM model was originally developed for state estimation and prediction of the upper ocean circulation in the GoM, including Loop Current evolution and eddy shedding. For these purposes, satellite-derived ocean surface observations and subsurface in situ observations were assimilated using a four-dimensional variational (4DVAR) method. For the data set used in this study, however, no data assimilation is performed. The model uses a telescopic grid with a horizontal resolution of 1/20° × 1/20° in the central GoM which decreases to 1/10° × 1/10° toward the boundaries and western part of the domain. For this study, daily output from the simulation for the four years forced by atmospheric and lateral boundary condition data from 2009 to 2012 are used. No tidal or atmospheric pressure forcing is applied. The in-situ data were obtained from the Dynamics of the Loop Current in U.S. Waters experiment (hereafter Dynloop), a comprehensive observational study of the Loop Current in the eastern GoM, ,. The observational array consisted of the instrumentation of nine tall-moorings and seven short-moorings, and an array of 25 pressure-equipped inverted echo sounders (PIES). The array system measured the water column velocity for 2.5 years, beginning in March 2009. Each tall-mooring included water velocity measurements made from an upward facing 75 kHz ADCP deployed near a depth of 450 m, as well as point current meters at five depths from around 600-3000 m, providing high resolution water velocity data between about 60-440 m depths and much lower resolution water velocity data below this. Additionally, each short-mooring included a single point current sensor located 100 m above the sea floor, providing near bottom water velocity data. The PIES array provided direct measurement of pressure and acoustic round trip travel times from the sea floor to the sea surface, which were used to create vertical profiles of density, salinity, and temperature. These pressure records combined with estimated horizontal density gradients were used to calculate geostrophic water velocities. This array was located to cover both east and west sides of the Loop Current between the West Florida Slope and the Mississippi Fan, and was also centered over the zone where Loop Current eddies typically separate from the Loop Current. The horizontal separation between moorings was around \(50-80\) km and between the PIES sensors was around \(40-50\) km. These recorded data were used to construct the measurement-based water velocity matrix used in this study. More details on the creation of the velocity matrix can be found in. The horizontal resolution varies between 30 and 50 km and only the first 500 m, corresponding to 26 vertical layers, was used in this study. The time resolution for the velocity data was 12 hours, which corresponds to 1810 data frames for each u and v velocity component. The final matrix dimensions were of 1810 × 26 × 29 × 36 for each component.
# Transform Model {#sec:Deep Learning .unnumbered}
The deep learning network of choice for our Transform Model is the u-shaped Convolution Neural Network (CNN), also known as U-Net. The typical use of convolutional networks is on classification tasks, where the output to an image is a single class label. However, in many visual tasks, especially in biomedical image processing, the desired output should include localization, i.e., a class label is supposed to be assigned to each pixel. The architecture of this network consists of two parts. The first part is the analysis path called the encoder, which is similar to the CNN architecture widely used for image feature extraction. The second part is the synthesis path called the decoder, where the compact features are expanded to the input dimension. In the image case, the feature set, or code, is decoded back to a (segmented) image of the same dimension as the input image. The name of U-Net is based on the U structure of the network. This type of network has been widely used in medical image segmentation applications, for land segmentation studies in remote sensing images, for cloud and shadow segmentation, for modeling and prediction of coastal weather events in the Netherlands, and for coastal wetland classification, among many other applications. For our application, the U-Net is slightly modified at the output layers to make it suitable for a regression problem instead of the original structure that was designed for image segmentation. The problem can be seen as a time series transformation through a regression analysis. Because the U-Net will learn both the spatial and temporal features of the training data, it is called hereafter the Spatial Temporal U-Net (STU-Net). Hence, the last layers after the convolution layers have been replaced by a regression layer, instead of a softmax and a segmentation layer in the original structure. The final model structure consists of four stages of encoders as shown in Figure [\[fig:lgraph\]](#fig:lgraph){reference-type="ref" reference="fig:lgraph"}. The input is made of a four-dimensional input layer, which is organized as \(x,y,u/v,t\). Here \(x,y\) are the coordinates of the sample location, \(u/v\) are the zonal/meridional velocities respectively, and \(t\) denotes the time of the measurements. The implementation of the STU-Net is two-fold (Fig. [\[fig:DL_model_tempora\]](#fig:DL_model_tempora){reference-type="ref" reference="fig:DL_model_tempora"}): first is the training phase, where both the numerical model and the observations are fed to the STU-Net. The relationship between the measurements and the numerical model is represented by the neural network structure, including its hyper parameters and weights. The errors between the two fields are used to adjust the weights of the model until the training process reaches a certain number of epochs. With a Mini-batch loss set at 0.1 and Mini-batch Root Mean Squared Error (RMSE) of 0.5, 120 epochs were necessary. The cost function limit used in the training stage was set by the RMSE. At the end of the training stage, the obtained regression layers constitute the Transform Model that will be used to correct the numerical model fields before, during and after observations started in the second step of the STU-Net implementation, which is the inferring phase. To apply the transformation back in time, namely before the observation period, training is conducted backward in time to learn the temporal features in reverse order. For the experiments conducted in this study, the hyperparameters were set to be as follows: \(InitialLearnRate = 5e-4, MiniBatchSize = 4, MaxEpochs = 300, LearnRateDropFactor = 0.1.\). This model was implemented with the Matlab Deep Learning Toolbox.
# Implementation of the Transform Model Concept {#sec:Methodology .unnumbered}
## Experimental strategy and input data formatting
In this section the strategy to evaluate our Transform Model concept is presented. First we will assess the transformation of the HYCOM and MITgcm-GoM velocity fields by the in-situ observations. The training is conducted with the first 855 days and the transformation results are evaluated with the remaining 50 days of measurements. The purpose of this experiment is to assess the transformative potential of limited observation periods and its efficacy on assimilated and non-assimilated numerical model fields, respectively. Second, the long-term transformation capability of the concept is assessed by using HYCOM as virtual observations to transform the MITgcm-GoM velocity fields. In this case, the training took place over a period of three years and the testing period was either the preceding or following year of the training period. In order to proceed, both the model and the observations must be interpolated on the same spatial and temporal grids. Therefore the time interval was set to daily, MITgcm-GoM's and the coarsest temporal resolution of the three time series. Because the spatial resolution of the Dynloop data set over the geographical area of the measurements is \(29 \times 36\), \(64 \times 88\) for HYCOM and \(52 \times 70\) for the MITgcm-GoM simulations, all three data sets were interpolated to \(64 \times 64\) grid using a bicubic interpolation. Vertically, the same depth layers as in the Dynloop dataset were used to reinterpolate the other two model fields, which resulted in 26 layers of 20m interval from the surface to 500m (maximum depth used in this study). For the short-term Transform Model, the period March 2009-June 2011 was used for training and July-August 2011 for testing. For the long-term transformation experiment, the years 2009-2011 (2012-2010) were used for training and 2012 (2009) for testing in the forward (backward) experiment.
## Comparison between observed and modeled fields {#subsec:The input data and training}
To evaluate how different the three concurrent field times series are, we calculated the Taylor diagrams of the spatially averaged velocity magnitude for each of them at four depths, 0, 20, 100, and 500m respectively (Figure [\[fig:Taylor\]](#fig:Taylor){reference-type="ref" reference="fig:Taylor"}). Because the MITgcm-GoM is a free-run, the model velocity exhibits at all depths the lowest correlation coefficient (CC), largest RMSE and standard deviation. While the assimilated HYCOM model exhibits a lower standard deviation than the MITgcm-GoM, the RMSE between the two models is not significantly different at 0, 20 and 100m, except at 500m, where HYCOM underestimates it (Figure [\[fig:Taylor\]](#fig:Taylor){reference-type="ref" reference="fig:Taylor"}d). A large RMSE signifies that the amplitude of the variation of the velocity components is overestimated. HYCOM's CC is also lower at 500m than at shallower depth, suggesting that the data correction at 500m is less efficient than in shallower waters.
## Short-term Transform Model evaluation {#sec:ST-TM}
For this evaluation, two types of experiments were conducted. In the first one the transformation was applied only to the two-dimensional surface velocity and in the second experiment the transformation was applied to three-dimensional tensor of the velocity field. The efficacy of the transformation was evaluated by calculating the MSE between the transform model solution and the reference, which is the observations in this case, over the last 50 days of the observational period, not used in the training phase. The same calculation was done for the original numerical model velocity field.
### Two-dimensional field transformation
The results of the MSE for both the original numerical model and the transformed model fields are shown in Figure [\[fig:RMSE_0\]](#fig:RMSE_0){reference-type="ref" reference="fig:RMSE_0"} for both HYCOM and MITgcm-GoM. Figure [\[fig:RMSE_HYCOM\]](#fig:RMSE_HYCOM){reference-type="ref" reference="fig:RMSE_HYCOM"} shows a significant reduction of the HYCOM model velocity MSE, both in the mean and in the variability. The MSE of the transformed HYCOM field remained almost constant throughout the 50 day period, unlike the original HYCOM field. Figure [\[fig:RMSE_ECCO\]](#fig:RMSE_ECCO){reference-type="ref" reference="fig:RMSE_ECCO"} shows that the transformation of the MITgcm-GoM fields was as efficient as for the HYCOM fields but the MSE of the transformed fields increased over time, at a rate similar to the one of the original fields. HYCOM's fields' MSE, although quite variable does not exhibit an increasing trend over time like the MITgcm-GoM fields. This difference is most likely due to the fact that data were assimilated in the HYCOM simulation and not in the MITgcm-GoM model. The degree of transformation also called correction gain is shown in Table [1](#tab:correction){reference-type="ref" reference="tab:correction"} and is calculated as follows: \[Gain= \frac{ |M_{transformed}-M_{model}|} {M_{model}} * 100 \label{eq:Gain}\] where (\(M_{model}= \frac{\sum_{n=1}^{T} { MSE_1(n)}} T\)), (\(M_{transformed}= \frac{\sum_{n=1}^{T} { MSE_2(n)}} T\)) ,\(T\) is the length of the time series, \(MSE_1\) is the MSE between the model and the observed data, and the \(MSE_2\) is the MSE between the transformed and observed data. A high (low) gain signifies that the transformed model field errors are low (high). The transformation gain was higher for the HYCOM model than for the MITgcm-GoM field. However, the non-assimilated model transformation exhibited a MSE \< 1.5 \(cm.s^{-1}\) for up to 30 days, which is the less than the MSE of the assimilated model fields (Figure [\[fig:RMSE_ECCO\]](#fig:RMSE_ECCO){reference-type="ref" reference="fig:RMSE_ECCO"}). This result suggests that the correction of the flow field with in-situ observations by the Transform Model can be as effective as if not better than the data assimilation method currently used in the HYCOM operational forecast. This result is significant because the transformation was made outside of the observation period. In terms of the spatial structure of the flow field variability, the transformation is also able to carry through the spatial modes contained in the observed field as shown in Figure [\[fig:EOF_0\]](#fig:EOF_0){reference-type="ref" reference="fig:EOF_0"}. The efficacy of the transformation is stronger for the HYCOM than for the MITgcm-GoM field, which also shows that there is room for improvement in the data assimilation process, whether in the NCODA system or by applying our Transform Model.
### Three-dimensional field transformation
Surface observations are in general more commonly available than subsurface observations, in particular from three-dimensional arrays. Therefore, predicting subsurface dynamics has remained more challenging than at the ocean surface because of the lack of continuous sampling at the same location. Vertical projections of the surface corrections is still reliant on methods described in ,. The transformation of three-dimensional fields is therefore critical to full water column prediction. For the three-dimensional velocity field transformation, a Transform Model was created for each depth level between the numerical model and the observations and the transformation was applied to the last 50 days not included in the training. The results in terms of MSE and correlation coefficient (CC) for each depth level for the transformed HYCOM model over the 50-day period are shown in Figure [\[fig:HYCOM_RMSE_CC\]](#fig:HYCOM_RMSE_CC){reference-type="ref" reference="fig:HYCOM_RMSE_CC"}. As seen in the two-dimensional case, the transformation efficacy is significant. The MSE is reduced by a factor four from the surface down to 200m and by half at the 500m level (Figure [\[fig:HYCOMdepth_RMSE\]](#fig:HYCOMdepth_RMSE){reference-type="ref" reference="fig:HYCOMdepth_RMSE"}). The HYCOM field's MSE decreases with depth along with the gain of the transformation (Table [1](#tab:correction){reference-type="ref" reference="tab:correction"}). While the MSE of the HYCOM meridional velocity field is higher than the one of the zonal velocity field, when transformed, the resulting MSEs of each components are nearly identical. Similarly, the CC of the meridional field is much less than the CC of the zonal field across all levels (Figure [\[fig:HYCOMdepth_CC\]](#fig:HYCOMdepth_CC){reference-type="ref" reference="fig:HYCOMdepth_CC"}). The difference persists in the transformed fields but is strongly reduced. In addition the CC of the transformed fields is larger than the CC of the original fields and varies less vertically as well. The difference between zonal and meridional velocity CC is not surprising since the current meridional velocities dominate the flow pattern of the Loop Current. For the MITgcm-GoM three-dimensional transformed velocity fields the MSE is reduced by a factor three from the surface down to 200m and by half at the 500m level (Figure [\[fig:Depth_RMSE\]](#fig:Depth_RMSE){reference-type="ref" reference="fig:Depth_RMSE"}) although that reduction is less for the meridional velocity below 150m. For the CC, the MITgcm-GoM original model show very small correlation with the observations. The transformed fields however, shows an increase for the zonal velocity by a factor sixteen (Figure [\[fig:Depth_CC\]](#fig:Depth_CC){reference-type="ref" reference="fig:Depth_CC"}), relatively constant over depth, in contrast with the CC of the transformed meridional velocity, which is not as uniformely improved, although the correction can reach a factor fifteen near 100 and 350m. The 50-days gain is therefore less for the MITgcm-GoM overall as shown by Table [1](#tab:correction){reference-type="ref" reference="tab:correction"} but the correction factor is much higher for the latter than for HYCOM. The MSE of the transformed MITgcm-GoM fields (red lines Figure [\[fig:Depth_RMSE\]](#fig:Depth_RMSE){reference-type="ref" reference="fig:Depth_RMSE"}) is similar to the HYCOM's original MSE (blue lines in Figure [\[fig:HYCOMdepth_RMSE\]](#fig:HYCOMdepth_RMSE){reference-type="ref" reference="fig:HYCOMdepth_RMSE"}). From a correlation standpoint, there is a similar improvement, in particular for the zonal velocity (red lines in Figure [\[fig:Depth_CC\]](#fig:Depth_CC){reference-type="ref" reference="fig:Depth_CC"} vs blue lines in Figure [\[fig:HYCOMdepth_CC\]](#fig:HYCOMdepth_CC){reference-type="ref" reference="fig:HYCOMdepth_CC"}) and to a lesser degree for the meridional velocity. This shows that our method is at least as efficient at correcting the free-run model with just the observed velocity field as the NCODA is with all the different data sources used for HYCOM. The latter on the other hand requires less correction by the Transform Model, but the improvements by the Transform Model can still be significant, especially near the surface.
[\[tab:correction\]]{#tab:correction label="tab:correction"}
:::
While the statistical metrics such as MSE and CC may indicate a good agreement between the quantities compared, the spatial features that comprise the system may look dissimilar. Therefore, we show here the efficacy of the transformation in the three-dimensional structures of the transformed field on day 50 after the end of the observation period (Figure [\[fig:3DHYCOM_trans\]](#fig:3DHYCOM_trans){reference-type="ref" reference="fig:3DHYCOM_trans"}, [\[fig:ThreeD_ECCO\]](#fig:ThreeD_ECCO){reference-type="ref" reference="fig:ThreeD_ECCO"}). For the HYCOM model, the vertical and horizontal flow features in the observations are present in the transformed model field, showing similarities in magnitude and location in space. The presence of these features and their timing is critical to the development of the instabilities that lead to the growth of the frontal cyclones near the surface and then in the deep waters and ultimately to the separation of a Loop Current eddy, , ,. The three-dimensional transformation confirms the results seen in the two-dimensional transformation. The Transform Model is capable of extending the correction of observations in HYCOM at least 50 days past their the last measurement. For the MITgcm-GoM model, flow field difference is quite significant as shown by Figure [\[fig:ThreeD_ECCO\]](#fig:ThreeD_ECCO){reference-type="ref" reference="fig:ThreeD_ECCO"}. The transformed flow field is shifted toward the path of the observed flow, both horizontally and vertically, although the magnitude of the transformed flow is reduced from the observations.
## Long-term Transform Model Evaluation
For this evaluation HYCOM is used as virtual observations to transform the MITgcm-GoM surface velocity field. In the forward transformation experiment the transformation was applied for up to 365 days after the virtual observations ended. As seen in the short-term transformation, the MSE of the transformed field is reduced by up to a factor four and exhibits less temporal variability than the MITgcm-GoM MSE (Figure [\[fig:ECCO_RMSE_F\_P\]](#fig:ECCO_RMSE_F_P){reference-type="ref" reference="fig:ECCO_RMSE_F_P"}). The MSE of the transformed field varies over time along with the evolution of the Loop Current System, whose dynamics can affect the efficacy of the transformation. Indeed between days 120 and 160, the Loop Current evolution between the HYCOM and the MITgcm-GoM significantly diverged. While the eddy shed and reattached in the HYCOM flow, the Loop Current remained in its extended position in the MITgcm-GoM model. This event affected the capacity of the Transform Model to correct the divergent dynamics, which increased the MSE of the transformed model. Nonetheless, the transformation is able to strongly reduce the MITcgm-GoM velocity field MSE even 300 days after the virtual observation ended, which reveals the long lasting potential of the transformation process. The transformation is most efficient when the difference between models is associated with a spatial shift of the flow features (days 200-365 on Figure [\[fig:Future_RMSEM\]](#fig:Future_RMSEM){reference-type="ref" reference="fig:Future_RMSEM"}) and not a dynamical state difference. The backward transformation also shows a strong decrease in the MSE of the transformed MITgcm-GoM velocity fields for most part except during periods when the original numerical model MSE is also low. In this case, both MITgcm-GoM and HYCOM circulation features exhibit dynamical similarities. In addition the transformed MSE is less variable than the original one. Although the transformation could lead to unrealistic features after several months, the EOF analysis of the velocity field reveals that the modal structure of the virtual observations is present in the transformed field, both in the forward and backward transformations (Figure [\[fig:ECCO_EOF_F\_P\]](#fig:ECCO_EOF_F_P){reference-type="ref" reference="fig:ECCO_EOF_F_P"}). The transformation is also able to convey the timeliness of the separation of the Loop Current eddy (Figure [\[Future_Samples\]](#Future_Samples){reference-type="ref" reference="Future_Samples"}) in the forward transformation, which is paramount to the generation of reliable and useful predictions of the Loop Current. In the backward direction, the Transform Model is capable of readjusting the Loop Current state in the MITgcm-GoM to closely follow the state of the Loop Current in the virtual observations (Figure [\[Past_Samples\]](#Past_Samples){reference-type="ref" reference="Past_Samples"}).
# Conclusions {#sec:Conclusion .unnumbered}
Reliable and long-term ocean observations are difficult to obtain due to a variety of factors such as but not limited to costs, data collection manned or unmanned platform availability, instruments failure and durability, accessibility, limited energy supplies, data storage and extraction. In the current state of environmental predictions, data used for prognostic activities such as ocean or weather forecast are only used at the time of the forecast. They become part of history as soon as the forecast is done and are no longer relevant to the forecast. What if we could extend the period of relevance both forward and backward in time, in other words the usability of environmental observations? In this study we have presented a new method based on deep learning that enables the extension of the usability period of real-world observations up to one year. We have focused on ocean observations, but we believe that our approach could be applied to any biogeochemical and physical prediction exercise, as long as observations and model output are concurrently available. Our approach is based on the re-purpose of an existing deep learning model, called U-Net, designed specifically for image segmentation analysis in the biomedical field. We have modified the network in such a way that its original segmentation layer is replaced by, instead a regression layer. That layer is used to create a Transform Model that retains the temporal and spatial evolution of the differences between model and observations to produce a correction in the form of regression weights that evolves spatially and temporally with the model both forward and backward in time, beyond the observation period. In classical data-assimilation problems, the corrections that the observations make to the first-guess/background must lie in the space spanned by \(P\), the error covariance matrix. As such, \(P\) has been the subject of much research because of the central role it plays in data assimilation. For example, estimating the actual level of uncertainty of the first-guess is very difficult, and choosing a \(P\) that accurately reflects the inhomogeneity and anisotropic nature of the errors across the broad range of space-and time-scales that characterize the ocean is challenging, although innovative methods for multi-scale data-assimilation are being explored ,. In addition there is no efficient method to spread information forward or backward in time and to evolve the error covariance matrix in time, which can be attempted by using ensemble-based Khalman filters methods. Even though we haven't applied our approach to assimilate ocean data in numerical models, our method could be applied to the estimation of \(P\) while solving the time dependency, the spatial inhomogeneity and anisotropy and at the same time increase the usability of observations beyond their time measurement for up to maybe one year as shown in this study. It would also be applicable to the many other challenges faced by data-assimilation, associated with model and observations errors. The evolving relationship between the best model estimate and the main causes of errors including model and instrument errors could be captured by a Transform Model as described in this study without previous knowledge of absent physics or small scale processes not resolved by the model. Ultimately, one could envision that the first-guess estimate could be entirely realized with our Transform Model. Our transformation could be applied sequentially, at each time step of the prediction, which would reduce the magnitude of the increments and the shock induced to the numerical model, by keeping the model on a track closer to the observations than when no correction has been applied to the previous step as done in this work. While very few data-assimilation systems consistently assimilate sub-surface ocean current velocity data, the results shown in this study also reveal the efficacy of the Transform Model at correcting the three-dimensional flow field, even on existing reanalyzes, such as the HYCOM data used herein. Finally, with the increased availability of surface current data from high-frequency (HF) radars, a Transform Model could be easily created for any numerical model with just a few years of concurrent model data and observations. Such model, which can be updated daily with new data would provide useful corrections to the model surface flow field that would be transferred vertically to other variables, either by the numerical model itself or through a similar type of AI technique as the one described here. | {'timestamp': '2022-06-06T02:15:21', 'yymm': '2206', 'arxiv_id': '2206.01599', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01599'} |
# Introduction
Given examples and background knowledge (BK), the goal of inductive logic programming (ILP) is to find a set of rules (a logic program ) which with the BK correctly generalises the examples. The fundamental challenge is to efficiently search a large hypothesis space (the set of all programs) for a *solution* (a program that correctly generalises the examples). The two main approaches are *divide-and-conquer* (D&C) and *separate-and-conquer* (S&C). D&C approaches divide the examples into disjoint sets and search for a program for each set. S&C approaches search for a program that generalises a subset of the examples, separate these examples, and then search for more rules to add to the program to generalise the remaining examples. These two approaches support noise and learning programs with many rules and literals. However, as they only learn from a subset of the examples, they struggle to learn recursive programs. Moreover, these approaches offer no guarantees about the optimality of their solutions and can learn overly specific programs. Finally, these systems cannot perform predicate invention. Many modern approaches overcome these limitations through meta-level search and can learn optimal and recursive programs. However, most modern approaches struggle to learn programs with many literals in a rule, many rules in a program, or both. For instance, [ASPAL]{.smallcaps} precomputes every possible rule in a program and uses an answer set solver to find a subset of rules that generalises the examples. However, this precomputation rule selection approach does not scale to rules with more than a few literals. Likewise, many modern systems struggle to learn programs with many rules. In this paper, our goal is to overcome the limitations of modern systems yet maintain the ability to learn recursive and optimal programs and support predicate invention. The three key ideas are to (i) decompose programs into *non-separable* fragments, (ii) learn fragments separately, and (iii) *combine* fragments to learn programs with many rules and literals. A non-separable program cannot be decomposed into smaller parts. For instance, consider this program: \[\begin{array}{l} p_1: \left\{ \begin{array}{l} \emph{happy(A) \(\leftarrow\) rich(A)}\\ \emph{happy(A) \(\leftarrow\) friend(A,B), famous(B).}\\ \emph{happy(A) \(\leftarrow\) married(A,B), beautiful(B)}\\ \end{array} \right\} \end{array}\] This program has three rules that can be evaluated independently. In other words, if we learn the three rules separately and union their individual logical consequences, we have a program that is logically equivalent to \(p_1\). This program is, therefore, a *separable* program. By contrast, consider this program: \[\begin{array}{l} p_2: \left\{ \begin{array}{l} \emph{happy(A) \(\leftarrow\) rich(A)}\\ \emph{happy(A) \(\leftarrow\) married(A,B), happy(B)}\\ \end{array} \right\} \end{array}\] This program has two rules that cannot be evaluated independently as it has a recursive rule. In other words, if we learn the two rules separately and union their individual logical consequences, we have a program that is logically different to \(p_2\). It is, therefore, a *non-separable* program. By decomposing programs into smaller non-separable fragments, we want to make them easier to learn and thus reduce learning times. To explore these ideas, we build on the *learning from failures* (LFF) approach. LFF frames the ILP problem as a constraint satisfaction problem. The goal of an LFF learner, such as [Popper]{.smallcaps}, is to accumulate constraints in a generate, test, and constrain loop. In other words, [Popper]{.smallcaps} continually generates programs, tests them on examples, and, if they are not solutions, builds constraints to explain why they are not solutions to guide future program generation. To improve learning performance, we add the three new previously mentioned ideas to LFF. We describe our approach with an example.
### Motivating example {#motivating-example .unnumbered}
Suppose we have positive and negative examples of lists of numbers of arbitrary length. We want to learn a program to say whether an unseen list is good or bad. We might want to learn a program such as: \[\begin{array}{l} h_0: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,7)}\\ \emph{f(A) \(\leftarrow\) head(A,4), tail(A,B), head(B,4)}\\ \emph{f(A) \(\leftarrow\) head(A,23), tail(A,B), head(B,24)}\\ \emph{f(A) \(\leftarrow\) tail(A,B), f(B)} \end{array} \right\} \end{array}\] This program says that a list is good if it contains the sequence \[7\] or \[4,4\] or \[23,24\]. The last rule is important. It is a recursive rule that allows the program to generalise to lists of arbitrary length. To find a program that generalises the examples, we use a *generate*, *test*, *combine*, and *constrain* loop. In the generate stage, we generate progressively larger programs, i.e. programs with one literal, two literals, etc. Importantly, in this step, we only generate non-separable programs. For this example, we would start by generating programs such as \(h_1\) and \(h_2\): \[\begin{array}{l} h_1: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,1)}\\ \end{array} \right\}\\ h_2: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,5)}\\ \end{array} \right\} \end{array}\] In the test stage, we test each program on the examples. If a program entails a negative example, we build a *generalisation* constraint to eliminate more general programs from the hypothesis space as they will also entail the negative example. For instance, if we have the negative example *f(\[5,5,5\])* we would eliminate \(h_2\) from the hypothesis space and anything more general than it, such as: \[\begin{array}{l} h_3: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,5)}\\ \emph{f(A) \(\leftarrow\) head(A,7)}\\ \end{array} \right\} \end{array}\] If a program entails no negative examples, we build a *specialisation* constraint to eliminate more specific programs from the hypothesis space, as they will also not cover any positive example. For instance, if we do not have a positive example where the first element is 5, we prune all specialisations of \(h_2\), such as \[\begin{array}{l} h_4: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,5), tail(A,B), head(B,9)} \end{array} \right\} \end{array}\] A key novelty of our approach is the combine stage. If a program entails *at least one* positive and none of the negative examples, we add the program to a set of *candidate* programs. Candidate programs are non-separable programs that we want to combine. For instance, when considering programs of size 5, we might generate the recursive program: \[\begin{array}{l} h_5: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,7)}\\ \emph{f(A) \(\leftarrow\) tail(A,B),f(B)}\\ \end{array} \right\} \end{array}\] As \(h_5\) covers at least one positive example and no negative examples we deem it a candidate program. In the combine stage, we search for a combination of candidate programs that covers all the positive examples and that is minimal in size. If we do not find a combination, we go to the constrain stage, where we use the discovered generalisation and specialisation constraints to generate a new program. If we find a combination, we deem it the best solution so far. For instance, suppose that after considering programs of size 7 we see the programs: \[\begin{array}{l} h_6: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,4), tail(A,B), head(B,4)}\\ \emph{f(A) \(\leftarrow\) tail(A,B),f(B)}\\ \end{array} \right\}\\ h_7: \left\{ \begin{array}{l} \emph{f(A) \(\leftarrow\) head(A,23), tail(A,B), head(B,24)}\\ \emph{f(A) \(\leftarrow\) tail(A,B),f(B)}\\ \end{array} \right\} \end{array}\] Then the combination of \(h_5 \cup h_6 \cup h_7\) is \(h_0\) (the solution we want to learn). We have, therefore, learned a program with 4 rules and 13 literals by only considering programs with 2 rules and 7 literals. As we show in our experiments, this decomposition and recombination of programs allows us to learn programs much faster than existing approaches. As this point, we have has not yet proven that the combination is optimal in terms of program size. In other words, we have not yet proven that there is no smaller solution. Therefore, to avoid over-fitting, we continue to the constrain stage and add a constraint on the maximum program size in the generate stae (12 literals) for future iterations. We repeat this loop until we prove the optimality of a solution.
# Related work
**D&C.** [TILDE]{.smallcaps} is a D&C approach. [TILDE]{.smallcaps} extends decision tree learning algorithms to learn first-order, rather than propositional, rules. Although [TILDE]{.smallcaps} can learn large programs and scale to large datasets, it cannot learn recursive programs and struggles to learn from small numbers of examples. **S&C.** Progol is a S&C approach that has inspired many other approaches, notably [Aleph]{.smallcaps}. Progol employs a set covering algorithm to incrementally find programs that covers an uncovered example, removes the example from the set, and repeats on the remaining uncovered examples. Progol struggles to learn recursive and optimal programs and does not support predicate invention. Progol variants, such as [Aleph]{.smallcaps}, inherit the same limitations. **Rule selection.** Many systems systems formulate the ILP problem as a rule selection problem. ASPAL is one of the first rule selection approaches. ASPAL precomputes every possible rule in the hypothesis space. ASPAL uses an ASP solver to find a subset of the rules that covers all the positive but none of the negative examples. A major issue with this approach is scalability in terms of the size of rules. As it involves pre-computing every possible rule in the hypothesis space, the approach does not scale to rules with many body literals, or, more generally, to problems with many rules. For instance, the *coins-goal* task in our experiments would require precomputing at least \(10^{12}\) rules, which is clearly intractable. In contrast to these approaches, we do not precompute every possible rule. In addition, we do not search for combinations of individual rules. We instead search for combinations of *programs*, potentially formed of multiple rules, including recursive rules and invented predicates. **[Popper]{.smallcaps}.** Rather than precompute every possible rule, [Popper]{.smallcaps} lazily generates *programs*, potentially with multiple rules. The key idea of [Popper]{.smallcaps} is to discover constraints from smaller programs to rule out larger programs. [Popper]{.smallcaps} searches for a single solution for all the examples. [Popper]{.smallcaps} struggles to learn solutions with many rules and many literals. We go beyond [Popper]{.smallcaps} by (i) only generating non-separable programs in the generate step, and (ii) adding a combine step. As we experimentally demonstrate, our [Popper+]{.smallcaps} approach can drastically outperform [Popper]{.smallcaps} and other systems. **[DCC]{.smallcaps}.** The most similar work to our approach is the *divide, constrain, and conquer* ([DCC]{.smallcaps}) approach. [DCC]{.smallcaps} combines classical D&C search with modern constraint-driven ILP. Similar to D&C approaches, [DCC]{.smallcaps} learns a program for each example separately. To do so, [DCC]{.smallcaps} calls [Popper]{.smallcaps} with a subset of the examples. As these programs are likely to be overly specific, [DCC]{.smallcaps} iteratively tries to learn more general programs. The key idea behind [DCC]{.smallcaps} is to reuse knowledge (constraints) between iterations to improve learning performance. Our approach is completely different to [DCC]{.smallcaps}. [DCC]{.smallcaps} tries to generate a single program that covers all the examples, including separable programs and [DCC]{.smallcaps} has no combine stage, i.e. does not search for combinations of programs. By contrast, [Popper+]{.smallcaps} never generates a separable program. Moreover, it searches for combinations of programs in the combine stage. As we experimentally show, [Popper+]{.smallcaps} can drastically outperform [DCC]{.smallcaps}, especially in terms of learning times.
# Problem Setting {#sec:setting}
We use the *learning from failures* (LFF) setting. LFF uses *hypothesis constraints* to restrict the hypothesis space. Let \(\mathcal{L}\) be a language that defines hypotheses, i.e. a meta-language. For instance, consider a meta-language formed of two literals *h_lit/3* and *b_lit/3* which represent *head* and *body* literals respectively. With this language, we can denote the rule *last(A,B) \(\leftarrow\) tail(A,C), head(C,B)* as the set of literals *{h_lit(0,last,(0,1)), b_lit(0,tail,(0,2)), b_lit(0,head,(2,1))}*. The first argument of each literal is the rule index, the second is the predicate symbol, and the third is the literal variables, where *0* represents *A*, *1* represents *B*, etc. A *hypothesis constraint* is a constraint (a headless rule) expressed in \(\mathcal{L}\). Let \(C\) be a set of hypothesis constraints written in a language \(\mathcal{L}\). A set of definite clauses \(H\) is *consistent* with \(C\) if, when written in \(\mathcal{L}\), \(H\) does not violate any constraint in \(C\). For instance, the rule *last(A,B) \(\leftarrow\) last(B,A)* violates the constraint *\(\leftarrow\) h_lit(0,last,(0,1)), b_lit(0,last,(1,0))*. We denote as \(\mathcal{H}_{C}\) the subset of the hypothesis space \(\mathcal{H}\) which does not violate any constraint in \(C\). We define the LFF problem:
We define a LFF solution:
If a hypothesis is not a solution then it is a *failure*. A hypothesis is *incomplete* when \(\exists e \in E^+, \; H \cup B \not \models e\). A hypothesis is *inconsistent* when \(\exists e \in E^-, \; H \cup B \models e\). A hypothesis is *partially complete* when \(\exists e \in E^+, \; H \cup B \models e\). A hypothesis is *totally incomplete* when \(\forall e \in E^+, \; H \cup B \not \models e\). Let \(cost: \mathcal{H} \mapsto R\) be an arbitrary cost function that measures the cost of a hypothesis. We define an *optimal* solution:
In this paper, our cost function is the number of literals in the hypothesis \(H\).
#### Constraints.
The goal of an LFF learner is to learn hypothesis constraints from failed hypotheses. introduce hypothesis constraints based on subsumption. A clause \(C_1\) *subsumes* a clause \(C_2\) (\(C_1 \preceq C_2\)) if and only if there exists a substitution \(\theta\) such that \(C_1\theta \subseteq C_2\). A clausal theory \(T_1\) subsumes a clausal theory \(T_2\) (\(T_1 \preceq T_2\)) if and only if \(\forall C_2 \in T_2, \exists C_1 \in T_1\) such that \(C_1\) subsumes \(C_2\). A clausal theory \(T_1\) is a *specialisation* of a clausal theory \(T_2\) if and only if \(T_2 \preceq T_1\). A clausal theory \(T_1\) is a *generalisation* of a clausal theory \(T_2\) if and only if \(T_1 \preceq T_2\). If a hypothesis \(H\) is incomplete, a *specialisation* constraint prunes specialisations of \(H\), as they are guaranteed to also be incomplete. If a hypothesis \(H\) is inconsistent, a *generalisation* constraint prunes generalisations of \(H\), as they are guaranteed to be inconsistent as well.
# [Popper+]{.smallcaps} {#sec:impl}
We now describe our [Popper+]{.smallcaps} algorithm. We first briefly describe [Popper]{.smallcaps}.
## [Popper]{.smallcaps} {#popper}
Algorithm [\[alg:popper\]](#alg:popper){reference-type="ref" reference="alg:popper"} shows the [Popper]{.smallcaps} algorithm, which solves the LFF problem (Definition [\[def:probin\]](#def:probin){reference-type="ref" reference="def:probin"}). [Popper]{.smallcaps} takes as input background knowledge (), positive () and negative () examples, and an upper () bound on hypothesis sizes. [Popper]{.smallcaps} uses a *generate*, *test*, and *constrain* loop to find an optimal solution. [Popper]{.smallcaps} starts with a ASP program \(\mathcal{P}\) (hidden in the generate function). The models of \(\mathcal{P}\) correspond to hypotheses (definite programs). In the generate stage (line 5), [Popper]{.smallcaps} uses Clingo, an ASP system, to search for a model of \(\mathcal{P}\). If there is no model, [Popper]{.smallcaps} increments the hypothesis size (line 7) and loops again. If there is a model, [Popper]{.smallcaps} converts it to a hypothesis (). In the test stage (line 9), [Popper]{.smallcaps} tests the hypothesis on the training examples. If the hypothesis is a solution (is complete and consistent), [Popper]{.smallcaps} returns it. If the hypothesis fails (is incomplete or inconsistent), then, in the constrain stage (line 12), [Popper]{.smallcaps} builds hypothesis constraints (represented as ASP constraints) from the failure. [Popper]{.smallcaps} adds these constraints to \(\mathcal{P}\) to prune models and thus reduce the hypothesis space. For instance, if a hypothesis is incomplete (does not entail all the positive examples), [Popper]{.smallcaps} builds a specialisation constraint to prune hypotheses that are logically more specific. If a hypothesis is inconsistent, [Popper]{.smallcaps} builds a generalisation constraint to prune hypotheses that are logically more general. [Popper]{.smallcaps} repeats this loop until it (i) finds an optimal solution, or (ii) there are no more hypotheses to test.
This program is incomplete and inconsistent. However, we do not want to prune specialisations of it as doing so would eliminate the following program from the hypothesis space:
Therefore, [Popper+]{.smallcaps} only builds a specialisation constraint when a program is either (i) totally incomplete, or (ii) consistent. If a program is partially complete and consistent, [Popper+]{.smallcaps} adds the program to a set of candidate programs (line 12). In the combine stage, [Popper+]{.smallcaps} searches for combinations of programs that cover the positive examples. If there is such a combination, [Popper+]{.smallcaps} updates its maximum program size and continues to search to prove optimality. For instance, suppose that the combination of two programs each with 3 literals is a solution. Then there is no point considering programs with more than 5 literals, as we already have a solution with 6 literals. [Popper+]{.smallcaps} repeats this loop until either it (i) finds an optimal solution, or (ii) there are no more hypotheses to test.
#### Combiner.
To find combinations of programs, we follow ASPAL and use an ASP encoding. We give each positive example a unique example id. Unlike ASPAL, which only assigns ids to individual rules, we assign each candidate program (potentially formed of multiple rules) a unique program id. We create an ASP choice rule (a decision variable in standard constraint terminology) that determines whether a program should be in a solution. We add facts about the coverage and size of each candidate program. We then ask the ASP solver to find a combination of programs that covers as many positive examples as possible. We also ask the ASP solver to find the smallest combination, as to learn the optimal program.
# Experiments {#sec:exp}
The evaluate the performance of [Popper+]{.smallcaps}, our experiments aim to answer the question:
Q1
: How does [Popper+]{.smallcaps} compare against other approaches?
To answer **Q1** we compare [Popper+]{.smallcaps} against [Popper]{.smallcaps}, [DCC]{.smallcaps}, and [Aleph]{.smallcaps}. A key motivation for introducing [Popper+]{.smallcaps} is to reduce learning times, especially the time required to provably learn optimal programs. Our experiments, therefore, aim to answer the question:
Q2
: How well does [Popper+]{.smallcaps} perform given more learning time?
To answer **Q2**, we measure the learning performance of [Popper+]{.smallcaps} and other systems when given progressively longer timeouts (maximum learning times). In other words, we ask each system to return the best solution found in the time given.
#### Methods.
We measure predictive accuracy and learning time given a maximum learning time. We use two maximum learning times: **60 seconds** and **10 minutes**. We repeat all the experiments 10 times and measure the mean and standard deviation. We use a 3.8 GHz 8-Core Intel Core i7 with 32GB of ram. All the systems use a single CPU.
#### Systems. {#sec:systems}
We compare [Popper+]{.smallcaps} against [Popper]{.smallcaps}, [DCC]{.smallcaps}, and [Aleph]{.smallcaps}[^2]. We use these systems because they can learn recursive definite programs. [DCC]{.smallcaps}, [Popper]{.smallcaps}, and [Popper+]{.smallcaps} use identical biases so the comparison between them is fair. [Aleph]{.smallcaps} uses a similar bias to [Popper+]{.smallcaps} but requires additional settings, such as an upper bound on the nodes to be explored when searching for a rule. We have tried to make a fair comparison but there will likely be different parameters that would improve the performance of [Aleph]{.smallcaps}.
## Experimental Domains
We use the following domains. More details are in the appendix. **Michalski Trains.** The Michalski trains problem is to find a hypothesis that distinguishes eastbound trains from westbound trains. We use this domain because we can easily generate progressively more difficult tasks to test the scalability of the approaches as the solution size grows. We consider four tasks of increasing complexity, where the subscript denotes the number of rules in the solution. There are 1000 examples but the distribution of positive and negative examples is different for each task. We randomly sample the examples and split them into 80/20 train/test partitions **IGGP.** In *inductive general game playing* (IGGP) agents are given game traces from the general game playing competition. The task is to induce a set of rules that could have produced these traces. We use four IGGP games: *minimal decay* (md), *rock, paper, scissors* (rps), *buttons*, and *coins*. We learn the *next* relation for each game. We also learn the *goal* relation for *buttons* and *coins*. These tasks are non-trivial for ILP systems as the search spaces can be large. We exclude the goal relations for *md* and *rps* and they require trivial two literal solutions. These tasks are non-trivial for ILP systems as the search spaces can be large. For instance, the *coins-goal* problem has 130 background relations and requires a solution with multiple rules and many literals. **Program synthesis.** Inducing complex recursive programs has long been considered a difficult problem and most ILP systems cannot learn recursive programs. We use the program synthesis dataset introduced by augmented with two new tasks *contains* and *reverse*.
## Experimental Results
Table [1](#tab:q1-60){reference-type="ref" reference="tab:q1-60"} shows the predictive accuracies of the solutions returned by the systems when given a maximum learning time of **60 seconds**. The results show that [Popper+]{.smallcaps} generally outperforms the other systems in terms of predictive accuracy. [Popper+]{.smallcaps} comfortably outperforms [Popper]{.smallcaps} and [DCC]{.smallcaps}, especially on the trains and IGGP tasks. For instance, for the *buttons* task, [Popper+]{.smallcaps} achieves 100% accuracy on the test set, whereas [Popper]{.smallcaps} and [DCC]{.smallcaps} only achieve 19%. [Aleph]{.smallcaps} performs reasonably well on the trains and IGGP tasks but struggles on the program synthesis tasks. This performance is expected as [Aleph]{.smallcaps}, as with all bottom clause approaches, struggle to learn recursive programs. [Popper+]{.smallcaps}, [Popper]{.smallcaps}, and [DCC]{.smallcaps} are all guaranteed to learn optimal programs given sufficient time, assuming one exists. Therefore, when these systems have low predictive accuracy, it is because they struggled to find a solution in the given time limit and thus return an empty or partial program--this explanation does not hold for [Aleph]{.smallcaps}, which always terminates in the given learning time, but tends to learn overly specific programs. For instance, the *buttons* tasks requires learning a solution with 10 rules and 61 literals, which is beyond [DCC]{.smallcaps} and [Popper]{.smallcaps}. This explanation is clearer when looking at the predictive accuracies of the solutions returned by the systems when given a maximum learning time of **10 minutes**, which are shown in Table [2](#tab:q1-600){reference-type="ref" reference="tab:q1-600"}. With this larger timeout, all three systems have higher predictive accuracy, as does [Aleph]{.smallcaps}. However, even with this larger timeout, [Popper+]{.smallcaps} is still the best performing system. The systems have large differences in learning times, which are shown in Table [3](#tab:q1times){reference-type="ref" reference="tab:q1times"}. The main difference is that [Popper+]{.smallcaps} provably learns the optimal solution for most tasks in under 90 seconds. By contrast, for many tasks, [DCC]{.smallcaps} and [Popper]{.smallcaps} do not terminate within 10 minutes. For instance, [DCC]{.smallcaps} and [Popper]{.smallcaps} do not find a solution for *buttons-goal* given 10 minutes The reasons for this drastic improvement in learning time are because [Popper+]{.smallcaps} (i) only generates non-separable programs, and (ii) then tries to combine them. The results of the *buttons-goal* task need explaining. [Popper+]{.smallcaps} is guaranteed to find the smallest program that covers all the positive and none of the negative examples. On the *buttons-goal* task, [Popper+]{.smallcaps} finds such a program in less than 60 seconds. However, for this task, [Aleph]{.smallcaps} outperforms [Popper+]{.smallcaps} in terms of predictive accuracy (100% vs 74%). The reason is that [Aleph]{.smallcaps} learns a very specific program that includes memorising certain examples. In other words, [Aleph]{.smallcaps} includes positive examples in the solution. For this task, memorising the training examples is suitable for high predictive accuracy on the test examples. [Popper+]{.smallcaps}, by contrast, learns a more general solution, i.e. without memorising the examples, but which does not generalise as well to the test data. Overall, the results from this experiment show that [Popper+]{.smallcaps} can outperform existing systems both in terms of predictive accuracies and learning times.
# Conclusions and Limitations
We have introduced an ILP approach that employs a *generate*, *test*, *combine*, and *constrain* loop. The three key ideas are to (i) decompose programs into *non-separable fragments*, (ii) learn fragments separately, and (iii) *combine* fragments to learn programs with many rules and literals. We have implemented our approach in [Popper+]{.smallcaps}, a new ILP system. [Popper+]{.smallcaps} can learn optimal, recursive, and large programs and perform predicate invention. Our experiments on three domains (classification, inductive general game playing, and program synthesis) show that our approach can improve predictive accuracies and reduce learning times compared to other ILP systems. The drastic improvements we have shown in learning performance should directly transfer to problems in which LFF has already been applied, such as learning to explain Rubix cubes and learning higher-order programs.
## Limitations and Future Work
#### Noise.
In the combine stage, [Popper+]{.smallcaps} searches for a combination of candidate programs that covers all the positive and none of the negative examples and that is minimal in size. To perform this search, we use Clingo's optimisation feature. This approach allows us to handle misclassified positive examples, i.e. false negatives. In future work, we want to extend the approach to handle negative examples, which early work has already explored. | {'timestamp': '2022-06-06T02:15:48', 'yymm': '2206', 'arxiv_id': '2206.01614', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01614'} |
## Homotopical smoothness and generalized purity
## Duality with compact support and the fundamental class of the diagonal
If \(f\) is smooth, the above commutative diagram shows that under the canonical isomorphism \[\begin{array}{rcl} [\htp_S(X),\htp^c_S(X)] & \simeq & [\htp_S(X) \otimes \htp_S(X,-\tau_f),\un_S] \\ & \simeq & [X \times_S X,p_1^*(\tau_f)] \\ & \equiv & \Pi^{2d,d}(X \times_S X,p_1^*(\tau_f)) \end{array}\] the map \(\alpha'_f\) can be computed as the fundamental class of the diagonal \[\eta_{X \times_S X}(\Delta_{X/S})=\delta_*(1)\]
implies the following more potent form of duality We will give concrete examples in .
## Stable motivic homotopy type at infinity of hyperplane arrangements {#subsection:smhtaioha}
Owing to and , we deduce the following computation of stable motivic homotopy types at infinity.
[^1]: Recall that if \(S=\Spec(k)\) where \(k\) is a field of characteristic \(0\), then implies that for any separated \(k\)-scheme \(X\) of finite type, \(\Pi_k(X,v)\) is rigid since it is constructible, see. The same holds in \(\SH(k)[1/p]\) if \(k\) has positive characteristic \(p\) using (see also ). | {'timestamp': '2022-06-06T02:14:26', 'yymm': '2206', 'arxiv_id': '2206.01564', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01564'} |
# Introduction
Neurorehabilitation is a widely used medical practice that aims to aid recovery from a nervous system injury. Its purpose is to maximize and maintain the subject's motor control while trying to restore motor functions in people with neurological impairments. Robot-assisted training has been widely investigated as an effective neurorehabilitation approach that helps augment physical therapy and facilitates motor recovery. According to the literature , such approaches can help therapists save time and effort while reproducing accurate and repetitive motions and delivering high-intensity training. In particular, upper-limb robotic rehabilitation is one of the fastest-growing areas in modern neurorehabilitation. For this approach, the authors chose to exploit the PlanArm2 prototype , a planar robotic device designed for domestic upper-limb neurorehabilitation. The system is envisioned to be used by patients at home after an initial calibration phase supervised by professionals. However, a crucial issue for rehabilitation training is user engagement and motivation , which may be lacking if the rehabilitation system is used without a human medical coach. In these terms, we believe that introducing socially interactive agents could represent a valid solution to restore and augment the social aspects typical of an in-person rehabilitation session. Moreover, inheriting results typical of social robotics, the enhancement of a rehabilitation system through physiological and social signals awareness may augment the system personalization level and further facilitate neuroplasticity. Therefore, the capability of a neurorehabilitation training system to model the patient's state and tune its behavior according to both the measured performance and the inferred social and physiological state could improve the engagement of the user and the outcome of the therapy. This paper investigates a technologically-supported upper-limb neurorehabilitation approach by introducing social interaction capabilities in a heterogeneous system composed of 1) a planar rehabilitation device for physical assistance interaction and 2) two socially interactive virtual agents in the role of a coaching assistant, one demanding, one supporting. In this context, physiological and social signals of the participant are analyzed and elaborated into attentiveness, stress, engagement, and pain information, which are then exploited to tune the rehabilitation session coherently.
# Background and Related Work
## Neurorehabilitation Best Practice {#sec:bpnt}
The constantly growing employment of technological devices in neurorehabilitation therapy can be explained by the introduced ease in reaching a significant number of movement repetitions in a chosen district and in obtaining higher patient engagement, fundamental aspects in the motor rehabilitation process. However, to do so, one must be able to reconcile movement repetition, often leading to boredom, with variability and incentive to enhance the patient's attention and commitment. This section reports \"neurorehabilitation best practices\" we collected from experienced physiotherapists how they manage these issues. Based on these, a system concept is proposed. During every therapy session, a therapist assesses the actual level of attention, commitment, engagement, and stress or pain currently experienced by the patient to deliver the correct amount of exercise and avoid the risk of too easy or too difficult exercises that may lead to a decay in interest or even frustration. In particular, the therapist monitoring the therapy session relies on the activity scores and the subject's behavior to reach this goal. For example, if the patient cannot achieve a particular performance, the selected activity is likely too difficult. On the other hand, if the patient can perform the exercise but, after some time, becomes very talkative and less performing, it is likely that decay in interest is being experienced. On this basis, the therapist must give feedback, if needed, support when the activities are too difficult or change exercise when the attention starts to decrease. Furthermore, considering patients affected by neurological impairments, attention problems are frequent. Recording the period to which the attention could last can be useful information to provide a correct dosage of exercise. Lastly, the management of neurological disorders can be considerably different between adults and children. In both cases, understanding when a pause or a change of exercise is necessary is crucial. However, considering adults, one can count on their responsibility to train towards an improvement, even if the activity could lead to boredom. On the other hand, children may not behave in the same way and, to augment their engagement, it is crucial to introduce gaming aspects to the exercise.
## Socially Interactive Agents as Medical Coaches {#sec:agents4medicaltraining}
As a use case for interactive social agents, technologically supported health care has been researched for about 15 years. One of the early systems is the Fit Track system with the relational agent Laura. Laura has the role of an exercise advisor that interacts with patients daily for one month to motivate them to exercise more. A more recent system employs the socially interactive agent Gloria for stress management training system using biofeedback. The training system is designed to run autonomously without a medical coach present during the training. The authors conducted an expert interview and a user study in which the novel approach was compared to the stress management training using stress diaries. The actual training combines two methods of feedback: 1) situation awareness and 2) social interaction feedback with an awareness tool and a socially interactive agent. This approach draws on theories of embodied learning, which proclaims an interaction between bodily and cognitive functions and foresee their concurrent stimulation for more intuitive, long-lasting, and transferable learning. In the study, the awareness tool helped the user to track their sensory-motor reactions to stressful situations, increasing situation (introspective) awareness. The agent first created a social background to embed this awareness. After a first awareness training, the awareness tool is faded out, and the social agent overtakes the feedback. This implements prominent theories on the socio-emotional routes of learning and implicit emotion regulation in particular. In the current context, we are looking into embedding the cognitive understanding of the task in a socio-emotional context to increase the learning effect of the neurorehabilitation training. The results of a study revealed, HRV biofeedback training supported by a socially interactive agent tends to lower stress levels and improve adaptability to stressful situations. For the presented neurorehabilitation approach, we rely on the Gloria biofeedback training system. However, the behavior and the training strategies are adapted to the specific needs of neurorehabilitation training.
## Robotic Neurorehabilitation
In the last two decades, robotic devices for neurorehabilitation have been widely investigated, developed, and introduced in the market to offer a valid alternative to conventional therapy and fill the constantly growing gap between supply and demand, both for upper limb and lower limb rehabilitation . The goal of control algorithms for robotic therapy is to control robotic devices designed for rehabilitation exercises so that the participant's selected exercises provoke motor plasticity and improve motor recovery. Thus, the capability of a robot to adapt the level of assistance according to the skill and the performance of the patient is one of the most crucial features . Furthermore, assistance is sometimes obtained not only from the biomechanical point of view but also evaluating physiological and psychological aspects , for a more fine-grained assessment of the patient's state. Visual feedback, sometimes including immersive and augmented reality, is an almost indispensable element for robotic rehabilitation devices, to administer therapeutic exercises to increase user engagement, to propose visually-augmented exercises, even resembling Activities of Daily Living (ADLs). Enriched communication, including virtual coaching, could increase the effect of personalized and motivating communication. This approach is more and more investigated and exploited in medical care, but its exploitation is still limited in the rehabilitation field .
## Social and Physiological Signals
The level of attention a user pays to a task is usually studied in fields such as education (e.g., measuring student attentiveness during a lecture), driver assistance (e.g., monitoring attentiveness of driver to prevent accidents), etc. As noted in, attention is a key modulator of neuroplasticity, and information regarding the attentiveness of the patient can be important for the rehabilitation strategy. demonstrated that a driver's attention state could be determined through the head and facial features. These features were extracted from a video sequence that was captured through a single camera. proposed a student attention monitoring system that uses facial and body features extracted from MS Kinect 2D and 3D data. Similar to these studies, we rely on facial and body features to predict the attention level of the patient. The patient may experience stress while performing the exercise, especially if they find the exercise hard or cannot complete the recommended exercise. Many studies have studied stress-induced in different scenarios like public speaking, mental arithmetic tests, driving on challenging routes, hyperventilation, etc. These studies found that features extracted from ECG data like Heart Rate Variability (HRV) and EDA data like Skin Conductance Response (SCR) are good stress indicators. Pain assessment and management are critical for a variety of illnesses and medical interventions. Recent efforts in affective computing suggest that automatic detection of pain from facial expression is a feasible goal. In this work, we adopt a deep learning method to develop a model to detect pain from facial expressions. Typically, this requires a large amount of training data. Still, the availability of pain data is limited due to patient contact, privacy concerns, and the adherence to strict ethical guidelines. However, previous studies like have demonstrated that a transfer learning approach can be adopted to develop pain recognition models that perform well, despite the limited training data.
# Concepts
The envisioned Empathetic Neurorehabilitation Trainer architecture is reported in Fig. [\[fig:deployment\]](#fig:deployment){reference-type="ref" reference="fig:deployment"}. The goal of the proposed structure is to provide the patient with socially aware neurorehabilitation therapy. For this purpose, the system is equipped with a rehabilitation-specific robotic device and a virtual socially interactive agent (Fig. [\[fig:system_concept\]](#fig:system_concept){reference-type="ref" reference="fig:system_concept"}). Both device and agent can adapt their behavior based on the patient's performance, as in most assistance-as-needed paradigms, and take into account the patient's physiological and social state. Moreover, these two socially interactive agents, together with the specific task to be carried out and the feedback media chosen to provide the patient with an explanation regarding the exercise, are intended to work as a single entity, actively collaborating to improve the rehabilitation session outcomes further. All software and hardware components are realized as nodes in a ROS framework. There, a closed-loop monitoring of a set of heterogeneous parameters is introduced. In fact, during the execution of the task, the robotic device is in charge of collecting data regarding the kinematics of the patient's movement (e.g., position, speed), and a wearable device is used to extract a series of physiological values (e.g., ECG, EDA). At the same time, a camera captures the patient's upper body for social signal interpretation purposes. These raw data represent the input for a signal interpretation module (Fig. [\[fig:deployment\]](#fig:deployment){reference-type="ref" reference="fig:deployment"}, left bottom), responsible for providing a series of higher-level quantities such as patient's performance, attentiveness, stress, amusement, and pain. As depicted (Fig. [\[fig:deployment\]](#fig:deployment){reference-type="ref" reference="fig:deployment"}, right bottom), a human physiotherapist has a central role in the proposed approach. In fact, professional expertise is required for the patient's initial assessment (e.g., residual mobility, attention span), used to define the backbone of a model for both the patient, the agent and a selection of suitable exercises. Moreover, a supervised machine learning module (Fig. [\[fig:deployment\]](#fig:deployment){reference-type="ref" reference="fig:deployment"}, center bottom) is employed to learn from the physiotherapist how to optimally balance the target execution performance for the exercise and the social experience for the specific patient. In fact, a certain level of stress is beneficial for the therapy, and therefore the system is supposed to tune the exercise difficulty accordingly. Also, both challenging and entertaining portions of the session are be included to maximize the patient's attention. Closing the loop, a Therapy Manager (Fig. [\[fig:deployment\]](#fig:deployment){reference-type="ref" reference="fig:deployment"}, center top) actively exploits the inferred information to decide how the behavior of the socially interactive agent should be changed, which explanations (cf. awareness) should be given, and which exercise and difficulty level should be activated to optimize the therapy experience and effectiveness. The Therapy Manager is a software framework for authoring, orchestrating, and executing scenario content with task specifications. This component will be implemented using the Visual SceneMaker (VSM) tool. In addition, VSM provides dedicated plugins for all other components using ROS communication protocol.
## Interactive Agent-based Training and Explanations
We employ two socially interactive agents for the proposed agent-based training, one supportive agent and one demanding agent. Both take the role of a coaching assistant. According to the patient's level of training, the agent could be chosen. The supportive agent aims to enhance the patient's motivation by creating a warm, friendly, and motivating atmosphere during the training sessions. The agent signals comprehension by head nods and brief verbal utterances, such as\" OK\" to create a pleasant atmosphere for the patient. The patient is also encouraged through the use of positive feedback, such as smiles. On the other hand, we have a demanding agent that appears when the patient is seen to be losing attention or interest (Sec. [3.3](#sec:ssi){reference-type="ref" reference="sec:ssi"}). Initially, the patient starts the training session with a good performance, but their performance starts deteriorating over time. When such a situation occurs, the demanding agent appears, where the patient is provided with activities with a higher level of complexity. There, the demanding agent exhibits a gaze behavior that is supposed to be perceived as dominant and shows fewer head tilts while speaking and listening. Based on the described concept of assessing the affective states such as attentiveness, stress, and pain (Sec. [3.3](#sec:ssi){reference-type="ref" reference="sec:ssi"}), the agent adapts its motivational strategy. Here, we are considering the possibility of having two agents. One agent supports the patient in the case of actual difficulty during the training. The other shows a more demanding behavior and can be used if the patient is losing attention or interest, which can be detected with the help of social and physiological signals. If the exchange of them during the training exercises supports the training regarding, for example, efficiency, it is part of future investigations. The coaching assistant follows the best practices of training professionals (Sec. [2.1](#sec:bpnt){reference-type="ref" reference="sec:bpnt"}). The training always starts with a pre-screening phase. This phase serves for the identification of patient issues and the selection of further training sessions. In the pre-screening session, the physiotherapist observes how the patient interacts with the agent and the device and what the patient can achieve. In addition, the movement space is personalized, e.g., space is increased if a noticeable improvement in the patient's movement is detected. Since every patient is different from the other, they might face difficulties in different areas, e.g., some patients may find it difficult to control the smoothness of movement. Others may have difficulties in gripping. The exercises selected for all patient resembles actual day-to-day useful movements. During the training sessions, time and difficulty are tailored automatically. The system adapts to the level of attention or stress. For example, if the level of attention of the patient decreases, more competitive or challenging exercises/tasks cannot be provided. Strategies to overcome this phase to suggest a training pause or offer playful intermezzos. The particular realization always follows best practices (Sec. [2.1](#sec:bpnt){reference-type="ref" reference="sec:bpnt"}), and known approaches (Sec. [2.2](#sec:agents4medicaltraining){reference-type="ref" reference="sec:agents4medicaltraining"}) are part of future investigations. Content-wise, each session is represented by small training fragments. By combining such fragments, the difficulty and time of a training session can be tuned. In addition, each session is followed by some relaxation period, which could involve playful, less stressful activities.
## Neurorehabilitation Training Device
This work exploits the PlanArm2 rehabilitation device , a 2-DOF planar robotic prototype validated for its active, passive, and assistive upper-limb rehabilitation capabilities. The selected device has been developed specifically to achieve an acceptable compromise regarding workspace symmetry concerning the sagittal plane, relatively large workspace, portability, and affordability, rendering it suitable for domestic training sessions. Furthermore, from a long-term point of view, the system could be autonomously used by patients directly at home after an initial calibration phase supervised by professionals. In these terms, the presence of a socially interactive agent could provide the missing social interaction and improve the patient's experience while still guaranteeing an effective training session. However, not all cases are suited for autonomous rehabilitation, which stresses the importance of the patient's initial assessment performed by a specialist. Besides the agent, the device itself is supposed to adapt in a social way. Therefore, the envisioned assistive control logic exposes a series of parameters intended to be automatically tuned not only based on the patient's performance, as in most assist-as-needed paradigms, but also considering the patient's physiological and social inferred state. An example of an adjustable parameter is the target speed of execution. By increasing this value, different effects can be expected: in some cases, the exercise may become too stressful and might generate frustration or pain in the patient. In other cases, it could render the activity more challenging and augment the patient's engagement. By monitoring performance, social and physiological signals, the system should be able to tune the parameter correctly, also based on the therapist's experience embedded in the Supervised Machine Learning (Fig. [\[fig:deployment\]](#fig:deployment){reference-type="ref" reference="fig:deployment"}, center bottom) module. Similarly, a series of additional parameters are available for adaption in the context of the proposed assistive rehabilitation controller. For instance, given an ideal target trajectory to be followed by the patient, the level of regular and tangential assistance force provided can be independently tuned to the patient's needs. By regulating these assistance parameters, one could range from entirely passive to completely active rehabilitation paradigms, meaning that the same device could be used for a wide range of patients after a therapist has properly tuned and trained the system.
## Signal Interpretation {#sec:ssi}
During the training sessions, the social signals from the patients can be used to infer useful information about their experience. We identified the following emotional states that would serve as an input to the agent's motivational strategy.
### Attentiveness
The patient may get distracted during the training session because of boredom, lack of motivation. Motivation and attention can be critical modulators of neuroplasticity, contributing to determining the actual outcome of a rehabilitation therapy . Therefore, the level of attention of the patient is an important input to the agent's motivational strategy in neurorehabilitation. Previous works have demonstrated that attentiveness can be predicted through physiological signals like EEG and facial and body pose features like gaze direction, head orientation, body posture. Deep learning models are trained to focus on facial and body features that can be extracted using the camera.
### Stress and amusement
These affective states are crucial as we introduce some elements of gamification to the training session. On the one hand, the patient may be stressed, for example, when the exercise is too hard. On the other hand, the patient may enjoy the gamification and experience flow. The agent's behavior would differ depending on which state is detected. Physiological signals such as ECG and EDA have been demonstrated to be very effective in detecting stress. The WESAD data set is suitable for our use case as it contains multiple physiological signals from wearable sensors and is annotated for the states we intent to classify (stress, amusement, neutral). We adapt the neural network architecture proposed in to train an LSTM (Long short-term memory) network to detect stress and amusement.
### Pain
Pain is an important social component, as the expression of pain triggers social reactions such as empathy and care. Many health-care related fields are deploying image or video based automatic pain detection. Our idea is to detect pain from the facial expressions of the patient captured by the front camera. To achieve this, we train a deep learning model that can discern pain and no-pain images. We use the images from UNBC-McMaster shoulder pain expression database to train our model. This data set contains images of 25 participants suffering from shoulder pain while performing a range of motions with their affected and unaffected limb-a scenario similar to our use case. The small number of unique samples in this data set would affect the performance of our model. Similar to, we use a transfer learning approach to tackle this. To ensure that our models are generalizable and not over-tuned on a specific dataset, we perform cross-dataset validations by testing our model on different datasets.
## Socially Interactive Agent and Visual Explanations
For the presented approach, a socially interactive agent together with the current training task and explanations will be displayed on a monitor Fig. [\[fig:system_concept\]](#fig:system_concept){reference-type="ref" reference="fig:system_concept"}). The agent acts as a mediator to motivate, inform, and help the patient carry out specific neurorehabilitation tasks. The agent's behavior and the rehabilitation device are tuned in a way that it seems like they are one entity, and the agent is helping the patient apply a certain amount of force. Another task of the agent is to verbally give explanations of the current (and past) states of the training if asked by the patient or if best practice strategies (Sec. [2.1](#sec:bpnt){reference-type="ref" reference="sec:bpnt"}) suggest giving explanations to overcome specific training situations pro-actively. Technology-wise, the agent's behavior is modeled with the open-source Visual SceneMaker (VSM, [scenemaker.dfki.de](http://scenemaker.dfki.de)) toolkit. The tool is explicitly designed to facilitate studies with interactive agents. It comes with a real-time execution and authoring component for modeling verbal and non-verbal behavior of virtual agents and system actions. VSM also allows adapting an agent's behavior relying on external information procedurally. The agent's contributions and behavior are influenced by the patient's social and physiological signals, and these signals are interpreted in real-time by a social signal interpretation framework (SSI, [openssi.net](http://openssi.net)) (Sec. [3.3](#sec:ssi){reference-type="ref" reference="sec:ssi"}). The Gloria agent is a high-quality agent with a natural human appearance and verbal and nonverbal dialogue skills. The agent software independently runs in a web browser ([charamel.com](https://www.charamel.com)). The agent is capable of performing social cue-based interaction with the user. It performs lip-sync speech output using the state-of-the-art CereProc ([cereproc.com](https://www.cereporc.com)) Text-To-Speech system. For a more advanced animation control, Gloria allows the direct manipulation of skeleton model joints (e.g., neck joint). In addition, it comes with 36 conversational motion-captured gestures, which can be modified during run-time in some aspects (e.g., overall speed). Moreover, it provides 14 facial expressions, which contains, among others, the six basic emotional expressions defined by Ekman.
# Summary and Future Work
In this paper, we have presented an approach for an empathetic robotic neurorehabilitation system to help the patients train at home without a therapist but with a socially interactive agent in the role of a coaching assistant. To ensure proper and suitable training exercises, the system is prepared for patients with the help of a professional physiotherapist before it can be used at home. Our vision is to maximize the training effect by measuring social signals, e.g., attentiveness and stress, and physiological signals, e.g., EEG and EDA, to tailor the time and difficulty of training exercises. The interactive social agent motivates and gives explanations. Patients can choose if the agent should behave supportive or demanding. The following steps consist of evaluating several aspects concerning training efficiency with objective measurements (e.g., overall speed, precision) and standard questionnaires. Furthermore, we investigate if the presence of an interactive agent during therapy sessions has a significant impact on the user's engagement and training progress. Finally, another future evaluation asks if the social and explanatory skills of the system (with or without an agent) impact the training progress? | {'timestamp': '2022-06-06T02:14:58', 'yymm': '2206', 'arxiv_id': '2206.01587', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01587'} |
# Adapted Experimental Setup {#sec:adapted_setup}
Based on the prior failed experiments, we adapt the experimental setup by committing the unused space with a cold file in an effort to fully utilize the ZNS device, instead of leaving the unused space empty. We additionally make use of two conventional SSDs, one of which is a Samsung flash-based SSD and another Optane-based SSD. The conventional SSDs are used to provide a performance comparison between the conventional namespace on the ZNS device and regular SSDs. Detailed characteristics of the additional SSDs are presented in Table [\[tab:SSD_architecture\]](#tab:SSD_architecture){reference-type="ref" reference="tab:SSD_architecture"}. The conventional SSDs do not support multiple namespaces, therefore separate partitions are used, with the same 100GB of experimental space, and the remaining space serving as storage for cold data. While the ZNS device used during the experiments supports setting sector sizes of 512B and 4KiB, we set the device to 512B, since the used conventional SSDs only support 512B sectors, and we aim to keep device configurations as similar as possible. Prior to running experiments devices are pre-conditioned to steady state performance by writing the entire space numerous times.
# General Information
Throughout this guide, commands and set up explanation contain names of the specific NVMe device which are depicting their configuration in our system. *nvme0n1p1* depicts the Samsung SSD, and *nvme1n1p2* the Optane SSD, both of which only support a single namespace, requiring a partition to set up the 100GB experimental space. The conventional namespace of the ZNS device is *nvme2n1* and the zoned namepace is *nvme2n2*. We provide scripts for automation of all these benchmarks [^1], as often numerous steps and retrieval of device specific information is required.
# Device Setup {#appendix:Device_Setup}
This section contains all required setup for devices, including namespace configuration, as well as the required applications for the different configurations used in this evaluation. To interact with ZNS devices, libnvme [^2], nvme-cli [^3], blkzone from util-linux [^4], libzbd [^5], and all their dependencies need to be installed. Note, ZNS integration is largely still new to applications used in this evaluation, therefore using the master branch is often required. Additionally, ZNS support was added to the Linux Kernel 5.9, therefore this version or newer one is required.
## ZNS Device Configuration
As we are using several namespaces on the ZNS device, one to expose a small amount of conventional randomly writable area, another of 100GiB (50 zones), and one for the remaining available space. The command for identifying the available size of the device is shown in listing [\[Listing:ZNS_info\]](#Listing:ZNS_info){reference-type="ref" reference="Listing:ZNS_info"}, note that the ZNS device is configured to a 512B sector size. See Appendix [10](#appendix:ZNS_info){reference-type="ref" reference="appendix:ZNS_info"} on how to retrieve the supported sector sizes for a device. Next, the creation of namespaces is depicted in listing [\[Listing:ZNS_setup\]](#Listing:ZNS_setup){reference-type="ref" reference="Listing:ZNS_setup"}. After creation of the namespaces, it is important to set the appropriate scheduler for all zones namespaces, as applications often do not do this automatically or check if the desired scheduler is set.
``` {#Listing:ZNS_setup .bash language="bash" caption="Setting up NVMe ZNS namespaces." label="Listing:ZNS_setup"}
# Create a 100GiB namespace, size is given in 512B sectors \( sudo nvme create-ns /dev/nvme2-s 209715200-c 209715200-b 512--csi=2 # Repeat for all namespaces with according size # Attach all namespaces to same controller (adapt-n argument with ns id) sudo nvme attach-ns /dev/nvme2-n 1-c 0 # Set the correct scheduler for all zoned namespaces (adapt device path for each ns) \) echo mq-deadline | sudo tee /sys/block/nvme2n2/queue/scheduler
```
## f2fs Configuration
Setting up of f2fs requires f2fs-tools [^6] to make the file system. Configurations of f2fs with a ZNS device require an additional regular block device that is randomly writable, due to f2fs using in place updates for metadata and the superblock. In addition, both devices have to be configured to the same sector size. The exact commands for creating of the f2fs file system and mounting it are shown in listing [\[Listing:f2fs\]](#Listing:f2fs){reference-type="ref" reference="Listing:f2fs"}. The order of devices specified lists one or more zoned device, followed by a single conventional block device that is randomly writable. The location of the superblock is used for mounting, hence only the randomly writable device used for file system creating is provided in the mount command.
``` {#Listing:f2fs .bash language="bash" caption="Creating and mounting of f2fs file system. Requires the ZNS device and an additional randomly writable block device." label="Listing:f2fs"}
# Format devices and create fs \( sudo mkfs.f2fs-f-m-c /dev/nvme2n2 /dev/nvme2n1 \) sudo mount-t f2fs /dev/nvme2n1 /mnt/f2fs/
```
## ZenFS Configuration
ZenFS [^7] provides the storage backend for RocksDB to provide usage of ZNS devices. The ZenFS file system allows to be backed up and recovered to avoid data loss in failure events. Setting up of ZenFS requires the zoned device and an additional auxiliary path on another device with a file system, where it places the backup files, as well as any LOG and LOCK files required during RocksDB runtime. The command to set up ZenFS on a zoned device is shown in listing [\[Listing:zenfs\]](#Listing:zenfs){reference-type="ref" reference="Listing:zenfs"}. The auxiliary path is placed on a conventional block device that is mounted with f2fs.
``` {#Listing:zenfs .bash language="bash" caption="Creating of ZenFS file system. Requires an auxiliary path to place metadata." label="Listing:zenfs"}
$ sudo./plugin/zenfs/util/zenfs mkfs--zbd=nvme2n2--aux_path=/home/nty/rocksdb_aux_path/zenfs2n2
```
## Namespace initialization
As mentioned previously, we utilize a 100GiB namespace (*nvme2n2*) for experiments and leave the remaining available space in a separate namespace (*nvme2n3*) to be filled with cold data. Listing [\[Listing:ZNS_Cold_Data_NS\]](#Listing:ZNS_Cold_Data_NS){reference-type="ref" reference="Listing:ZNS_Cold_Data_NS"} shows how this is achieved using fio [^8]. For fio to be able to write the entire namespace on the device, it requires the block size to be a multiple of the zone capacity (see listing [\[Listing:ZNS_info\]](#Listing:ZNS_info){reference-type="ref" reference="Listing:ZNS_info"} on how to retrieve it). Similarly, the conventional devices (Optane and Samsung SSDs) also have their free space filled with cold data with the second shown command. The command is set up to write 2TiB however, fio will quit once the device is full. Additionally, note that as mentioned earlier the conventional SSDs only support a single namespace and hence have separate partitions set up for the experimental and cold data space.
``` {#Listing:ZNS_Cold_Data_NS .bash language="bash" caption="Filling namespace 3 with cold data." label="Listing:ZNS_Cold_Data_NS"}
# Fill ZNS free space with cold data \( sudo fio--name=zns-fio--filename=/dev/nvme2n3--direct=1--size=\)((4194304*512*`cat /sys/block/nvme2n3/queue/nr_zones`))--ioengine=libaio--zonemode=zbd--iodepth=8--rw=write--bs=512K # Fill conventional SSD free space with cold data $ sudo fio--name=zns-fio--filename=/dev/nvme0np2--direct=1--size=2T--ioengine=libaio--iodepth=8--rw=write--bs=512K
```
# Getting ZNS Device Information {#appendix:ZNS_info}
There are several attributes to the ZNS device that are required for later experiments, such as the zone capacity and the number of allowed active zones. Listing [\[Listing:ZNS_info\]](#Listing:ZNS_info){reference-type="ref" reference="Listing:ZNS_info"} illustrates how to retrieve these. Supported sector sizes can be checked with the provided command, and are presented in powers of 2. Hence, a *lbads:9* is equivalent to \(2^9=512\) Bytes. We additionally retrieve the NUMA node at which the device is attached to, in order to pin workloads to this specific NUMA node.
``` {#Listing:ZNS_info .bash language="bash" caption="Retrieving information about the ZNS device." label="Listing:ZNS_info"}
# Get the available device capacity in 512B sectors \( sudo nvme id-ctrl /dev/nvme2 | grep tnvmcap | awk '{print \)3/512}' # Get the zone capacity in MiB \( sudo nvme zns report-zones /dev/nvme2n2-d 1 | grep-o 'Cap:.*\)' | awk '{print strtonum(\(2)*512/1024/1024}' # Get maximum supported active zones \) cat /sys/block/nvme2n2/queue/max_active_zones # Get the supported sector sizes in powers of 2 \( sudo nvme id-ns /dev/nvme2n2 | grep-o "lbads:[0-9]*" # Get NUMA node device is attached at \) cat/sys/block/nvme2n1/device/numa_node
```
# ZNS Block-Level I/O Performance
This section contains the commands used to establish the baseline maximum performance of the device, as well as extracted metrics for comparison to later experiments. All experiments for this are using fio as benchmarking tool. Appendix [11.1](#appendix:Conv_device_performance){reference-type="ref" reference="appendix:Conv_device_performance"} shows the set-up and commands for the provided evaluation in [22.1](#sec:block-level_ZNS_IO_Conv_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Conv_ns"}, and Appendix [11.2](#appendix:Zoned_device_performance){reference-type="ref" reference="appendix:Zoned_device_performance"} shows the same evaluation on the ZNS device, as presented in [22.2](#sec:block-level_ZNS_IO_Zoned_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Zoned_ns"}.
## Conventional Device Performance {#appendix:Conv_device_performance}
**Device throughput:** We run sequential and random writing and reading benchmarks for a block size (I/O size) of 4KiB under varying I/O queue depths to identify the maximum achievable IOPs of the devices. We run this on all three devices, since they all expose the regular conventional block device interface without write constraints, hence we only use the conventional namespace of the ZNS device in this section. The commands are shown in Listing [\[listing:conv_performance\]](#listing:conv_performance){reference-type="ref" reference="listing:conv_performance"}. The order of benchmarks is intentional such that randomwrite first runs, then the namespace is reset and written with the sequential benchmark. Overwrite benchmarks are done similarly with sequential and random writing after the entire namespace is filled with a write benchmark. Note that benchmarks are pinned to the NUMA node where the device is attached (see Appendix [10](#appendix:ZNS_info){reference-type="ref" reference="appendix:ZNS_info"} for retrieval of this). All defined variables (indicated by the \$ before a variable) are set up by our script, however for manual running of these commands require to simply be replaced by their associated value. The name and output argument depict naming for our plotting script to parse, however can simply be changed. **ZNS device bandwidth:** We additionally showed the bandwidth scaling for the conventional namespace on the ZNS device for a I/O queue depth of 4 and increasing block sizes, which are shown in Listing [\[listing:conv_bandwidth\]](#listing:conv_bandwidth){reference-type="ref" reference="listing:conv_bandwidth"}. Again, replace all defined variables with their respective value.
``` {#listing:conv_performance .bash language="bash" caption="Establishing the peak IOPs for the conventional devices with 4KiB block size and varying I/O queue depths." label="listing:conv_performance"}
# We define several variables, replace these with values # DEV: device name (e.g., nvme2n1) # depth: I/O queue depth (from 1-1024 in powers of 2) # DEV_NUMA_NODE: NUMA Node of the device # SIZE: device size (e.g., 100G) \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randwrite_4Ki_queue-depth-\)depth--output=\(DEV_randwrite_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=randwrite--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Reset the namespace between write benchmarks (only on namespaces, not partitions) \( sudo nvme format /dev/\)DEV-f # Run remaining benchmarks \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_write_4Ki_queue-depth-\)depth--output=\(DEV_write_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_read_4Ki_queue-depth-\)depth--output=\(DEV_read_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=read--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randread_4Ki_queue-depth-\)depth--output=\(DEV_randread_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=randread--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Namespace is still full so run overwrite benchs \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-seq_4Ki_queue-depth-\)depth--output=\(DEV_overwrite-seq_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-rand_4Ki_queue-depth-\)depth--output=\(DEV_overwrite-rand_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=randwrite--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95
```
``` {#listing:conv_bandwidth .bash language="bash" caption="Establishing the maximum achievable bandwidth of the ZNS convention namespace with I/O queue depth of 4 and increasing block size." label="listing:conv_bandwidth"}
# We define several variables, replace these with values # DEV: device name (e.g., nvme2n1) # block_size: I/O size (from 4KiB to 128KiB in powers of 2) # DEV_NUMA_NODE: NUMA Node of the device # SIZE: device size (e.g., 100G) \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randwrite_\)block_size_queue-depth-4--output=\(DEV_randwrite_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=randwrite--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Reset the namespace between write benchmarks (only on namespaces, not partitions) \( sudo nvme format /dev/\)DEV-f # Run remaining benchmarks \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_write_\)block_size_queue-depth-4--output=\(DEV_write_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=write--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_read_\)block_size_queue-depth-4--output=\(DEV_read_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=read--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randread_\)block_size_queue-depth-4--output=\(DEV_randread_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=randread--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Namespace is still full so run overwrite benchs \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-seq_\)block_size_queue-depth-4--output=\(DEV_overwrite-seq_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=write--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-rand_\)block_size_queue-depth-4--output=\(DEV_overwrite-rand_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=randwrite--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95
```
## Zoned Device Performance {#appendix:Zoned_device_performance}
We run various workloads on the zoned namespace of the ZNS device, and compare the performance of the scheduler set to *mq-deadline* and *none*. Listing [\[Listing:ZNS_setup\]](#Listing:ZNS_setup){reference-type="ref" reference="Listing:ZNS_setup"} showed how to change the scheduler for a namespace. For all benchmarks we utilize fio configured to use 4KiB I/Os. **Write performance:** The benchmarks are shown in Listing [\[listing:ZNS_baseline_performance\]](#listing:ZNS_baseline_performance){reference-type="ref" reference="listing:ZNS_baseline_performance"}. The listing also shows the write benchmark that was used to produce Figure [\[fig:ZNS_concur_write_seq\]](#fig:ZNS_concur_write_seq){reference-type="ref" reference="fig:ZNS_concur_write_seq"}. This benchmark writes a single 4KiB I/O to a zone and we increase the number of concurrent writes that are issued to the increasing number of active zones. Therefore, the benchmarks for the different schedulers have the exact same command, and only requires to change the scheduler in between iterations, when the number of concurrent jobs is increased. Note, we also use a 50 zone namespace and have a maximum of 14 active zones, which is why the *offset_increment* flag is set to 3z, such that each additional thread starts at an increasing offset of 3 zones and all 14 threads can still fit into the namespace when running concurrently. In addition, the write size of each thread is 3z. Also note that the benchmarks with I/O queue depth of 1 and increasing threads over active zones utilize the *psync* ioengine, since I/Os are synchronous, as opposed to *libaio* with asynchronous benchmarks where I/O queue depth is larger than 1.
``` {#listing:ZNS_baseline_performance .bash language="bash" caption="Measuring write latency for the schedulers with increasing number of active zones and a single outstanding I/O per zone." label="listing:ZNS_baseline_performance"}
# We define several variables, replace these with values # DEV: device name (e.g., nvme2n2) # DEV_NUMA_NODE: NUMA Node of the device # scheduler: current scheduler # jobs: [1-14] number of active zones # Write benchmark, change the scheduler between iterations \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_\({scheduler}_4Ki_nummjobs-\){jobs}")--output=\((echo "\){DEV}_\({scheduler}_4Ki_numjobs-\){jobs}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=3z--offset_increment=3z--ioengine=psync--zonemode=zbd--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--group_reporting--numjobs=$jobs--percentile_list=50:95
```
The next write benchmark, as was shown in Figure [\[fig:ZNS_concur_write_seq_iodepth\]](#fig:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="fig:ZNS_concur_write_seq_iodepth"}, runs *mq-deadline* with a single zone and an increasing I/O queue depth and *none* runs just as before, with a single I/O per zone and an increasing number of active zones. Listing [\[listing:ZNS_concur_write_seq_iodepth\]](#listing:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_write_seq_iodepth"} shows these specific commands. This benchmark defines the size as the total space of the device, since it runs a single thread.
``` {#listing:ZNS_concur_write_seq_iodepth .bash language="bash" caption="Measuring write latency for the schedulers with increasing number of active zones and a single outstanding I/O per zone." label="listing:ZNS_concur_write_seq_iodepth"}
# In addition to the prior defined variables we also define # SIZE: device size (e.g., 100G) # depth: [1-14] I/O queue depth, # and numjobs for none # mq-deadline benchmark \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_mq-deadline_\({BS}_iodepth-\){depth}")--output=\((echo "\){DEV}_mq-deadline_\({BS}_iodepth-\){depth}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--zonemode=zbd--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\(DEV_NUMA_NODE--ramp_time=10--iodepth=\)depth--time_based--percentile_list=50:95 # none benchmark \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_none_4Ki_nummjobs-\({jobs}")--output=\)(echo "\({DEV}_none_4Ki_numjobs-\){jobs}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--offset_increment=3z--ioengine=psync--zonemode=zbd--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\(DEV_NUMA_NODE--ramp_time=10--numjobs=\)depth--time_based--group_reporting--percentile_list=50:95
```
**Read performance:** Read performance was measured with sequential reading of 4KiB on a single zone and an increasing I/O queue depth for both schedulers. Recall, that ZNS devices only have a sequential write constraint, and hence read requests can be issued with increasing I/O queue depth under any scheduler and do not have to be at the write pointer of a zone. Listing [\[listing:ZNS_concur_read_seq_iodepth\]](#listing:ZNS_concur_read_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_read_seq_iodepth"} shows the command for this particular benchmark. The benchmark increases only the I/O queue depth from 1 to 14, and commands are the same under both scheduler configurations, however the correct scheduler has to be set before each iteration. Next, we run the random read benchmark. The *mq-deadline* configuration uses the exact same set up as for sequential reading, namely issuing 4KiB I/Os in a single zone with an increasing I/O queue depth. However, *none* issues a single 4KiB I/O per zone, with an increasing number of concurrent threads. The command from Listing [\[listing:ZNS_concur_read_seq_iodepth\]](#listing:ZNS_concur_read_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_read_seq_iodepth"} can be used for the *mq-deadline* scheduler, and the *none* scheduler can use the command in Listing [\[listing:ZNS_concur_write_seq_iodepth\]](#listing:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_write_seq_iodepth"}, with the benchmark parameter *--rw=randread*. Note, that similar to the prior read benchmark, the namesapce needs to be full in order to be able to read data.
``` {#listing:ZNS_concur_read_seq_iodepth .bash language="bash" caption="Measuring sequential read performance for 4KiB I/Os in a single zone with increasing I/O queue depth." label="listing:ZNS_concur_read_seq_iodepth"}
# Defined variables # scheduler: current scheduler # depth: [1-14] I/O queue depth # Change scheduler between iterations \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_\({scheduler}_4Ki_iodepth-\){depth}")--output=\((echo "\){DEV}_\({scheduler}_4Ki_iodepth-\){depth}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=SIZE--ioengine=libaio--zonemode=zbd--rw=read--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--iodepth=$depth--time_based--percentile_list=50:95
```
# Footnotes, Verbatim, and Citations
# Adapted Experimental Setup {#sec:adapted_setup}
Based on the prior failed experiments, we adapt the experimental setup by committing the unused space with a cold file in an effort to fully utilize the ZNS device, instead of leaving the unused space empty. We additionally make use of two conventional SSDs, one of which is a Samsung flash-based SSD and another Optane-based SSD. The conventional SSDs are used to provide a performance comparison between the conventional namespace on the ZNS device and regular SSDs. Detailed characteristics of the additional SSDs are presented in Table [\[tab:SSD_architecture\]](#tab:SSD_architecture){reference-type="ref" reference="tab:SSD_architecture"}. The conventional SSDs do not support multiple namespaces, therefore separate partitions are used, with the same 100GB of experimental space, and the remaining space serving as storage for cold data. While the ZNS device used during the experiments supports setting sector sizes of 512B and 4KiB, we set the device to 512B, since the used conventional SSDs only support 512B sectors, and we aim to keep device configurations as similar as possible. Prior to running experiments devices are pre-conditioned to steady state performance by writing the entire space numerous times.
# General Information
Throughout this guide, commands and set up explanation contain names of the specific NVMe device which are depicting their configuration in our system. *nvme0n1p1* depicts the Samsung SSD, and *nvme1n1p2* the Optane SSD, both of which only support a single namespace, requiring a partition to set up the 100GB experimental space. The conventional namespace of the ZNS device is *nvme2n1* and the zoned namepace is *nvme2n2*. We provide scripts for automation of all these benchmarks [^9], as often numerous steps and retrieval of device specific information is required.
# Device Setup {#appendix:Device_Setup}
This section contains all required setup for devices, including namespace configuration, as well as the required applications for the different configurations used in this evaluation. To interact with ZNS devices, libnvme [^10], nvme-cli [^11], blkzone from util-linux [^12], libzbd [^13], and all their dependencies need to be installed. Note, ZNS integration is largely still new to applications used in this evaluation, therefore using the master branch is often required. Additionally, ZNS support was added to the Linux Kernel 5.9, therefore this version or newer one is required.
## ZNS Device Configuration
As we are using several namespaces on the ZNS device, one to expose a small amount of conventional randomly writable area, another of 100GiB (50 zones), and one for the remaining available space. The command for identifying the available size of the device is shown in listing [\[Listing:ZNS_info\]](#Listing:ZNS_info){reference-type="ref" reference="Listing:ZNS_info"}, note that the ZNS device is configured to a 512B sector size. See Appendix [10](#appendix:ZNS_info){reference-type="ref" reference="appendix:ZNS_info"} on how to retrieve the supported sector sizes for a device. Next, the creation of namespaces is depicted in listing [\[Listing:ZNS_setup\]](#Listing:ZNS_setup){reference-type="ref" reference="Listing:ZNS_setup"}. After creation of the namespaces, it is important to set the appropriate scheduler for all zones namespaces, as applications often do not do this automatically or check if the desired scheduler is set.
``` {#Listing:ZNS_setup .bash language="bash" caption="Setting up NVMe ZNS namespaces." label="Listing:ZNS_setup"}
# Create a 100GiB namespace, size is given in 512B sectors \( sudo nvme create-ns /dev/nvme2-s 209715200-c 209715200-b 512--csi=2 # Repeat for all namespaces with according size # Attach all namespaces to same controller (adapt-n argument with ns id) sudo nvme attach-ns /dev/nvme2-n 1-c 0 # Set the correct scheduler for all zoned namespaces (adapt device path for each ns) \) echo mq-deadline | sudo tee /sys/block/nvme2n2/queue/scheduler
```
## f2fs Configuration
Setting up of f2fs requires f2fs-tools [^14] to make the file system. Configurations of f2fs with a ZNS device require an additional regular block device that is randomly writable, due to f2fs using in place updates for metadata and the superblock. In addition, both devices have to be configured to the same sector size. The exact commands for creating of the f2fs file system and mounting it are shown in listing [\[Listing:f2fs\]](#Listing:f2fs){reference-type="ref" reference="Listing:f2fs"}. The order of devices specified lists one or more zoned device, followed by a single conventional block device that is randomly writable. The location of the superblock is used for mounting, hence only the randomly writable device used for file system creating is provided in the mount command.
``` {#Listing:f2fs .bash language="bash" caption="Creating and mounting of f2fs file system. Requires the ZNS device and an additional randomly writable block device." label="Listing:f2fs"}
# Format devices and create fs \( sudo mkfs.f2fs-f-m-c /dev/nvme2n2 /dev/nvme2n1 \) sudo mount-t f2fs /dev/nvme2n1 /mnt/f2fs/
```
## ZenFS Configuration
ZenFS [^15] provides the storage backend for RocksDB to provide usage of ZNS devices. The ZenFS file system allows to be backed up and recovered to avoid data loss in failure events. Setting up of ZenFS requires the zoned device and an additional auxiliary path on another device with a file system, where it places the backup files, as well as any LOG and LOCK files required during RocksDB runtime. The command to set up ZenFS on a zoned device is shown in listing [\[Listing:zenfs\]](#Listing:zenfs){reference-type="ref" reference="Listing:zenfs"}. The auxiliary path is placed on a conventional block device that is mounted with f2fs.
``` {#Listing:zenfs .bash language="bash" caption="Creating of ZenFS file system. Requires an auxiliary path to place metadata." label="Listing:zenfs"}
$ sudo./plugin/zenfs/util/zenfs mkfs--zbd=nvme2n2--aux_path=/home/nty/rocksdb_aux_path/zenfs2n2
```
## Namespace initialization
As mentioned previously, we utilize a 100GiB namespace (*nvme2n2*) for experiments and leave the remaining available space in a separate namespace (*nvme2n3*) to be filled with cold data. Listing [\[Listing:ZNS_Cold_Data_NS\]](#Listing:ZNS_Cold_Data_NS){reference-type="ref" reference="Listing:ZNS_Cold_Data_NS"} shows how this is achieved using fio [^16]. For fio to be able to write the entire namespace on the device, it requires the block size to be a multiple of the zone capacity (see listing [\[Listing:ZNS_info\]](#Listing:ZNS_info){reference-type="ref" reference="Listing:ZNS_info"} on how to retrieve it). Similarly, the conventional devices (Optane and Samsung SSDs) also have their free space filled with cold data with the second shown command. The command is set up to write 2TiB however, fio will quit once the device is full. Additionally, note that as mentioned earlier the conventional SSDs only support a single namespace and hence have separate partitions set up for the experimental and cold data space.
``` {#Listing:ZNS_Cold_Data_NS .bash language="bash" caption="Filling namespace 3 with cold data." label="Listing:ZNS_Cold_Data_NS"}
# Fill ZNS free space with cold data \( sudo fio--name=zns-fio--filename=/dev/nvme2n3--direct=1--size=\)((4194304*512*`cat /sys/block/nvme2n3/queue/nr_zones`))--ioengine=libaio--zonemode=zbd--iodepth=8--rw=write--bs=512K # Fill conventional SSD free space with cold data $ sudo fio--name=zns-fio--filename=/dev/nvme0np2--direct=1--size=2T--ioengine=libaio--iodepth=8--rw=write--bs=512K
```
# Getting ZNS Device Information {#appendix:ZNS_info}
There are several attributes to the ZNS device that are required for later experiments, such as the zone capacity and the number of allowed active zones. Listing [\[Listing:ZNS_info\]](#Listing:ZNS_info){reference-type="ref" reference="Listing:ZNS_info"} illustrates how to retrieve these. Supported sector sizes can be checked with the provided command, and are presented in powers of 2. Hence, a *lbads:9* is equivalent to \(2^9=512\) Bytes. We additionally retrieve the NUMA node at which the device is attached to, in order to pin workloads to this specific NUMA node.
``` {#Listing:ZNS_info .bash language="bash" caption="Retrieving information about the ZNS device." label="Listing:ZNS_info"}
# Get the available device capacity in 512B sectors \( sudo nvme id-ctrl /dev/nvme2 | grep tnvmcap | awk '{print \)3/512}' # Get the zone capacity in MiB \( sudo nvme zns report-zones /dev/nvme2n2-d 1 | grep-o 'Cap:.*\)' | awk '{print strtonum(\(2)*512/1024/1024}' # Get maximum supported active zones \) cat /sys/block/nvme2n2/queue/max_active_zones # Get the supported sector sizes in powers of 2 \( sudo nvme id-ns /dev/nvme2n2 | grep-o "lbads:[0-9]*" # Get NUMA node device is attached at \) cat/sys/block/nvme2n1/device/numa_node
```
# ZNS Block-Level I/O Performance
This section contains the commands used to establish the baseline maximum performance of the device, as well as extracted metrics for comparison to later experiments. All experiments for this are using fio as benchmarking tool. Appendix [11.1](#appendix:Conv_device_performance){reference-type="ref" reference="appendix:Conv_device_performance"} shows the set-up and commands for the provided evaluation in [22.1](#sec:block-level_ZNS_IO_Conv_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Conv_ns"}, and Appendix [11.2](#appendix:Zoned_device_performance){reference-type="ref" reference="appendix:Zoned_device_performance"} shows the same evaluation on the ZNS device, as presented in [22.2](#sec:block-level_ZNS_IO_Zoned_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Zoned_ns"}.
## Conventional Device Performance {#appendix:Conv_device_performance}
**Device throughput:** We run sequential and random writing and reading benchmarks for a block size (I/O size) of 4KiB under varying I/O queue depths to identify the maximum achievable IOPs of the devices. We run this on all three devices, since they all expose the regular conventional block device interface without write constraints, hence we only use the conventional namespace of the ZNS device in this section. The commands are shown in Listing [\[listing:conv_performance\]](#listing:conv_performance){reference-type="ref" reference="listing:conv_performance"}. The order of benchmarks is intentional such that randomwrite first runs, then the namespace is reset and written with the sequential benchmark. Overwrite benchmarks are done similarly with sequential and random writing after the entire namespace is filled with a write benchmark. Note that benchmarks are pinned to the NUMA node where the device is attached (see Appendix [10](#appendix:ZNS_info){reference-type="ref" reference="appendix:ZNS_info"} for retrieval of this). All defined variables (indicated by the \$ before a variable) are set up by our script, however for manual running of these commands require to simply be replaced by their associated value. The name and output argument depict naming for our plotting script to parse, however can simply be changed. **ZNS device bandwidth:** We additionally showed the bandwidth scaling for the conventional namespace on the ZNS device for a I/O queue depth of 4 and increasing block sizes, which are shown in Listing [\[listing:conv_bandwidth\]](#listing:conv_bandwidth){reference-type="ref" reference="listing:conv_bandwidth"}. Again, replace all defined variables with their respective value.
``` {#listing:conv_performance .bash language="bash" caption="Establishing the peak IOPs for the conventional devices with 4KiB block size and varying I/O queue depths." label="listing:conv_performance"}
# We define several variables, replace these with values # DEV: device name (e.g., nvme2n1) # depth: I/O queue depth (from 1-1024 in powers of 2) # DEV_NUMA_NODE: NUMA Node of the device # SIZE: device size (e.g., 100G) \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randwrite_4Ki_queue-depth-\)depth--output=\(DEV_randwrite_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=randwrite--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Reset the namespace between write benchmarks (only on namespaces, not partitions) \( sudo nvme format /dev/\)DEV-f # Run remaining benchmarks \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_write_4Ki_queue-depth-\)depth--output=\(DEV_write_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_read_4Ki_queue-depth-\)depth--output=\(DEV_read_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=read--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randread_4Ki_queue-depth-\)depth--output=\(DEV_randread_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=randread--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Namespace is still full so run overwrite benchs \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-seq_4Ki_queue-depth-\)depth--output=\(DEV_overwrite-seq_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-rand_4Ki_queue-depth-\)depth--output=\(DEV_overwrite-rand_4Ki_queue-depth-\)depth.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=\(depth--rw=randwrite--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95
```
``` {#listing:conv_bandwidth .bash language="bash" caption="Establishing the maximum achievable bandwidth of the ZNS convention namespace with I/O queue depth of 4 and increasing block size." label="listing:conv_bandwidth"}
# We define several variables, replace these with values # DEV: device name (e.g., nvme2n1) # block_size: I/O size (from 4KiB to 128KiB in powers of 2) # DEV_NUMA_NODE: NUMA Node of the device # SIZE: device size (e.g., 100G) \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randwrite_\)block_size_queue-depth-4--output=\(DEV_randwrite_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=randwrite--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Reset the namespace between write benchmarks (only on namespaces, not partitions) \( sudo nvme format /dev/\)DEV-f # Run remaining benchmarks \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_write_\)block_size_queue-depth-4--output=\(DEV_write_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=write--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_read_\)block_size_queue-depth-4--output=\(DEV_read_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=read--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_randread_\)block_size_queue-depth-4--output=\(DEV_randread_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=randread--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 # Namespace is still full so run overwrite benchs \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-seq_\)block_size_queue-depth-4--output=\(DEV_overwrite-seq_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=write--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95 \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\(DEV_overwrite-rand_\)block_size_queue-depth-4--output=\(DEV_overwrite-rand_\)block_size_queue-depth-4.json--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--iodepth=4--rw=randwrite--bs=\(block_size--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--time_based--percentile_list=50:95
```
## Zoned Device Performance {#appendix:Zoned_device_performance}
We run various workloads on the zoned namespace of the ZNS device, and compare the performance of the scheduler set to *mq-deadline* and *none*. Listing [\[Listing:ZNS_setup\]](#Listing:ZNS_setup){reference-type="ref" reference="Listing:ZNS_setup"} showed how to change the scheduler for a namespace. For all benchmarks we utilize fio configured to use 4KiB I/Os. **Write performance:** The benchmarks are shown in Listing [\[listing:ZNS_baseline_performance\]](#listing:ZNS_baseline_performance){reference-type="ref" reference="listing:ZNS_baseline_performance"}. The listing also shows the write benchmark that was used to produce Figure [\[fig:ZNS_concur_write_seq\]](#fig:ZNS_concur_write_seq){reference-type="ref" reference="fig:ZNS_concur_write_seq"}. This benchmark writes a single 4KiB I/O to a zone and we increase the number of concurrent writes that are issued to the increasing number of active zones. Therefore, the benchmarks for the different schedulers have the exact same command, and only requires to change the scheduler in between iterations, when the number of concurrent jobs is increased. Note, we also use a 50 zone namespace and have a maximum of 14 active zones, which is why the *offset_increment* flag is set to 3z, such that each additional thread starts at an increasing offset of 3 zones and all 14 threads can still fit into the namespace when running concurrently. In addition, the write size of each thread is 3z. Also note that the benchmarks with I/O queue depth of 1 and increasing threads over active zones utilize the *psync* ioengine, since I/Os are synchronous, as opposed to *libaio* with asynchronous benchmarks where I/O queue depth is larger than 1.
``` {#listing:ZNS_baseline_performance .bash language="bash" caption="Measuring write latency for the schedulers with increasing number of active zones and a single outstanding I/O per zone." label="listing:ZNS_baseline_performance"}
# We define several variables, replace these with values # DEV: device name (e.g., nvme2n2) # DEV_NUMA_NODE: NUMA Node of the device # scheduler: current scheduler # jobs: [1-14] number of active zones # Write benchmark, change the scheduler between iterations \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_\({scheduler}_4Ki_nummjobs-\){jobs}")--output=\((echo "\){DEV}_\({scheduler}_4Ki_numjobs-\){jobs}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=3z--offset_increment=3z--ioengine=psync--zonemode=zbd--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--group_reporting--numjobs=$jobs--percentile_list=50:95
```
The next write benchmark, as was shown in Figure [\[fig:ZNS_concur_write_seq_iodepth\]](#fig:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="fig:ZNS_concur_write_seq_iodepth"}, runs *mq-deadline* with a single zone and an increasing I/O queue depth and *none* runs just as before, with a single I/O per zone and an increasing number of active zones. Listing [\[listing:ZNS_concur_write_seq_iodepth\]](#listing:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_write_seq_iodepth"} shows these specific commands. This benchmark defines the size as the total space of the device, since it runs a single thread.
``` {#listing:ZNS_concur_write_seq_iodepth .bash language="bash" caption="Measuring write latency for the schedulers with increasing number of active zones and a single outstanding I/O per zone." label="listing:ZNS_concur_write_seq_iodepth"}
# In addition to the prior defined variables we also define # SIZE: device size (e.g., 100G) # depth: [1-14] I/O queue depth, # and numjobs for none # mq-deadline benchmark \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_mq-deadline_\({BS}_iodepth-\){depth}")--output=\((echo "\){DEV}_mq-deadline_\({BS}_iodepth-\){depth}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--ioengine=libaio--zonemode=zbd--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\(DEV_NUMA_NODE--ramp_time=10--iodepth=\)depth--time_based--percentile_list=50:95 # none benchmark \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_none_4Ki_nummjobs-\({jobs}")--output=\)(echo "\({DEV}_none_4Ki_numjobs-\){jobs}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=\)SIZE--offset_increment=3z--ioengine=psync--zonemode=zbd--rw=write--bs=4Ki--runtime=30--numa_cpu_nodes=\(DEV_NUMA_NODE--ramp_time=10--numjobs=\)depth--time_based--group_reporting--percentile_list=50:95
```
**Read performance:** Read performance was measured with sequential reading of 4KiB on a single zone and an increasing I/O queue depth for both schedulers. Recall, that ZNS devices only have a sequential write constraint, and hence read requests can be issued with increasing I/O queue depth under any scheduler and do not have to be at the write pointer of a zone. Listing [\[listing:ZNS_concur_read_seq_iodepth\]](#listing:ZNS_concur_read_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_read_seq_iodepth"} shows the command for this particular benchmark. The benchmark increases only the I/O queue depth from 1 to 14, and commands are the same under both scheduler configurations, however the correct scheduler has to be set before each iteration. Next, we run the random read benchmark. The *mq-deadline* configuration uses the exact same set up as for sequential reading, namely issuing 4KiB I/Os in a single zone with an increasing I/O queue depth. However, *none* issues a single 4KiB I/O per zone, with an increasing number of concurrent threads. The command from Listing [\[listing:ZNS_concur_read_seq_iodepth\]](#listing:ZNS_concur_read_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_read_seq_iodepth"} can be used for the *mq-deadline* scheduler, and the *none* scheduler can use the command in Listing [\[listing:ZNS_concur_write_seq_iodepth\]](#listing:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="listing:ZNS_concur_write_seq_iodepth"}, with the benchmark parameter *--rw=randread*. Note, that similar to the prior read benchmark, the namesapce needs to be full in order to be able to read data.
``` {#listing:ZNS_concur_read_seq_iodepth .bash language="bash" caption="Measuring sequential read performance for 4KiB I/Os in a single zone with increasing I/O queue depth." label="listing:ZNS_concur_read_seq_iodepth"}
# Defined variables # scheduler: current scheduler # depth: [1-14] I/O queue depth # Change scheduler between iterations \( sudo numactl-m \)DEV_NUMA_NODE fio--name=\((echo "\){DEV}_\({scheduler}_4Ki_iodepth-\){depth}")--output=\((echo "\){DEV}_\({scheduler}_4Ki_iodepth-\){depth}.json")--output-format=json--filename=/dev/\(DEV--direct=1--size=SIZE--ioengine=libaio--zonemode=zbd--rw=read--bs=4Ki--runtime=30--numa_cpu_nodes=\)DEV_NUMA_NODE--ramp_time=10--iodepth=$depth--time_based--percentile_list=50:95
```
# Background {#sec:background}
Applications and file systems rely on the Linux block layer to provide interfaces and abstractions for accessing the underlying storage media. Originally designed to match the hardware characteristics of HDDs the block layer presents the storage as a linear address space, allowing for sequential and random writes. Flash storage however has different write constraints due to its architecture. Relying on the FTL to hide the device management idiosyncrasies however leads to negative performance impacts due to the unpredictable performance from the device garbage collection , large tail latency it causes , and increased write amplification . The increased write amplification additionally reduces the device lifetime, since flash cells on SSDs have limited program/erase cycles. Furthermore, garbage collection requires the device to maintain a certain amount of free space, called the *overprovisioning space*, such that the FTL is able to move valid pages. Most commonly used overprovisioning takes between 10-28% of the device capacity. One of the possible FTL design uses a fully-associative mapping of host logical block addresses (LBAs) to physical addresses in order to provide the LBAs of the page(s) that contain valid data. Such a design requires significant resources in order to store all mappings.
## Zoned Storage {#sec:zoned_storage}
The arrival of ZNS SSDs eliminates the need for on device garbage collection done by the FTL, pushing this responsibility to the host. This provides the host with more opportunity for optimized data placement, through mechanisms such as data grouping, and makes garbage collection overheads predictable. The concept of exposing storage as zones is not new, as it was already introduced when Shingled Magnetic Recording (SMR) HDDs appeared, which also enforce a sequential write constraint. The zoned storage model was established with the addition of Zoned Block Device (ZBD) support in the Linux Kernel 4.10.0 , in an effort to avoid the mismatch between the block layer and the sequential write constraint on devices. The standards defining the management of SMR HDDs in the zoned storage model came through the Zoned Device ATA Command Set (ZAC) and the Zoned Block Command (ZBC) specifications. Since zones require sequential writing, the address of the current write is managed with a *write pointer*. The write pointer is incremented to the LBA of the next write only after a successful write. In order to manage zones, each zone has a state associated to it. These are to identify the condition of a zone, which can be any of the following; *EMPTY* to indicate a zone is empty, *OPEN* which is required for a zone to allow writes, *FULL* to indicate the write pointer is at the zone capacity, *CLOSED* which is used to release resources for the zone (e.g., write buffers on the ZNS device) without resetting a zone, this additionally does not allow writes to continue in the zone until it is transitioned to *OPEN* again, *OFFLINE* which makes all data in the zone inaccessible until the zone is reset, and *READ-ONLY*. The command sets provide the proper mechanisms to transition zones between any of the states. While the majority of available space for both SMR and ZNS devices is utilized by sequential write required zones, they can also expose a limited amount of randomly writable area. This is mainly intended for metadata as this space only occupies a very small percentage of the device capacity. For example, the ZNS device used in this evaluation could expose 4 zones as randomly writable, equivalent to approximately 4GiB, compared to the total device size of 7.2TiB. The newly introduced ZNS SSDs are standardized through the NVMe specification , which builds on the foundations established with ZBD support. While the zoned storage model aims to provide a unified software stack for all zoned devices (SMR and ZNS devices), the NVMe specification introduces several new concepts particular to ZNS devices. Firstly, it defines a *zone capacity* for each zone, stating the usable capacity of a zone, which is less than or equal to the size of the zone. Figure [\[fig:ZNS_HW\]](#fig:ZNS_HW){reference-type="ref" reference="fig:ZNS_HW"} shows an example layout of zones on a ZNS SSD and each zone's associated state. Zone capacity is kept separate from the zone size such that the capacity can be aligned to the erase unit of the underlying flash media, and such that the zone size can be kept as a power of two value. This is required for easier conversions between LBAs and zone offsets. Secondly, the *active zones* limit specifies the maximum number of zones that can be active (in *OPEN* or *CLOSED* state) at any point in time. As the SSD requires to allocate resources for each active zone, such as the write buffers, it enforces a maximum on the active zones. Lastly, the NVMe specification introduces the *zone append* command , providing the option for the host to maintain multiple outstanding I/Os in a zone. The ZNS controller writes the data at the write pointer, and returns the LBA of the write to the host. Therefore, making append especially beneficial if a large number of small writes are issued. If applications do not utilize the append command, they are required to ensure correct ordering among I/Os such that writes happen at the correct LBA, equal to the write pointer of the zone. ZNS devices additionally require the use of direct I/O, bypassing the page cache. This enforces to align with the sequential write constraint of zones, such that pages from the page cache are not written at out-of-order LBAs.
## ZNS I/O Scheduling
Adhering to the sequential write requirement with ZNS devices under multiple outstanding I/O requests requires an appropriate scheduler to be set. It is responsible for ensuring writes are submitted at consecutive LBAs on the device . However, requests can additionally be reordered on the device , making the host responsible for ensuring command ordering. For this the *mq-deadline* scheduler has to be enabled. It enforces that just a single write I/O is submitted and holds all subsequent I/Os until completion of the submitted I/O. This allows to submit numerous asynchronous write requests while adhering to the sequential write requirement and enforcing correct command ordering. Additionally, the scheduler set to *none* can also be used with ZNS devices, bypassing the Linux Kernel I/O scheduler, however this does not enforce sequential write ordering or command ordering, as I/O requests are directly issued to the device. Hence, if this scheduler is set writes have to be issued synchronously and at the correct LBA. As ZNS devices only enforce sequential write ordering, reading can be done with both schedulers asynchronously under any number of outstanding I/Os, and random or sequential accesses.
## ZNS Application Integration
With the zoned storage model providing a unified software stack for SMR and ZNS devices and support for SMR having been in numerous applications for some time, required changes for added support of ZNS devices was minimal. We focus primarily on f2fs (with f2fs-tools [^17]) , and ZenFS (commit *5ca1df7*) , a RocksDB storage backend for zoned storage devices. **f2fs:** f2fs had existing support for the ZBC/ZAC specification , making the changes for supporting ZNS devices minimal . The changes include adding the zone capacity and limiting the maximum number of active zones. f2fs manages the storage as a number of segments (typically 2MiB), of which one or more are contained in a section. Sections are the erase unit of f2fs, and segments in a section are written sequentially. The segments are aligned to the zone size, such that they do not span across zones. Since ZNS devices have a zone capacity, which is possibly smaller than the zone size, an additional segment type *unusable* was added in order identify segments outside the usable zone capacity. Partially usable segments are also supported with the *partial* segment type, in order to fully utilize the entire zone capacity if a segment is not aligned to zone capacity. The maximum active zones was already implemented by limiting the maximum number of open segments at any point in time. By default, this is set to 6, however if the device supports less active zones, f2fs will decrease this at file system creation time. While f2fs supports ZNS devices, it requires an additional randomly writable block device for metadata, which is updated in-place, as well as caching of writes prior to writing to the zoned device. **ZenFS:** ZenFS provides the file system plugin for RocksDB to utilize zoned block devices. RocksDB is an LSM-tree based persistent key-value store optimized for flash based SSDs. It works by maintaining tables at different levels in the LSM tree, of which the first level is in memory and all other levels are on the storage device. Writes initially go into the table in the first level, called the *memtable*, which gets flushed to the next level periodically or when it is full. Flushing will merge the flushed table with one from the next level, such that keys are ordered and do not overlap. This process is called *compaction*. Tables at lower levels than the memtable are referred to as Sorted String Tables (SSTs). SSTs are immutable, written sequentially, and erased as a single unit, hence making it a flash friendly architecture . With zoned storage devices the RocksDB data files need to be aligned to the zone capacity for most efficient device utilization. ZenFS maps RocksDB data files to a number of *extents*, which are contiguous regions that are written sequentially. Extents are written to a single zone, such that they do not span across multiple zones, and multiple extents can be written into one zone, depending on the extent size and zone capacity. Selection of extent placement into zones relies on the provided lifetime hints that RocksDB gives with its data files. ZenFS places an extent into a zone where the to be written extent's lifetime is smaller than the largest lifetime of the other extents in the zone, such that it is not unnecessarily delaying the resetting of a zone. ZenFS resets a zone when all the files that have extents in that particular zone have been invalidated. While RocksDB provides the option to set the maximum size for data files, data files will not have precisely this size due to compaction resulting in varying sized data files. ZenFS manages this by setting a configurable utilization percentage for the zones, which it fills up to this percentage, leaving space if files are larger than specified. While ZenFS requires at least 3 zones to run, of which one is for journaling, another is the metadata, and the last is a data zone, if the device supports more active zones, the active zones can be increased by setting a larger number of concurrent compactions in RocksDB. This can be up to the value of maximum active zones (minus the metadata and journaling zone), however performance gains for more than 12 active zones on particular ZNS devices are insignificant .
# Conclusion {#sec:conclusion}
The newly standardized ZNS device present a unique new addition to the host storage stack. Pushing the garbage collection responsibility up in the stack to the host allows for more optimal data placement, providing predictable garbage collection overheads, compared to overheads from conventional flash storage. We provide one of the first systematic studies analyzing the performance implications of ZNS integration, and present an initial set of development guidelines for ZNS devices. While we mainly focus on the block-level performance of ZNS devices, we additionally provide our unsuccessful experiments and pitfalls to avoid, and furthermore propose numerous future work ideas to evaluate ZNS device integration further and extend the guidelines provided in this work. Main findings in this evaluation show that sequential reads on ZNS devices achieve almost double the peak bandwidth of writes, and larger I/Os (\(\ge\) 16KiB) are required for fully saturating the device bandwidth. Additionally, the selection of scheduler for ZNS devices can provide workload dependent performance gains.
# Block-Level Device Performance {#sec:evaluation}
Focusing on the block-level performance of ZNS devices, we design several benchmarks using the fio benchmarking tool . In particular, we evaluate the following aspects of ZNS performance:
- **ZNS block I/O performance for the conventional namespace ([22.1](#sec:block-level_ZNS_IO_Conv_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Conv_ns"}).** We establish the baseline performance of block I/O on the ZNS device for its conventional namespace, which is exposed as a small randomly writable space. We measure the achievable throughput and compare it to performance of conventional SSDs. Main findings show that, for achieving peak write bandwidth of the device larger block sizes are required.
- **ZNS block I/O performance for the zoned namespace ([22.2](#sec:block-level_ZNS_IO_Zoned_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Zoned_ns"}).** We measure the performance of the ZNS device, benchmarking its zoned namespace. Specifically, we identify the performance of the *mq-deadline* and *none* scheduler under various read and write workloads. Results show that sequential write performance is higher with the *mq-deadline* scheduler, and read performance achieves lower median and tail latency with the scheduler set to *none*.
## Conventional Device Performance {#sec:block-level_ZNS_IO_Conv_ns}
To measure the block-level I/O performance of the conventional namespace we run several fio workloads over the namespace and compare its performance to that of the Optane and Samsung SSDs. **Device throughput:** First, we measure the performance of the Optane and Samsung SSDs. For this, we run a fio benchmark that issues 4KiB read and write I/Os (I/O size is commonly referred to as the block size throughout this section). Specifically, we run the following benchmarks; sequential write, random write, sequential read, random read, sequential overwrite, and random overwrite. The overwrite benchmarks are achieved by fully writing the entire namespace and running a sequential and random write benchmark on the full namespace. Benchmarks are repeated with varying I/O queue depths, indicating the number of outstanding I/Os to maintain . To avoid performance impact of NUMA effects on the results, we pin each workload to the NUMA node where the respective device is attached. Figures [\[fig:samsung_iops\]](#fig:samsung_iops){reference-type="ref" reference="fig:samsung_iops"} and [\[fig:optane_iops\]](#fig:optane_iops){reference-type="ref" reference="fig:optane_iops"} show the performance of the benchmarks for the Samsung SSD and the Optane based SSD, respectively. Both devices show a stable peak performance of 300KIOPs for small queue depths of 4, except for random reading on the Samsung SSD, which requires 16 outstanding I/Os to reach peak IOPs. Next, we run the same benchmarks on the conventional namespace exposed by the ZNS device. Figure [\[fig:ZNS_conv_iops\]](#fig:ZNS_conv_iops){reference-type="ref" reference="fig:ZNS_conv_iops"} shows that the ZNS device only reaches a peak of 296KIOPs for read benchmarks at deeper queue depths of 8 and 64 for sequential and random reading, respectively. Write benchmarks reach peak performance at a shallow queue depth of 2, however performance is only 19% of the peak device throughput of 296 KIOPs. **ZNS device bandwidth:** As write performance for the conventional namespace is 81% below the peak throughput of 296KIOPs, we additionally measure the achievable bandwidth for the ZNS device for larger block sizes. For this, we increase the block size to power of 2 values from 4KiB to 128KiB, and maintain a lower queue depth of 4. The resulting performance is depicted in Figure [\[fig:ZNS_scaled_bw\]](#fig:ZNS_scaled_bw){reference-type="ref" reference="fig:ZNS_scaled_bw"}, showing an increase to a peak write bandwidth of 1GiB for block sizes from 16KiB and larger. Note, the throughput in Figure [\[fig:ZNS_conv_iops\]](#fig:ZNS_conv_iops){reference-type="ref" reference="fig:ZNS_conv_iops"} reached peak performance of 296 KIOPS on sequential reading with a block size of 4KiB and a queue depth of 4, which is equivalent to \(296\text{KIOPs}*4\text{KiB}=1.13\text{GiB}\). However, with an increasing block size the bandwidth increases to a peak of 2GiB for sequential reads. Unlike the sequential read performance, the write performance only reaches a peak bandwidth of 1GiB. **Recommendations:** Based on this evaluation we can identify that for the particular ZNS device evaluated **(i)** sequential reading achieves a 94.2% larger peak bandwidth than write performance at block size \(\geq\) 16KiB, (**ii**) peak throughput of 296KIOPs for sequential reading is reached at lower queue depth of 4, while random reading requires deeper queues \(\geq 64\) to achieve the same throughput, and **(iii)** for achieving peak write bandwidth of the device larger block sizes (\(\geq\) 16KiB) are required.
## Zoned Device Performance {#sec:block-level_ZNS_IO_Zoned_ns}
These benchmarks focus purely on the zoned namespace performance, and quantify the overheads of using *mq-deadline* and *none* schedulers with ZNS devices. The *mq-deadline* scheduler can utilize a higher I/O queue depth (\(>1\)), as it holds back I/Os and only submits a single I/O at a time, and with *none* the host needs to ensure the I/O queue depth is equal to one. **Read performance:** First, we measure the performance of sequential and random reading with both schedulers. Recall that ZNS does not enforce reading constraints, and thus both schedulers can have any number of outstanding read requests, with sequential or random accesses. The sequential read benchmark is configured to issue 4KiB read I/Os in a single zone, under both schedulers, and an increasing I/O queue depth, ranging from 1-14. Results presented in Figure [\[fig:ZNS_concur_read_seq_iodepth\]](#fig:ZNS_concur_read_seq_iodepth){reference-type="ref" reference="fig:ZNS_concur_read_seq_iodepth"} show that the scheduler set to *none* achieves a 9.95% lower median and 9.66% lower tail latency at an I/O queue depth of 14. This is due to the added overhead of having the *mq-deadline* scheduler compared to *none* bypassing the Linux I/O scheduler. Next, we benchmark random read performance. However, as sequential reading in a single zoned showed that bypassing the Linux I/O scheduler provides lower median and tail latency, we measure random reading by utilizing multiple zones for the *none* scheduler. Specifically, *mq-deadline* is set up with the same configuration as with sequential reading, issuing 4KiB I/Os in a single zone with an increasing I/O queue depth, while *none* is set to issue a single I/O to a zone with an increasing number of concurrent threads (also ranging from 1-14, up to the maximum number of active zones). Thus, both schedulers have a particular number of outstanding I/Os (x-axis). Results shown in Figure [\[fig:ZNS_concur_read_rand\]](#fig:ZNS_concur_read_rand){reference-type="ref" reference="fig:ZNS_concur_read_rand"} show that up to 4 outstanding I/Os *mq-deadline* has a higher median latency, ranging between 1.23-4.79% higher than median latency of *none*. However, for more outstanding I/Os (\(> 4\)) both schedulers similar performance. **Write performance:** Next, we measure the performance of the ZNS device under write workloads with both schedulers. ZNS devices enforce write constraints, therefore only *mq-deadline* can utilize I/O queue depths \(> 1\). We first benchmark the performance of both schedulers by issuing a single 4KiB sequential write I/O in a zone with an increasing number of concurrent threads split across available zones on the device. The number of concurrent threads ranges from 1-14, equivalent to the maximum number of active zones on the device. Figure [\[fig:ZNS_concur_write_seq\]](#fig:ZNS_concur_write_seq){reference-type="ref" reference="fig:ZNS_concur_write_seq"} shows that the resulting performance of both schedulers is nearly identical over all numbers of outstanding I/O requests. Next, we measure the performance of *mq-deadline* issuing I/Os in a single zone and increasing the I/O queue depth, rather than splitting individual I/Os concurrently across zones. The benchmark with the *none* scheduler is the same as for the prior write benchmark, namely issuing 4KiB I/Os concurrently with an increasing number of threads split across the active zones. This allows both benchmarks to have a specific number of outstanding I/O requests (x-axis). Figure [\[fig:ZNS_concur_write_seq_iodepth\]](#fig:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="fig:ZNS_concur_write_seq_iodepth"} shows that the scheduler set to *mq-deadline* achieves a 72.74% lower median and 83.27% lower tail latency. The lower latency for *mq-deadline* as the number of outstanding I/Os increases is due to the merging of I/O requests into larger I/Os, as all the I/Os are at consecutive LBAs. This allows it to issue overall less I/Os than with *none* scheduling in this configuration. **Recommendations:** With this evaluation we can identify that **(i)** random read heavy workloads should avoid the Linux I/O scheduler by setting it to *none*, providing up to 9.95% lower median latency, and **(ii)** multiple outstanding write I/Os should utilize the *mq-deadline* scheduler to merge I/Os in a single zone, rather than splitting I/Os concurrently over multiple zones.
# Experimental Setup {#sec:exp_setup}
For the experiments we utilize a ZNS device, whose details and properties are depicted in Table [\[tab:SSD_architecture\]](#tab:SSD_architecture){reference-type="ref" reference="tab:SSD_architecture"}. The initial goal of this evaluation was to evaluate all possible integrations of ZNS devices. For this we establish the following configuration: **f2fs.** The f2fs (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(c)) parameters are mostly kept at its default options, with the only change being the enforcing of 10% overprovisioning. Since f2fs requires a randomly writable device for its metadata, we expose 4 of the zones on the ZNS device as randomly writable space, corresponding to the maximum amount that the ZNS device can expose as randomly writable space. This space is then used by f2fs for metadata and write caching. However, this requires the zoned space of f2fs (where the actual file system data will be) to align with the size of the randomly writable space, i.e., the randomly writable space has to be large enough to fit all the metadata for the file system. Therefore, the resulting largest possible size that successfully formats the f2fs file system on the zoned space is 100GiB. As a result, we create a 100GiB namespace from the available zoned capacity for the f2fs file system, which we utilize for all experiments. We do not use an additional larger randomly writable device, as we aim to avoid performance implications of multiple devices, which would make it difficult to differentiate between performance effects from the ZNS device and the additional block device. **ZenFS.** ZenFS (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(d)) requires an auxiliary path for its metadata to store LOG and LOCK files for RocksDB at runtime. With this it additionally allows to backup and recover a file system through its command line options, however we do not utilize this option in our evaluation. For the auxiliary path we use the 4GiB randomly writable space exposed by the ZNS device and create a f2fs file system on it which is mounted for solely the ZenFS auxiliary path to be placed on. As the largest zoned space that successfully formats f2fs is 100GiB, we utilize a 100GiB namespace for all experiments, alongside the 4GiB randomly writable namespace, and all remaining capacity is left in an unused namespace. Throughout the evaluation we refer to the 4GiB randomly writable space exposed by the ZNS device as the *conventional namespace*, and the namespace containing the zoned storage on the ZNS device is referred to as the *zoned namespace*. Lastly, since the ZNS specification was integrated in Linux Kernel 5.9.0+, we use a later Kernel version 5.12.0.
# Future Work {#sec:future_work}
With ZNS devices having just been introduced, they leave a plethora of avenues to explore. In particular, as we showcased in Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}, there are numerous possibilities of integrating ZNS devices into the host software stack. In this evaluation we focus mainly on the block-level ZNS performance and provide several initial guidelines for developers. Expansion of these is left as future work by exploring the additional levels of integration. Specifically this includes evaluating the following aspects of ZNS devices:
- What is the performance of the varying levels of integration for ZNS devices in terms of achievable IOPs, latency, transactions/sec, and additional instructions required? Under sequential and random, read and write workloads, are there performance implications of a specific integration, and what rules can be established when building applications for a particular integration?
- How is garbage collection influenced by the different levels of integration? Does building application specific garbage collection policies provide superior performance over other levels of integration? In particular how does an almost fully utilized device (e.g., 95% utilization) affect garbage collection performance at the different levels?
- Do multiple applications running on the same ZNS device interfere with each other? This includes a shared device with multiple namespaces, shared block-level interface with for example multiple concurrently running file systems, and lastly a shared file system on ZNS with multiple concurrent applications. Does in each case an application's garbage collection impact the performance of concurrently running applications?
As we evaluated some of these aspects and failed to produce insightful results, we propose to evaluate them by taking into account our shortcomings and considering the assumptions we provide in Section [28.2](#sec:pitfalls_to_avoid){reference-type="ref" reference="sec:pitfalls_to_avoid"}. An additional exploration that became apparent during this study was that there is currently no enforcing on the number of active zones at a time. The device only allows a maximum number of zones to be active at any point in time, however there is no managing of how active zones are split across namespaces on the device, or across applications running on the same namespace. Especially, as the device supports a larger number of namespaces to be created than active zones that can be open, a zone manager is required to assign zones across namespaces and applications, and provide fair resource sharing across namespaces and applications, enforcing that the maximum number is not exceeded, even if there are more concurrent applications running. Lastly, we suggest the exploration of file system improvements for f2fs, and other ZNS specific file systems, with grouping of files by creation time, death time, or owner as was suggested in and similarly evaluated for conventional SSDs in . This would provide the possibility for optimizing garbage collection and improving the device performance at the file system integration level.
# Introduction
The introduction of flash storage provided significant changes in the storage hierarchy. Achieving as low as single digit \(\mu\)-second latency, several GB/s of bandwidth, and millions of I/O operations per second (IOPS) , they offer significant performance gains over prior storage technologies, such as Hard Disk Drives (HDDs). Flash storage is organized in pages (typically 16KiB in size) , representing the unit of read and write accesses, of which multiple pages are combined into a block (typically multiple MiB in size). Blocks are further packed in planes and dies to manage data and control connectivity to the host. Flash pages do not support in-place updates. As a result, pages have to be erased prior to being written again. However, erase operations require substantially more time than read and write operations . Therefore, erase operations are done at block granularity to amortize the erase overhead. Additionally, flash storage requires pages within a block to be written sequentially. Flash storage therefore includes complex firmware, called the Flash Translation Layer (FTL), to provide the seemingly in-place updates of data and hide the sequential write constraints of devices by exposing a sector/page-addressable SSD . Furthermore, a crucial task of the FTL is to run *garbage collection* (GC), in order to erase blocks with invalid pages and free up space. For this, the FTL reads out valid pages of data from a block and relocates the data to a free block, followed by erasing of the original block. Garbage collection is triggered periodically, or when the device is running low on free space to write data. The resulting interface exposed by conventional flash storage allows it to mimic the behavior of HDDs, thus requiring no changes in the host storage software to access the underlying flash storage. However, recent research results made evident that hiding the flash management complexities from the host leads to suboptimal data placement, unpredictable performance overheads, and shortens the lifetime of flash devices . Therefore, researchers have proposed to open the flash storage interface and expose device internals to the host. This allows for the host to optimize storage management with workload-specific decisions . Zoned Namespace (ZNS) SSDs are the latest addition in these efforts, which are now standardized in the NVMe 2.0 specification and are commercially available . In order to better match the underlying properties of flash chips, ZNS exposes the address space with numerous *zones*, where each zone requires append-only sequential writes. Zones are aligned to the erase unit of a block. With this new interface, the host storage software is now responsible for resetting a zone (i.e., trigger garbage collection) after which a zone becomes writable again from the starting address. Apart from the write restrictions, there are operational parameters such as the zone capacity, the maximum number of active zones limits, and the append limits, which the host software must be aware of. The new interface of these devices necessitates changes to the host storage stack. However, there is more than one way these devices can be integrated within systems. In a classical setup as shown in Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(a), at the bottom is a block device with a file system on top of it, and an application running on top of the file system. Integrating ZNS devices into this storage stack can be done in three different configurations. Firstly, integration at the block-device level (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(b)), while keeping the rest of the stack above the same. An example of this setup is the *dm-zap* zoned device mapper project within the Linux Kernel . This way of integration is the least intrusive one and requires the minimum amount of changes to anything running on top of the block device. Secondly, integration at the file system level (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(c)). For this integration a file system is made aware of the zoned device characteristics such as zone capacity, number of active zones limits, and append limits. Example of such a project is the added ZNS support in f2fs . With such integration the knowledge and required changes for ZNS devices are pushed higher in the stack, from block-device level to the file system level. Lastly, ZNS-aware application integration (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(d)). In this case, there are ZNS-specific application-level changes at the very top of the storage stack. By pushing the customization higher up in the stack, the expectation is to deliver better performance together with the best case application-specific customization and integration with ZNS devices. An example of such an integration is the ZenFS file system module for RocksDB and MySQL . With such configuration possibilities, it is not immediately clear which integration one should choose for their workload. In this research work, we aim to systematically understand the impact of the ZNS integration in the host storage system. This work is largely inspired by related work that has provided unwritten contracts of storage devices for Optane based SSDs and flash based SSDs with guidelines for developers to optimize storage performance. Before we can synthesize actionable design guidelines for storage stack developers, in this work, we first start with systematic benchmarking in the presence of the OS I/O scheduler. In particular, we make the following contributions:
- We provide information on the newly standardized NVMe ZNS devices, how these devices work, how they are integrated into the host storage stack, and how existing applications are modified to support ZNS devices.
- We measure the block-level ZNS device performance, comparing it to conventional block devices in terms of achievable IOPs and bandwidth, and benchmark the possible scheduler configurations for ZNS devices, depicting their implications and limitations.
- We present pitfalls and failed experiments during this evaluation in an effort for others to learn from and to avoid the obstacles we encountered.
- We provide a set of initial guidelines for optimizing ZNS integration into systems, and propose several future work ideas to further explore ZNS integration implications and expand our initial set of guidelines.
- All collected datasets and benchmarking scripts are made publicly available at <https://github.com/nicktehrany/ZNS-Study>. The appendix provides more detailed setup and benchmarking information.
The remainder of the paper is organized as follows. Section [20](#sec:background){reference-type="ref" reference="sec:background"} provides background information on ZNS devices, and their integration into systems. Next, Sections [23](#sec:exp_setup){reference-type="ref" reference="sec:exp_setup"} explains the experimental setup, followed by the first round of unsuccessful experiments in Section [28](#sec:unsuccessful_exps){reference-type="ref" reference="sec:unsuccessful_exps"}. Based on this we provide an adapted experimental setup in Section [7](#sec:adapted_setup){reference-type="ref" reference="sec:adapted_setup"}, and Section [22](#sec:evaluation){reference-type="ref" reference="sec:evaluation"} presents the various benchmarks. Lastly, Section [27](#sec:related){reference-type="ref" reference="sec:related"} provides the related work, followed by future work ideas in Section [24](#sec:future_work){reference-type="ref" reference="sec:future_work"}, and Section [21](#sec:conclusion){reference-type="ref" reference="sec:conclusion"} concludes the paper.
# Related Work {#sec:related}
While ZNS has just recently been standardized , there have been several initial evaluations and discussions of ZNS devices. Bjorling et al. present modifications made to RocksDB and f2fs to support ZNS devices, and showcase the performance gains for these devices. Stavrinos et al. provide an initial discussion on the benefits of ZNS devices and possible improvements for applications on them. Shin et al. show a performance study of ZNS devices, depicting the need for large request sizes for achieving increased on-device parallelism. Similar to the results in our study, where the conventional namespace required larger block sizes (\(\geq\) 16KiB) to reach peak device bandwidth. Han et al. provide a discussion on the ZNS interface, and propose an improved interface particularly optimized for log-structured file systems with segment compaction. However, there currently is no work that systematically studies the possible ways to integrate ZNS devices into the host software stack. We are the first to present a start at such a study. As the ZBD model was originally introduced for SMR devices, there have been several performance evaluations of SMR devices and their integration into systems. Wu et al. showcase an extensive evaluation of host-aware SMR drives, which similarly to ZNS device expose device characteristics to the host. In particular, the authors focus on evaluating performance characteristics of the zoned interface. Additionally, Wu et al. provide a performance study on implications of different properties of host-aware SMR devices, including performance implications under the number of open zones. While not focusing on ZNS devices, past work presented similar evaluations for conventional SSDs, where He et al. provide an unwritten contract for flash-based SSDs, depicting numerous guidelines on performance improvements for such devices. Similarly, Wu et al. present such an unwritten contract for Optane SSDs. Both of these unwritten contracts were inspiration for us to provide a set of developer guidelines for the new ZNS devices. Yang et al. characterize performance implications of building log-structured applications for flash-based SSDs, showcasing the negated benefits of optimizing application data structures for flash, caused by the device characteristics. Such an evaluation showcases application-level integration, which presents insightful results that should be reproduced on ZNS devices to further expand the developer guidelines we provide.
# Unsuccessful Experiments {#sec:unsuccessful_exps}
Designing of experiments to evaluate the performance of the varying levels of integration proved challenging, as initial experiments did not provide insightful results. We provide the iterations of experimental design and why experiments failed in an effort for others evaluating ZNS performance to avoid these pitfalls. Section [28.1](#sec:failed_setup){reference-type="ref" reference="sec:failed_setup"} provides the configurations of the various benchmark, followed by Section [28.2](#sec:pitfalls_to_avoid){reference-type="ref" reference="sec:pitfalls_to_avoid"} describing the failures of these benchmarks and pointing out possible causes for this. Lastly, we give additional lessons learned during this evaluation in Section [28.3](#sec:lessons_learned){reference-type="ref" reference="sec:lessons_learned"}.
## Benchmark Setup {#sec:failed_setup}
We run several benchmarks to evaluate different performance aspects of ZNS devices. While each of the evaluations relies on db_bench and RocksDB, they have slightly different benchmarking configurations. Below we describe the workload parameters for the different benchmarks.
### Integration Level
We first benchmark the different possible integration levels, as shown in Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}. Benchmarks initially fill the database in random and sequential key order, followed by reading from the database in random and sequential key order. We additionally run an overwrite benchmark, which overwrites existing keys in random order, and an updaterandom workload that modifies values of random keys. The overwrite benchmark and the updaterandom differ in the way values are accessed and modified. Overwrite issues asynchronous writes to the key, whereas updaterandom uses a read-modify-write approach. Keys are 16B each, values are configured to be 100B, and we disable any data compression. As mentioned, ZNS devices require direct I/O, we therefore set appropriate db_bench flags to issue direct reads and use direct I/O for flushing and compaction.
### ZenFS Benchmark {#sec:zenfs_exp}
Next, to verify correctness of our setup we attempt to repeat an experiment depicted in , comparing db_bench performance on a configuration with f2fs to a configuration with ZenFS. The focus of this benchmark is write intensive workloads, as this is meant to trigger increased garbage collection and showcase gains of managing ZNS at the application-level with ZenFS. The benchmark is configured to run fillrandom and overwrite with a key size of 100B and value size of 800B. We additionally use data compression to compress data down to 400B. The original paper uses the entire device for its benchmark, however we scale it to the maximum possible that successfully fits into the namespace, which is equivalent to 50 million keys. Lastly, we set the target file size of SST files to be equal to the zone capacity, and again utilize appropriate flags for direct I/O.
### Multi-Tenancy
Lastly, we run an experiment to evaluate how multiple concurrently running namespaces affect the performance of one another. For this we mount f2fs on the 100GiB zoned namespace and create an additional 100GiB namespace on which ZenFS with a db_bench benchmark is running. We compare the performance of running the ZenFS namespace alone, without any interference from another namespace, to running the f2fs namespace concurrently. The ZenFS namespace runs the same workload as describe in the previous benchmark (Section [28.1.2](#sec:zenfs_exp){reference-type="ref" reference="sec:zenfs_exp"}). The goal being that the f2fs namespace creates substantial device traffic through write I/Os, especially during garbage collection from the overwrite benchmark, and thus show the performance impact on the ZenFS namespace.
## Pitfalls to Avoid {#sec:pitfalls_to_avoid}
Results of all experiments showed little to no performance difference in their benchmarks. Especially the ZenFS benchmark (Section [28.1.2](#sec:zenfs_exp){reference-type="ref" reference="sec:zenfs_exp"}), where the original paper showed substantial performance gains with ZenFS, in particular on overwrite benchmarks where GC is being triggered heavily. We failed to reproduce these exact results and only had minor performance gains from ZenFS. The multi-tenant evaluation showed very similar results, which appear contrary to prior expectations. That is if one namespace fully utilizes the device, then another namespace attempting to use the same device will have some performance implications, as they would now be sharing the device resources. As the prior experiments proved ineffective in their evaluation, we propose the assumption that for performance differences to appear on the ZNS device, it has to be largely utilized, such that LBA mappings are fully setup and more garbage collection is triggered. If the device is not largely utilized, LBA mappings are not fully setup and the device is able to provide peak performance without showing effects of garbage collection, as there is a large amount of free space it can utilize. Additionally, the device bandwidth has to be utilized to the extent that it competes with the garbage collection happening in the background. We therefore believe that with our setup we failed to produce enough device utilization in bandwidth and space, and thus evaluations showed that there are no performance differences. While we did not evaluate all configurations under increased garbage collection and device utilization, we believe it to have an effect on performance and thus suggest its evaluation as future work (discussed in detail in Section [24](#sec:future_work){reference-type="ref" reference="sec:future_work"}). Per suggestions of Western Digital, for significant performance advantages to appear for ZNS devices, utilization of available storage capacity should be \(>50\%\) (preferably \(80\%+\)) and bandwidth utilization of \(>30\%\) of the device bandwidth, such that the write workload competes with ongoing garbage collection. However, we have not evaluated these specific configurations and thus leave it as future work.
## Lessons Learned {#sec:lessons_learned}
In addition to unsuccessful experiments we encountered several obstacles that were not immediately obvious to debug. Again, we provide our experiences in order for others to avoid these pitfalls. Firstly, by default the ZNS devices do not set a scheduler, neither do the applications such as f2fs or ZenFS, nor is there an error on an invalid scheduler being set. Thus setting up of the applications and formatting the ZNS device completes successfully, while as soon as writes are issued to the device I/O errors appear. Therefore, it is important to ensure the correct scheduler is always set, that is the *mq-deadline* scheduler, and every namespace requires it to be set individually after creation. The majority of applications utilize multiple outstanding I/Os per zone, thus the *mq-deadline* scheduler is required, as it will hold back I/Os such that only one outstanding I/O is submitted to the device at any point in time. If the application issues a single I/O synchronously, the scheduler could be left at the default configuration set to *none* (default setting in Linux 5.12.0), however here again the application must enforce the sequential write constraint. Secondly, device mapper support is not there yet. A device mapper implements a host-side FTL that makes the sequential write required zones randomly writable by exposing the zoned device as a conventional block device. It controls data placement, maintains data mappings, and runs garbage collection, just as the FTL on traditional flash storage. An existing implementation, such as *dm-zoned* is ZBC and ZAC compliant but not compliant to the new concepts of zone capacity from the ZNS specification. The *dm-zap* device mapper aims to be ZNS compliant, however it is still a prototype and not fully functional with ZNS devices yet. As a result, we are currently not able to evaluate performance of one level of integration (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(b)), requiring a future evaluation when the device mapper support is functional.
# Background {#sec:background}
Applications and file systems rely on the Linux block layer to provide interfaces and abstractions for accessing the underlying storage media. Originally designed to match the hardware characteristics of HDDs the block layer presents the storage as a linear address space, allowing for sequential and random writes. Flash storage however has different write constraints due to its architecture. Relying on the FTL to hide the device management idiosyncrasies however leads to negative performance impacts due to the unpredictable performance from the device garbage collection , large tail latency it causes , and increased write amplification . The increased write amplification additionally reduces the device lifetime, since flash cells on SSDs have limited program/erase cycles. Furthermore, garbage collection requires the device to maintain a certain amount of free space, called the *overprovisioning space*, such that the FTL is able to move valid pages. Most commonly used overprovisioning takes between 10-28% of the device capacity. One of the possible FTL design uses a fully-associative mapping of host logical block addresses (LBAs) to physical addresses in order to provide the LBAs of the page(s) that contain valid data. Such a design requires significant resources in order to store all mappings.
## Zoned Storage {#sec:zoned_storage}
The arrival of ZNS SSDs eliminates the need for on device garbage collection done by the FTL, pushing this responsibility to the host. This provides the host with more opportunity for optimized data placement, through mechanisms such as data grouping, and makes garbage collection overheads predictable. The concept of exposing storage as zones is not new, as it was already introduced when Shingled Magnetic Recording (SMR) HDDs appeared, which also enforce a sequential write constraint. The zoned storage model was established with the addition of Zoned Block Device (ZBD) support in the Linux Kernel 4.10.0 , in an effort to avoid the mismatch between the block layer and the sequential write constraint on devices. The standards defining the management of SMR HDDs in the zoned storage model came through the Zoned Device ATA Command Set (ZAC) and the Zoned Block Command (ZBC) specifications. Since zones require sequential writing, the address of the current write is managed with a *write pointer*. The write pointer is incremented to the LBA of the next write only after a successful write. In order to manage zones, each zone has a state associated to it. These are to identify the condition of a zone, which can be any of the following; *EMPTY* to indicate a zone is empty, *OPEN* which is required for a zone to allow writes, *FULL* to indicate the write pointer is at the zone capacity, *CLOSED* which is used to release resources for the zone (e.g., write buffers on the ZNS device) without resetting a zone, this additionally does not allow writes to continue in the zone until it is transitioned to *OPEN* again, *OFFLINE* which makes all data in the zone inaccessible until the zone is reset, and *READ-ONLY*. The command sets provide the proper mechanisms to transition zones between any of the states. While the majority of available space for both SMR and ZNS devices is utilized by sequential write required zones, they can also expose a limited amount of randomly writable area. This is mainly intended for metadata as this space only occupies a very small percentage of the device capacity. For example, the ZNS device used in this evaluation could expose 4 zones as randomly writable, equivalent to approximately 4GiB, compared to the total device size of 7.2TiB. The newly introduced ZNS SSDs are standardized through the NVMe specification , which builds on the foundations established with ZBD support. While the zoned storage model aims to provide a unified software stack for all zoned devices (SMR and ZNS devices), the NVMe specification introduces several new concepts particular to ZNS devices. Firstly, it defines a *zone capacity* for each zone, stating the usable capacity of a zone, which is less than or equal to the size of the zone. Figure [\[fig:ZNS_HW\]](#fig:ZNS_HW){reference-type="ref" reference="fig:ZNS_HW"} shows an example layout of zones on a ZNS SSD and each zone's associated state. Zone capacity is kept separate from the zone size such that the capacity can be aligned to the erase unit of the underlying flash media, and such that the zone size can be kept as a power of two value. This is required for easier conversions between LBAs and zone offsets. Secondly, the *active zones* limit specifies the maximum number of zones that can be active (in *OPEN* or *CLOSED* state) at any point in time. As the SSD requires to allocate resources for each active zone, such as the write buffers, it enforces a maximum on the active zones. Lastly, the NVMe specification introduces the *zone append* command , providing the option for the host to maintain multiple outstanding I/Os in a zone. The ZNS controller writes the data at the write pointer, and returns the LBA of the write to the host. Therefore, making append especially beneficial if a large number of small writes are issued. If applications do not utilize the append command, they are required to ensure correct ordering among I/Os such that writes happen at the correct LBA, equal to the write pointer of the zone. ZNS devices additionally require the use of direct I/O, bypassing the page cache. This enforces to align with the sequential write constraint of zones, such that pages from the page cache are not written at out-of-order LBAs.
## ZNS I/O Scheduling
Adhering to the sequential write requirement with ZNS devices under multiple outstanding I/O requests requires an appropriate scheduler to be set. It is responsible for ensuring writes are submitted at consecutive LBAs on the device . However, requests can additionally be reordered on the device , making the host responsible for ensuring command ordering. For this the *mq-deadline* scheduler has to be enabled. It enforces that just a single write I/O is submitted and holds all subsequent I/Os until completion of the submitted I/O. This allows to submit numerous asynchronous write requests while adhering to the sequential write requirement and enforcing correct command ordering. Additionally, the scheduler set to *none* can also be used with ZNS devices, bypassing the Linux Kernel I/O scheduler, however this does not enforce sequential write ordering or command ordering, as I/O requests are directly issued to the device. Hence, if this scheduler is set writes have to be issued synchronously and at the correct LBA. As ZNS devices only enforce sequential write ordering, reading can be done with both schedulers asynchronously under any number of outstanding I/Os, and random or sequential accesses.
## ZNS Application Integration
With the zoned storage model providing a unified software stack for SMR and ZNS devices and support for SMR having been in numerous applications for some time, required changes for added support of ZNS devices was minimal. We focus primarily on f2fs (with f2fs-tools [^18]) , and ZenFS (commit *5ca1df7*) , a RocksDB storage backend for zoned storage devices. **f2fs:** f2fs had existing support for the ZBC/ZAC specification , making the changes for supporting ZNS devices minimal . The changes include adding the zone capacity and limiting the maximum number of active zones. f2fs manages the storage as a number of segments (typically 2MiB), of which one or more are contained in a section. Sections are the erase unit of f2fs, and segments in a section are written sequentially. The segments are aligned to the zone size, such that they do not span across zones. Since ZNS devices have a zone capacity, which is possibly smaller than the zone size, an additional segment type *unusable* was added in order identify segments outside the usable zone capacity. Partially usable segments are also supported with the *partial* segment type, in order to fully utilize the entire zone capacity if a segment is not aligned to zone capacity. The maximum active zones was already implemented by limiting the maximum number of open segments at any point in time. By default, this is set to 6, however if the device supports less active zones, f2fs will decrease this at file system creation time. While f2fs supports ZNS devices, it requires an additional randomly writable block device for metadata, which is updated in-place, as well as caching of writes prior to writing to the zoned device. **ZenFS:** ZenFS provides the file system plugin for RocksDB to utilize zoned block devices. RocksDB is an LSM-tree based persistent key-value store optimized for flash based SSDs. It works by maintaining tables at different levels in the LSM tree, of which the first level is in memory and all other levels are on the storage device. Writes initially go into the table in the first level, called the *memtable*, which gets flushed to the next level periodically or when it is full. Flushing will merge the flushed table with one from the next level, such that keys are ordered and do not overlap. This process is called *compaction*. Tables at lower levels than the memtable are referred to as Sorted String Tables (SSTs). SSTs are immutable, written sequentially, and erased as a single unit, hence making it a flash friendly architecture . With zoned storage devices the RocksDB data files need to be aligned to the zone capacity for most efficient device utilization. ZenFS maps RocksDB data files to a number of *extents*, which are contiguous regions that are written sequentially. Extents are written to a single zone, such that they do not span across multiple zones, and multiple extents can be written into one zone, depending on the extent size and zone capacity. Selection of extent placement into zones relies on the provided lifetime hints that RocksDB gives with its data files. ZenFS places an extent into a zone where the to be written extent's lifetime is smaller than the largest lifetime of the other extents in the zone, such that it is not unnecessarily delaying the resetting of a zone. ZenFS resets a zone when all the files that have extents in that particular zone have been invalidated. While RocksDB provides the option to set the maximum size for data files, data files will not have precisely this size due to compaction resulting in varying sized data files. ZenFS manages this by setting a configurable utilization percentage for the zones, which it fills up to this percentage, leaving space if files are larger than specified. While ZenFS requires at least 3 zones to run, of which one is for journaling, another is the metadata, and the last is a data zone, if the device supports more active zones, the active zones can be increased by setting a larger number of concurrent compactions in RocksDB. This can be up to the value of maximum active zones (minus the metadata and journaling zone), however performance gains for more than 12 active zones on particular ZNS devices are insignificant .
# Conclusion {#sec:conclusion}
The newly standardized ZNS device present a unique new addition to the host storage stack. Pushing the garbage collection responsibility up in the stack to the host allows for more optimal data placement, providing predictable garbage collection overheads, compared to overheads from conventional flash storage. We provide one of the first systematic studies analyzing the performance implications of ZNS integration, and present an initial set of development guidelines for ZNS devices. While we mainly focus on the block-level performance of ZNS devices, we additionally provide our unsuccessful experiments and pitfalls to avoid, and furthermore propose numerous future work ideas to evaluate ZNS device integration further and extend the guidelines provided in this work. Main findings in this evaluation show that sequential reads on ZNS devices achieve almost double the peak bandwidth of writes, and larger I/Os (\(\ge\) 16KiB) are required for fully saturating the device bandwidth. Additionally, the selection of scheduler for ZNS devices can provide workload dependent performance gains.
# Block-Level Device Performance {#sec:evaluation}
Focusing on the block-level performance of ZNS devices, we design several benchmarks using the fio benchmarking tool . In particular, we evaluate the following aspects of ZNS performance:
- **ZNS block I/O performance for the conventional namespace ([22.1](#sec:block-level_ZNS_IO_Conv_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Conv_ns"}).** We establish the baseline performance of block I/O on the ZNS device for its conventional namespace, which is exposed as a small randomly writable space. We measure the achievable throughput and compare it to performance of conventional SSDs. Main findings show that, for achieving peak write bandwidth of the device larger block sizes are required.
- **ZNS block I/O performance for the zoned namespace ([22.2](#sec:block-level_ZNS_IO_Zoned_ns){reference-type="ref" reference="sec:block-level_ZNS_IO_Zoned_ns"}).** We measure the performance of the ZNS device, benchmarking its zoned namespace. Specifically, we identify the performance of the *mq-deadline* and *none* scheduler under various read and write workloads. Results show that sequential write performance is higher with the *mq-deadline* scheduler, and read performance achieves lower median and tail latency with the scheduler set to *none*.
## Conventional Device Performance {#sec:block-level_ZNS_IO_Conv_ns}
To measure the block-level I/O performance of the conventional namespace we run several fio workloads over the namespace and compare its performance to that of the Optane and Samsung SSDs. **Device throughput:** First, we measure the performance of the Optane and Samsung SSDs. For this, we run a fio benchmark that issues 4KiB read and write I/Os (I/O size is commonly referred to as the block size throughout this section). Specifically, we run the following benchmarks; sequential write, random write, sequential read, random read, sequential overwrite, and random overwrite. The overwrite benchmarks are achieved by fully writing the entire namespace and running a sequential and random write benchmark on the full namespace. Benchmarks are repeated with varying I/O queue depths, indicating the number of outstanding I/Os to maintain . To avoid performance impact of NUMA effects on the results, we pin each workload to the NUMA node where the respective device is attached. Figures [\[fig:samsung_iops\]](#fig:samsung_iops){reference-type="ref" reference="fig:samsung_iops"} and [\[fig:optane_iops\]](#fig:optane_iops){reference-type="ref" reference="fig:optane_iops"} show the performance of the benchmarks for the Samsung SSD and the Optane based SSD, respectively. Both devices show a stable peak performance of 300KIOPs for small queue depths of 4, except for random reading on the Samsung SSD, which requires 16 outstanding I/Os to reach peak IOPs. Next, we run the same benchmarks on the conventional namespace exposed by the ZNS device. Figure [\[fig:ZNS_conv_iops\]](#fig:ZNS_conv_iops){reference-type="ref" reference="fig:ZNS_conv_iops"} shows that the ZNS device only reaches a peak of 296KIOPs for read benchmarks at deeper queue depths of 8 and 64 for sequential and random reading, respectively. Write benchmarks reach peak performance at a shallow queue depth of 2, however performance is only 19% of the peak device throughput of 296 KIOPs. **ZNS device bandwidth:** As write performance for the conventional namespace is 81% below the peak throughput of 296KIOPs, we additionally measure the achievable bandwidth for the ZNS device for larger block sizes. For this, we increase the block size to power of 2 values from 4KiB to 128KiB, and maintain a lower queue depth of 4. The resulting performance is depicted in Figure [\[fig:ZNS_scaled_bw\]](#fig:ZNS_scaled_bw){reference-type="ref" reference="fig:ZNS_scaled_bw"}, showing an increase to a peak write bandwidth of 1GiB for block sizes from 16KiB and larger. Note, the throughput in Figure [\[fig:ZNS_conv_iops\]](#fig:ZNS_conv_iops){reference-type="ref" reference="fig:ZNS_conv_iops"} reached peak performance of 296 KIOPS on sequential reading with a block size of 4KiB and a queue depth of 4, which is equivalent to \(296\text{KIOPs}*4\text{KiB}=1.13\text{GiB}\). However, with an increasing block size the bandwidth increases to a peak of 2GiB for sequential reads. Unlike the sequential read performance, the write performance only reaches a peak bandwidth of 1GiB. **Recommendations:** Based on this evaluation we can identify that for the particular ZNS device evaluated **(i)** sequential reading achieves a 94.2% larger peak bandwidth than write performance at block size \(\geq\) 16KiB, (**ii**) peak throughput of 296KIOPs for sequential reading is reached at lower queue depth of 4, while random reading requires deeper queues \(\geq 64\) to achieve the same throughput, and **(iii)** for achieving peak write bandwidth of the device larger block sizes (\(\geq\) 16KiB) are required.
## Zoned Device Performance {#sec:block-level_ZNS_IO_Zoned_ns}
These benchmarks focus purely on the zoned namespace performance, and quantify the overheads of using *mq-deadline* and *none* schedulers with ZNS devices. The *mq-deadline* scheduler can utilize a higher I/O queue depth (\(>1\)), as it holds back I/Os and only submits a single I/O at a time, and with *none* the host needs to ensure the I/O queue depth is equal to one. **Read performance:** First, we measure the performance of sequential and random reading with both schedulers. Recall that ZNS does not enforce reading constraints, and thus both schedulers can have any number of outstanding read requests, with sequential or random accesses. The sequential read benchmark is configured to issue 4KiB read I/Os in a single zone, under both schedulers, and an increasing I/O queue depth, ranging from 1-14. Results presented in Figure [\[fig:ZNS_concur_read_seq_iodepth\]](#fig:ZNS_concur_read_seq_iodepth){reference-type="ref" reference="fig:ZNS_concur_read_seq_iodepth"} show that the scheduler set to *none* achieves a 9.95% lower median and 9.66% lower tail latency at an I/O queue depth of 14. This is due to the added overhead of having the *mq-deadline* scheduler compared to *none* bypassing the Linux I/O scheduler. Next, we benchmark random read performance. However, as sequential reading in a single zoned showed that bypassing the Linux I/O scheduler provides lower median and tail latency, we measure random reading by utilizing multiple zones for the *none* scheduler. Specifically, *mq-deadline* is set up with the same configuration as with sequential reading, issuing 4KiB I/Os in a single zone with an increasing I/O queue depth, while *none* is set to issue a single I/O to a zone with an increasing number of concurrent threads (also ranging from 1-14, up to the maximum number of active zones). Thus, both schedulers have a particular number of outstanding I/Os (x-axis). Results shown in Figure [\[fig:ZNS_concur_read_rand\]](#fig:ZNS_concur_read_rand){reference-type="ref" reference="fig:ZNS_concur_read_rand"} show that up to 4 outstanding I/Os *mq-deadline* has a higher median latency, ranging between 1.23-4.79% higher than median latency of *none*. However, for more outstanding I/Os (\(> 4\)) both schedulers similar performance. **Write performance:** Next, we measure the performance of the ZNS device under write workloads with both schedulers. ZNS devices enforce write constraints, therefore only *mq-deadline* can utilize I/O queue depths \(> 1\). We first benchmark the performance of both schedulers by issuing a single 4KiB sequential write I/O in a zone with an increasing number of concurrent threads split across available zones on the device. The number of concurrent threads ranges from 1-14, equivalent to the maximum number of active zones on the device. Figure [\[fig:ZNS_concur_write_seq\]](#fig:ZNS_concur_write_seq){reference-type="ref" reference="fig:ZNS_concur_write_seq"} shows that the resulting performance of both schedulers is nearly identical over all numbers of outstanding I/O requests. Next, we measure the performance of *mq-deadline* issuing I/Os in a single zone and increasing the I/O queue depth, rather than splitting individual I/Os concurrently across zones. The benchmark with the *none* scheduler is the same as for the prior write benchmark, namely issuing 4KiB I/Os concurrently with an increasing number of threads split across the active zones. This allows both benchmarks to have a specific number of outstanding I/O requests (x-axis). Figure [\[fig:ZNS_concur_write_seq_iodepth\]](#fig:ZNS_concur_write_seq_iodepth){reference-type="ref" reference="fig:ZNS_concur_write_seq_iodepth"} shows that the scheduler set to *mq-deadline* achieves a 72.74% lower median and 83.27% lower tail latency. The lower latency for *mq-deadline* as the number of outstanding I/Os increases is due to the merging of I/O requests into larger I/Os, as all the I/Os are at consecutive LBAs. This allows it to issue overall less I/Os than with *none* scheduling in this configuration. **Recommendations:** With this evaluation we can identify that **(i)** random read heavy workloads should avoid the Linux I/O scheduler by setting it to *none*, providing up to 9.95% lower median latency, and **(ii)** multiple outstanding write I/Os should utilize the *mq-deadline* scheduler to merge I/Os in a single zone, rather than splitting I/Os concurrently over multiple zones.
# Experimental Setup {#sec:exp_setup}
For the experiments we utilize a ZNS device, whose details and properties are depicted in Table [\[tab:SSD_architecture\]](#tab:SSD_architecture){reference-type="ref" reference="tab:SSD_architecture"}. The initial goal of this evaluation was to evaluate all possible integrations of ZNS devices. For this we establish the following configuration: **f2fs.** The f2fs (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(c)) parameters are mostly kept at its default options, with the only change being the enforcing of 10% overprovisioning. Since f2fs requires a randomly writable device for its metadata, we expose 4 of the zones on the ZNS device as randomly writable space, corresponding to the maximum amount that the ZNS device can expose as randomly writable space. This space is then used by f2fs for metadata and write caching. However, this requires the zoned space of f2fs (where the actual file system data will be) to align with the size of the randomly writable space, i.e., the randomly writable space has to be large enough to fit all the metadata for the file system. Therefore, the resulting largest possible size that successfully formats the f2fs file system on the zoned space is 100GiB. As a result, we create a 100GiB namespace from the available zoned capacity for the f2fs file system, which we utilize for all experiments. We do not use an additional larger randomly writable device, as we aim to avoid performance implications of multiple devices, which would make it difficult to differentiate between performance effects from the ZNS device and the additional block device. **ZenFS.** ZenFS (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(d)) requires an auxiliary path for its metadata to store LOG and LOCK files for RocksDB at runtime. With this it additionally allows to backup and recover a file system through its command line options, however we do not utilize this option in our evaluation. For the auxiliary path we use the 4GiB randomly writable space exposed by the ZNS device and create a f2fs file system on it which is mounted for solely the ZenFS auxiliary path to be placed on. As the largest zoned space that successfully formats f2fs is 100GiB, we utilize a 100GiB namespace for all experiments, alongside the 4GiB randomly writable namespace, and all remaining capacity is left in an unused namespace. Throughout the evaluation we refer to the 4GiB randomly writable space exposed by the ZNS device as the *conventional namespace*, and the namespace containing the zoned storage on the ZNS device is referred to as the *zoned namespace*. Lastly, since the ZNS specification was integrated in Linux Kernel 5.9.0+, we use a later Kernel version 5.12.0.
# Future Work {#sec:future_work}
With ZNS devices having just been introduced, they leave a plethora of avenues to explore. In particular, as we showcased in Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}, there are numerous possibilities of integrating ZNS devices into the host software stack. In this evaluation we focus mainly on the block-level ZNS performance and provide several initial guidelines for developers. Expansion of these is left as future work by exploring the additional levels of integration. Specifically this includes evaluating the following aspects of ZNS devices:
- What is the performance of the varying levels of integration for ZNS devices in terms of achievable IOPs, latency, transactions/sec, and additional instructions required? Under sequential and random, read and write workloads, are there performance implications of a specific integration, and what rules can be established when building applications for a particular integration?
- How is garbage collection influenced by the different levels of integration? Does building application specific garbage collection policies provide superior performance over other levels of integration? In particular how does an almost fully utilized device (e.g., 95% utilization) affect garbage collection performance at the different levels?
- Do multiple applications running on the same ZNS device interfere with each other? This includes a shared device with multiple namespaces, shared block-level interface with for example multiple concurrently running file systems, and lastly a shared file system on ZNS with multiple concurrent applications. Does in each case an application's garbage collection impact the performance of concurrently running applications?
As we evaluated some of these aspects and failed to produce insightful results, we propose to evaluate them by taking into account our shortcomings and considering the assumptions we provide in Section [28.2](#sec:pitfalls_to_avoid){reference-type="ref" reference="sec:pitfalls_to_avoid"}. An additional exploration that became apparent during this study was that there is currently no enforcing on the number of active zones at a time. The device only allows a maximum number of zones to be active at any point in time, however there is no managing of how active zones are split across namespaces on the device, or across applications running on the same namespace. Especially, as the device supports a larger number of namespaces to be created than active zones that can be open, a zone manager is required to assign zones across namespaces and applications, and provide fair resource sharing across namespaces and applications, enforcing that the maximum number is not exceeded, even if there are more concurrent applications running. Lastly, we suggest the exploration of file system improvements for f2fs, and other ZNS specific file systems, with grouping of files by creation time, death time, or owner as was suggested in and similarly evaluated for conventional SSDs in . This would provide the possibility for optimizing garbage collection and improving the device performance at the file system integration level.
# Introduction
The introduction of flash storage provided significant changes in the storage hierarchy. Achieving as low as single digit \(\mu\)-second latency, several GB/s of bandwidth, and millions of I/O operations per second (IOPS) , they offer significant performance gains over prior storage technologies, such as Hard Disk Drives (HDDs). Flash storage is organized in pages (typically 16KiB in size) , representing the unit of read and write accesses, of which multiple pages are combined into a block (typically multiple MiB in size). Blocks are further packed in planes and dies to manage data and control connectivity to the host. Flash pages do not support in-place updates. As a result, pages have to be erased prior to being written again. However, erase operations require substantially more time than read and write operations . Therefore, erase operations are done at block granularity to amortize the erase overhead. Additionally, flash storage requires pages within a block to be written sequentially. Flash storage therefore includes complex firmware, called the Flash Translation Layer (FTL), to provide the seemingly in-place updates of data and hide the sequential write constraints of devices by exposing a sector/page-addressable SSD . Furthermore, a crucial task of the FTL is to run *garbage collection* (GC), in order to erase blocks with invalid pages and free up space. For this, the FTL reads out valid pages of data from a block and relocates the data to a free block, followed by erasing of the original block. Garbage collection is triggered periodically, or when the device is running low on free space to write data. The resulting interface exposed by conventional flash storage allows it to mimic the behavior of HDDs, thus requiring no changes in the host storage software to access the underlying flash storage. However, recent research results made evident that hiding the flash management complexities from the host leads to suboptimal data placement, unpredictable performance overheads, and shortens the lifetime of flash devices . Therefore, researchers have proposed to open the flash storage interface and expose device internals to the host. This allows for the host to optimize storage management with workload-specific decisions . Zoned Namespace (ZNS) SSDs are the latest addition in these efforts, which are now standardized in the NVMe 2.0 specification and are commercially available . In order to better match the underlying properties of flash chips, ZNS exposes the address space with numerous *zones*, where each zone requires append-only sequential writes. Zones are aligned to the erase unit of a block. With this new interface, the host storage software is now responsible for resetting a zone (i.e., trigger garbage collection) after which a zone becomes writable again from the starting address. Apart from the write restrictions, there are operational parameters such as the zone capacity, the maximum number of active zones limits, and the append limits, which the host software must be aware of. The new interface of these devices necessitates changes to the host storage stack. However, there is more than one way these devices can be integrated within systems. In a classical setup as shown in Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(a), at the bottom is a block device with a file system on top of it, and an application running on top of the file system. Integrating ZNS devices into this storage stack can be done in three different configurations. Firstly, integration at the block-device level (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(b)), while keeping the rest of the stack above the same. An example of this setup is the *dm-zap* zoned device mapper project within the Linux Kernel . This way of integration is the least intrusive one and requires the minimum amount of changes to anything running on top of the block device. Secondly, integration at the file system level (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(c)). For this integration a file system is made aware of the zoned device characteristics such as zone capacity, number of active zones limits, and append limits. Example of such a project is the added ZNS support in f2fs . With such integration the knowledge and required changes for ZNS devices are pushed higher in the stack, from block-device level to the file system level. Lastly, ZNS-aware application integration (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(d)). In this case, there are ZNS-specific application-level changes at the very top of the storage stack. By pushing the customization higher up in the stack, the expectation is to deliver better performance together with the best case application-specific customization and integration with ZNS devices. An example of such an integration is the ZenFS file system module for RocksDB and MySQL . With such configuration possibilities, it is not immediately clear which integration one should choose for their workload. In this research work, we aim to systematically understand the impact of the ZNS integration in the host storage system. This work is largely inspired by related work that has provided unwritten contracts of storage devices for Optane based SSDs and flash based SSDs with guidelines for developers to optimize storage performance. Before we can synthesize actionable design guidelines for storage stack developers, in this work, we first start with systematic benchmarking in the presence of the OS I/O scheduler. In particular, we make the following contributions:
- We provide information on the newly standardized NVMe ZNS devices, how these devices work, how they are integrated into the host storage stack, and how existing applications are modified to support ZNS devices.
- We measure the block-level ZNS device performance, comparing it to conventional block devices in terms of achievable IOPs and bandwidth, and benchmark the possible scheduler configurations for ZNS devices, depicting their implications and limitations.
- We present pitfalls and failed experiments during this evaluation in an effort for others to learn from and to avoid the obstacles we encountered.
- We provide a set of initial guidelines for optimizing ZNS integration into systems, and propose several future work ideas to further explore ZNS integration implications and expand our initial set of guidelines.
- All collected datasets and benchmarking scripts are made publicly available at <https://github.com/nicktehrany/ZNS-Study>. The appendix provides more detailed setup and benchmarking information.
The remainder of the paper is organized as follows. Section [20](#sec:background){reference-type="ref" reference="sec:background"} provides background information on ZNS devices, and their integration into systems. Next, Sections [23](#sec:exp_setup){reference-type="ref" reference="sec:exp_setup"} explains the experimental setup, followed by the first round of unsuccessful experiments in Section [28](#sec:unsuccessful_exps){reference-type="ref" reference="sec:unsuccessful_exps"}. Based on this we provide an adapted experimental setup in Section [7](#sec:adapted_setup){reference-type="ref" reference="sec:adapted_setup"}, and Section [22](#sec:evaluation){reference-type="ref" reference="sec:evaluation"} presents the various benchmarks. Lastly, Section [27](#sec:related){reference-type="ref" reference="sec:related"} provides the related work, followed by future work ideas in Section [24](#sec:future_work){reference-type="ref" reference="sec:future_work"}, and Section [21](#sec:conclusion){reference-type="ref" reference="sec:conclusion"} concludes the paper.
# Footnotes, Verbatim, and Citations
# Related Work {#sec:related}
While ZNS has just recently been standardized , there have been several initial evaluations and discussions of ZNS devices. Bjorling et al. present modifications made to RocksDB and f2fs to support ZNS devices, and showcase the performance gains for these devices. Stavrinos et al. provide an initial discussion on the benefits of ZNS devices and possible improvements for applications on them. Shin et al. show a performance study of ZNS devices, depicting the need for large request sizes for achieving increased on-device parallelism. Similar to the results in our study, where the conventional namespace required larger block sizes (\(\geq\) 16KiB) to reach peak device bandwidth. Han et al. provide a discussion on the ZNS interface, and propose an improved interface particularly optimized for log-structured file systems with segment compaction. However, there currently is no work that systematically studies the possible ways to integrate ZNS devices into the host software stack. We are the first to present a start at such a study. As the ZBD model was originally introduced for SMR devices, there have been several performance evaluations of SMR devices and their integration into systems. Wu et al. showcase an extensive evaluation of host-aware SMR drives, which similarly to ZNS device expose device characteristics to the host. In particular, the authors focus on evaluating performance characteristics of the zoned interface. Additionally, Wu et al. provide a performance study on implications of different properties of host-aware SMR devices, including performance implications under the number of open zones. While not focusing on ZNS devices, past work presented similar evaluations for conventional SSDs, where He et al. provide an unwritten contract for flash-based SSDs, depicting numerous guidelines on performance improvements for such devices. Similarly, Wu et al. present such an unwritten contract for Optane SSDs. Both of these unwritten contracts were inspiration for us to provide a set of developer guidelines for the new ZNS devices. Yang et al. characterize performance implications of building log-structured applications for flash-based SSDs, showcasing the negated benefits of optimizing application data structures for flash, caused by the device characteristics. Such an evaluation showcases application-level integration, which presents insightful results that should be reproduced on ZNS devices to further expand the developer guidelines we provide.
# Unsuccessful Experiments {#sec:unsuccessful_exps}
Designing of experiments to evaluate the performance of the varying levels of integration proved challenging, as initial experiments did not provide insightful results. We provide the iterations of experimental design and why experiments failed in an effort for others evaluating ZNS performance to avoid these pitfalls. Section [28.1](#sec:failed_setup){reference-type="ref" reference="sec:failed_setup"} provides the configurations of the various benchmark, followed by Section [28.2](#sec:pitfalls_to_avoid){reference-type="ref" reference="sec:pitfalls_to_avoid"} describing the failures of these benchmarks and pointing out possible causes for this. Lastly, we give additional lessons learned during this evaluation in Section [28.3](#sec:lessons_learned){reference-type="ref" reference="sec:lessons_learned"}.
## Benchmark Setup {#sec:failed_setup}
We run several benchmarks to evaluate different performance aspects of ZNS devices. While each of the evaluations relies on db_bench and RocksDB, they have slightly different benchmarking configurations. Below we describe the workload parameters for the different benchmarks.
### Integration Level
We first benchmark the different possible integration levels, as shown in Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}. Benchmarks initially fill the database in random and sequential key order, followed by reading from the database in random and sequential key order. We additionally run an overwrite benchmark, which overwrites existing keys in random order, and an updaterandom workload that modifies values of random keys. The overwrite benchmark and the updaterandom differ in the way values are accessed and modified. Overwrite issues asynchronous writes to the key, whereas updaterandom uses a read-modify-write approach. Keys are 16B each, values are configured to be 100B, and we disable any data compression. As mentioned, ZNS devices require direct I/O, we therefore set appropriate db_bench flags to issue direct reads and use direct I/O for flushing and compaction.
### ZenFS Benchmark {#sec:zenfs_exp}
Next, to verify correctness of our setup we attempt to repeat an experiment depicted in , comparing db_bench performance on a configuration with f2fs to a configuration with ZenFS. The focus of this benchmark is write intensive workloads, as this is meant to trigger increased garbage collection and showcase gains of managing ZNS at the application-level with ZenFS. The benchmark is configured to run fillrandom and overwrite with a key size of 100B and value size of 800B. We additionally use data compression to compress data down to 400B. The original paper uses the entire device for its benchmark, however we scale it to the maximum possible that successfully fits into the namespace, which is equivalent to 50 million keys. Lastly, we set the target file size of SST files to be equal to the zone capacity, and again utilize appropriate flags for direct I/O.
### Multi-Tenancy
Lastly, we run an experiment to evaluate how multiple concurrently running namespaces affect the performance of one another. For this we mount f2fs on the 100GiB zoned namespace and create an additional 100GiB namespace on which ZenFS with a db_bench benchmark is running. We compare the performance of running the ZenFS namespace alone, without any interference from another namespace, to running the f2fs namespace concurrently. The ZenFS namespace runs the same workload as describe in the previous benchmark (Section [28.1.2](#sec:zenfs_exp){reference-type="ref" reference="sec:zenfs_exp"}). The goal being that the f2fs namespace creates substantial device traffic through write I/Os, especially during garbage collection from the overwrite benchmark, and thus show the performance impact on the ZenFS namespace.
## Pitfalls to Avoid {#sec:pitfalls_to_avoid}
Results of all experiments showed little to no performance difference in their benchmarks. Especially the ZenFS benchmark (Section [28.1.2](#sec:zenfs_exp){reference-type="ref" reference="sec:zenfs_exp"}), where the original paper showed substantial performance gains with ZenFS, in particular on overwrite benchmarks where GC is being triggered heavily. We failed to reproduce these exact results and only had minor performance gains from ZenFS. The multi-tenant evaluation showed very similar results, which appear contrary to prior expectations. That is if one namespace fully utilizes the device, then another namespace attempting to use the same device will have some performance implications, as they would now be sharing the device resources. As the prior experiments proved ineffective in their evaluation, we propose the assumption that for performance differences to appear on the ZNS device, it has to be largely utilized, such that LBA mappings are fully setup and more garbage collection is triggered. If the device is not largely utilized, LBA mappings are not fully setup and the device is able to provide peak performance without showing effects of garbage collection, as there is a large amount of free space it can utilize. Additionally, the device bandwidth has to be utilized to the extent that it competes with the garbage collection happening in the background. We therefore believe that with our setup we failed to produce enough device utilization in bandwidth and space, and thus evaluations showed that there are no performance differences. While we did not evaluate all configurations under increased garbage collection and device utilization, we believe it to have an effect on performance and thus suggest its evaluation as future work (discussed in detail in Section [24](#sec:future_work){reference-type="ref" reference="sec:future_work"}). Per suggestions of Western Digital, for significant performance advantages to appear for ZNS devices, utilization of available storage capacity should be \(>50\%\) (preferably \(80\%+\)) and bandwidth utilization of \(>30\%\) of the device bandwidth, such that the write workload competes with ongoing garbage collection. However, we have not evaluated these specific configurations and thus leave it as future work.
## Lessons Learned {#sec:lessons_learned}
In addition to unsuccessful experiments we encountered several obstacles that were not immediately obvious to debug. Again, we provide our experiences in order for others to avoid these pitfalls. Firstly, by default the ZNS devices do not set a scheduler, neither do the applications such as f2fs or ZenFS, nor is there an error on an invalid scheduler being set. Thus setting up of the applications and formatting the ZNS device completes successfully, while as soon as writes are issued to the device I/O errors appear. Therefore, it is important to ensure the correct scheduler is always set, that is the *mq-deadline* scheduler, and every namespace requires it to be set individually after creation. The majority of applications utilize multiple outstanding I/Os per zone, thus the *mq-deadline* scheduler is required, as it will hold back I/Os such that only one outstanding I/O is submitted to the device at any point in time. If the application issues a single I/O synchronously, the scheduler could be left at the default configuration set to *none* (default setting in Linux 5.12.0), however here again the application must enforce the sequential write constraint. Secondly, device mapper support is not there yet. A device mapper implements a host-side FTL that makes the sequential write required zones randomly writable by exposing the zoned device as a conventional block device. It controls data placement, maintains data mappings, and runs garbage collection, just as the FTL on traditional flash storage. An existing implementation, such as *dm-zoned* is ZBC and ZAC compliant but not compliant to the new concepts of zone capacity from the ZNS specification. The *dm-zap* device mapper aims to be ZNS compliant, however it is still a prototype and not fully functional with ZNS devices yet. As a result, we are currently not able to evaluate performance of one level of integration (Figure [\[fig:ZNS_configurations\]](#fig:ZNS_configurations){reference-type="ref" reference="fig:ZNS_configurations"}(b)), requiring a future evaluation when the device mapper support is functional.
[^1]: Available at <https://github.com/nicktehrany/ZNS-Study>
[^2]: Available at <https://github.com/linux-nvme/libnvme>
[^3]: Available at <https://github.com/linux-nvme/nvme-cli>
[^4]: Available at <https://github.com/util-linux/util-linux>
[^5]: Available at <https://github.com/westerndigitalcorporation/libzbd>
[^6]: Available at <https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs-tools.git/about/>
[^7]: Available at <https://github.com/westerndigitalcorporation/zenfs>
[^8]: Available at <https://github.com/axboe/fio>
[^9]: Available at <https://github.com/nicktehrany/ZNS-Study>
[^10]: Available at <https://github.com/linux-nvme/libnvme>
[^11]: Available at <https://github.com/linux-nvme/nvme-cli>
[^12]: Available at <https://github.com/util-linux/util-linux>
[^13]: Available at <https://github.com/westerndigitalcorporation/libzbd>
[^14]: Available at <https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs-tools.git/about/>
[^15]: Available at <https://github.com/westerndigitalcorporation/zenfs>
[^16]: Available at <https://github.com/axboe/fio>
[^17]: f2fs-tools provides the mkfs.f2fs functionality to format the storage device in order to mount the f2fs file system. ZNS support was added in version 1.14.0. Available at <https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs-tools.git/about>
[^18]: f2fs-tools provides the mkfs.f2fs functionality to format the storage device in order to mount the f2fs file system. ZNS support was added in version 1.14.0. Available at <https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs-tools.git/about> | {'timestamp': '2022-06-06T02:13:28', 'yymm': '2206', 'arxiv_id': '2206.01547', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01547'} |
# Introduction {#sec:intro}
The standard model of particle physics predicts that the electromagnetic (EM) interaction is invariant under the parity operator. Nonetheless, models of parity-violating electromagnetism have been proposed, reflecting the phenomenology of weak interactions . A popular model extends Maxwell's Lagrangian with a parity-violating Chern-Simons term : \[\mathcal{L} = \mathcal{L}_{EM} + \mathcal{L}_{CS} =-\dfrac{1}{4} F_{\mu\nu}F^{\mu\nu}-\dfrac{1}{2} p_\mu A_\nu \widetilde{F}^{\mu\nu},\] where \(F_{\mu\nu}\) is the EM field tensor, \(p_\mu\) is a constant kinematic four-vector with dimensions of mass, \(A_\nu\) is the four-potential and \(\widetilde{F}^{\mu\nu} = \frac{1}{2}\epsilon^{\mu\nu\alpha\beta}F_{\alpha\beta}\), \(\epsilon^{\mu\nu\alpha\beta}\) being the four-dimensional Levi-Civita symbol. The additional parity-violating term \(\mathcal{L}_{CS}\) is responsible for the Cosmic Birefringence (CB) effect, the in-vacuo rotation of the linear polarization plane of photons during propagation.
Hence, to observe such an effect one needs a source of linear polarised light. The Cosmic Microwave Background (CMB) is linearly polarised at the level of 1% to 10% due to Thomson scattering off free electrons at the surface of last scattering and it is therefore a good observable to constrain this phenomenon. Moreover, the CMB is the farthest EM radiation available in nature and this increases the chances of detecting the CB effect, as the polarization plane had more time to rotate, since the effect accumulates during propagation.
The CB effect induces correlations between the temperature field T and the B-modes of the polarization field or between the E-and B-modes of the polarization field, that are null with standard EM as the parity symmetry prevents these correlations to be active. Thus, we can use TB and/or EB correlations to detect the CB effect. However, the signal-to-noise ratio due to TB correlation is about a factor of two smaller than EB . Hence, in this paper we focus only on the EB correlation.
If the parity-violating cross-correlations of the CMB fields do not depend on the direction of observation , the CB effect is isotropic and it can be modelled with a single angle \(\alpha\). Otherwise, the CB effect is anisotropic and it has to be modelled with a set of angles \(\alpha(\hat{n})\). Recent constraints on the CB effect can be found in for the isotropic case[^1] and in for the anisotropic case.
Parity violating electromagnetism Lagrangian also induce a non-null cross-correlation between the CB field and the CMB temperature and polarization maps, see e.g. and references therein for a detailed treatment of cross-spectra and bi-spectra. Furthermore, in recent years, early dark energy models inspired by ultra-light axion fields have been proposed to alleviate the \(H_0\) tension . These models are known to produce CB signal and a non-zero cross-correlation between CB and the CMB fields, see e.g..
In this paper we analyse CB extending what was done in Ref. . In particular, using the Public Data Release 3 (PR3) products, we provide for the first time constraints on the cross-correlation spectra \(C_L^{\alpha E}\) and \(C_L^{\alpha B}\) between the CB angle maps and the CMB polarization maps at large angular scales. We also provide the CB auto-correlation spectra \(C_L^{\alpha\alpha}\) and its cross-correlation spectra \(C_L^{\alpha T}\) with the CMB temperature map up to \(L=24\), exploiting the Public Data Release 4 (NPIPE) data in addition to the PR3 products.
This paper is organised as follows: in Section [2](#sec:Dataset and Simulations){reference-type="ref" reference="sec:Dataset and Simulations"} we review the features of the datasets and simulations employed for this study; in Section [3](#sec:Analysis pipeline){reference-type="ref" reference="sec:Analysis pipeline"} we describe the analysis pipeline; in Section [4](#sec:Results){reference-type="ref" reference="sec:Results"} we present the results of this work; in Section [5](#sec:Conclusions){reference-type="ref" reference="sec:Conclusions"} we draw the conclusions; in Appendix [\[app:effective beam calculation\]](#app:effective beam calculation){reference-type="ref" reference="app:effective beam calculation"} we explain the effective beam calculation used for the CMB spectra; in Appendix [\[app:results on simulations\]](#app:results on simulations){reference-type="ref" reference="app:results on simulations"} we report the results of the analysis on simulations; in Appendix [\[app:Analysis robustness tests\]](#app:Analysis robustness tests){reference-type="ref" reference="app:Analysis robustness tests"} we perform robustness tests for the analysis pipeline; in Appendix [\[app:2D plots\]](#app:2D plots){reference-type="ref" reference="app:2D plots"} we set 2-dimensional (2D) joint constraints on \(A^{\alpha \alpha}\) and \(A^{\alpha T}\).
# Datasets and Simulations {#sec:Dataset and Simulations}
In this work we employ CMB maps provided by the collaboration in the PR3 analysis and in the subsequent NPIPE re-analysis. The PR3 maps have been cleaned from foregrounds by the four official *Planck* component separation methods `Commander`, , and `SMICA` while NPIPE maps have been processed only by `Commander` and Ṫhe corresponding simulations, provided by the *Planck* collaboration, have been also employed. The resolution of the maps, following the [^2] pixelization scheme, is \(N_{side} = 2048\). The NPIPE products have never been used to estimate the CB auto-correlation spectra \(C_L^{\alpha\alpha}\) before this work.
The PR3 simulation set is composed of 1000 CMB Monte Carlo simulations generated using the \(\Lambda\)CDM best-fit model, 300 noise simulations for the full-mission (FM) and 300 noise simulations for each of the two half-mission (HM) data splits, and for each of the component separation methods. We sum the first 300 CMB realizations with each set of 300 HM noise simulations that realistically describe the 2018 data. We employ the two data splits in the construction of the \(\alpha\) map. We also sum the first 300 CMB realizations with the 300 FM noise simulations in order to obtain 300 FM simulations. We use these maps for the cross-correlation spectra \(C_L^{\alpha E}\) and \(C_L^{\alpha B}\) estimation. The PR3 simulations contain beam leakage effects , ADC non linearities, thermal fluctuations and other systematic effects .
The NPIPE data and simulation sets are an evolution over PR3 . One of the effects of these improvements is lower levels of noise and systematics in the CMB maps at essentially all angular scales . We employ the 400 simulations available for `Commander`[^3] and the 600 simulations available for . The NPIPE `Commander` simulations contain the PR3 CMB signal and the instrumental noise. To extract the noise-only maps, we subtract the signal-only PR3 simulations to the complete NPIPE simulations. The NPIPE CMB simulations are divided in noise-only and signal-only as the PR3 ones, which we sum for a set of noisy CMB simulations. The NPIPE data splits, where the systematics between the splits are expected to be uncorrelated, are called A and B (see again for a full description). We calculate the NPIPE CMB Angular Power Spectra (APS) by cross-correlating those A and B maps. The NPIPE simulations employed in this work are the "noise aligned" ones, see again for details.
# Analysis pipeline {#sec:Analysis pipeline}
As first step of the analysis we need to define the processing masks. We divide the sky at \(N_{side} = 2048\) in small regions following the pixelization scheme at \(N_{side} = 8\). In the rest of the paper we refer to these regions as "patches". The sky fraction covered by one single patch is \(f_{sky,patch} \simeq 0.13\%\) and the total number of available patches is 768. Working at resolution \(N_{side}=8\) is an improvement with respect to , where they used \(N_{side}=4\). We tried to increase the resolution to \(N_{side}=16\), but the patch sky fraction was too low to get reliable CMB spectra.
We then combine the masks of these patches with a polarization mask that excludes the regions unobserved or heavily contaminated by foregrounds. For PR3 we combine the foreground mask (Galactic and point sources) with the HM missing pixels mask. For NPIPE only the foreground mask is employed. The polarization masks used are shown in the top panels of Fig. [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"} (left for PR3, right for NPIPE).
Some of the aforementioned patches contain masked pixels. A large fraction of masked pixels in a patch prevents the extraction of reliable CMB spectra. For this reason, we select and analyse only the patches that have at least 50% of active (i.e. non masked) pixels. This corresponds to \(f_{sky,patch} \gtrsim 0.065\%\). The maps showing the selected patches and the fraction of active pixels in each of them are shown in the bottom panels of Fig. [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"} (left for PR3, right for NPIPE), where the grey patches are the excluded ones. As expected, among the selected patches, the ones close to the Galactic mask borders or the ones containing a large number of point sources have a lower percentage of active pixels. The number of selected patches is 571 for PR3 and 592 for NPIPE. The final masks are not apodized because the apodization process reduces the effective sky fraction covered by each patch and, for the sky fraction considered, it does not provide any benefit in the CMB computed APS, see also the discussion about polarization purification in the next paragraph.
For each selected patch, the CMB APS is calculated with the Python package Pymaster[^4] by cross-correlating the two splits both for data and simulations. We recall that the data splits are HM 1 and 2 for PR3, A and B for NPIPE, as explained in Sec. [2](#sec:Dataset and Simulations){reference-type="ref" reference="sec:Dataset and Simulations"}. We compute the APS in cross-mode to reduce systematic effects and noise mismatches . When analysing a masked CMB map to produce the APS, E to B leakage can be generated . We have the possibility to purify the E and/or B modes in order to remove the misinterpreted modes, at a cost of information loss. We choose not to purify E and B modes: for the noise level of *Planck*, the size of the patches considered and for the multipole range of interest, the purification does not provide any substantial advantage, and does increase the error bars. We calculate the APS for multipoles from 2 to 2047. We deconvolve a non-Gaussian effective beam obtained from simulations, as explained in App. [\[app:effective beam calculation\]](#app:effective beam calculation){reference-type="ref" reference="app:effective beam calculation"}. The APS is binned with \(\Delta l = 60\) multipoles in order to reduce the errors and the correlations induced by the cut-sky .
We show in the top panels of Fig. [\[fig:CMB_spectra_validation\]](#fig:CMB_spectra_validation){reference-type="ref" reference="fig:CMB_spectra_validation"} the PR3 `Commander` validation plots for the EE, BB and EB power spectra for three patches, in terms of the bandpowers \(D_\ell = \ell(\ell+1)C_\ell/2\pi\). The three patches, following the RING ordering scheme, have indexes 1 (blue), 207 (orange) and 542 (green) and fractions of patch active pixels 100%, 74.5% and 51.1%, respectively. The dots represent the mean of the simulations and the error bars show the uncertainty on the mean. The fiducial power spectra are represented by the black lines in the plots. These plots show that the simulations CMB spectra are validated. The scatter around the fiducial at the highest multipoles is large and, for this reason, we do consider the APS with \(\ell \lesssim 1500\). We show in the central panels of Fig. [\[fig:CMB_spectra_validation\]](#fig:CMB_spectra_validation){reference-type="ref" reference="fig:CMB_spectra_validation"} the residuals plots, i.e. the difference between the simulation average and the fiducial spectra, divided by the error of the mean. For all the cases considered the simulation average is less than \(3.5\sigma_{\rm mean}\) away from the fiducial. The residuals do not show any systematic behaviour. We report in the bottom panels of Fig. [\[fig:CMB_spectra_validation\]](#fig:CMB_spectra_validation){reference-type="ref" reference="fig:CMB_spectra_validation"} the error of the mean. As expected, the error increases for higher multipoles and for lower numbers of active pixels in the patch. We verified that the behaviour of other patches and other component separation methods is similar to the one shown in Fig. [\[fig:CMB_spectra_validation\]](#fig:CMB_spectra_validation){reference-type="ref" reference="fig:CMB_spectra_validation"}. For the NPIPE case the simulations are validated with a similar procedure, the only difference is that the errors are slightly smaller. This is due to the lower level of noise with respect to the PR3 simulation set .
Given the low number of simulations, the CMB covariance matrices for each patch and for each component separation method are obtained via the analytical implementation of Pymaster . The module that produces the matrices needs the CMB signal APS, the noise APS and the smoothing function. The CMB signal APS used for the covariance matrix calculation is the FFP10 fiducial . The noise APS for each of the two data splits (HM1 and HM2 for PR3, A and B for NPIPE) is obtained by averaging the auto-correlation of the corresponding noise-only simulations. The smoothing function considered is obtained as discussed in App. [\[app:effective beam calculation\]](#app:effective beam calculation){reference-type="ref" reference="app:effective beam calculation"}. We verified that these analytical covariance matrices are a good approximation of the ones obtained using the simulations. This is not valid for the first multipole bin, so we exclude the latter from the analysis and use the CMB spectra starting from \(\ell_{min}=62\).
After obtaining the CMB APS and the corresponding covariance matrices, we estimate the CB angle in each patch assuming that the CB effect is isotropic inside each patch. The CMB APS is rotated by the isotropic CB effect in the following way : \[\begin{aligned}
C_\ell^{TT,obs} &= C_\ell^{TT},\\ C_\ell^{TE,obs} &= C_\ell^{TE}\cos(2\alpha),\\ C_\ell^{TB,obs} &= C_\ell^{TE}\sin(2\alpha),\\ C_\ell^{EE,obs} &= C_\ell^{EE}\cos^2(2\alpha) + C_\ell^{BB}\sin^2(2\alpha),\\ C_\ell^{BB,obs} &= C_\ell^{BB}\cos^2(2\alpha) + C_\ell^{EE}\sin^2(2\alpha),\\ C_\ell^{EB,obs} &= \frac{1}{2} \left(C_\ell^{EE}-C_\ell^{BB}\right)\sin(4\alpha), \end{aligned}\] where \(\alpha\) is the CB angle, the APS with the \(obs\) tag is the rotated (observed) one and the APS without the \(obs\) tag is the unrotated one. Without the CB effect, i.e. \(\alpha = 0\), the TB and EB spectra are null, as expected.
Harmonic based estimators for \(\alpha\) can be built exploiting the observed EB and/or TB spectra, see e.g. . Since the signal-to-noise ratio is smaller for the TB-based estimator as compared to EB one, we use the EB-only D-estimator defined in : \[D_\ell^{EB,obs} = C_\ell^{EB,obs} \cos(4\alpha)-\frac{1}{2} \left(C_\ell^{EE,obs}-C_\ell^{BB,obs}\right)\sin(4\alpha) \,. \label{eq:Dest}\]
We obtain an estimate of the isotropic CB angle \(\alpha\) in each patch by minimizing the following \(\chi^2_{EB}\): \[\chi^2_{EB} = \sum_{\ell,\ell'}D_\ell^{EB,obs} M_{\ell\ell'}^{EB} D_{\ell'}^{EB,obs},\] where \(D_\ell^{EB,obs}\) are the D-estimators defined in Eq. ([\[eq:Dest\]](#eq:Dest){reference-type="ref" reference="eq:Dest"}) and \(M_{\ell\ell'}^{EB} = \langle D_{\ell}^{EB,obs}D_{\ell'}^{EB,obs}\rangle^{-1}\) is their covariance matrix. The \(M_{\ell\ell'}^{EB}\) matrix reduces to the EB covariance matrix, i.e. without the CB effect , when the noise on the E and B CMB fields is equivalent. Since the latter is valid for the Planck maps we are employing , we use the EB covariance matrix in the \(\chi^2_{EB}\) minimization. Only the CMB spectra and covariance matrix related to a specific patch are used in the \(\chi^2_{EB}\) to estimate the CB angle in that patch.
For each CMB map we minimize the \(\chi^2_{EB}\) in each patch obtaining a map of CB angles. From these maps we firstly estimate the CB monopole averaging the angle values in the pixels outside the masks shown in the bottom panels of Fig. [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"}. Then, we calculate their auto-correlation spectra \(C_L^{\alpha\alpha}\) and their cross-correlation spectra \(C_L^{\alpha T}\), \(C_L^{\alpha E}\) and \(C_L^{\alpha B}\) with the CMB T, Q, U maps employing a Quadratic Maximum Likelihood (QML) estimator . The maps and the associated covariance matrices used in the QML estimator are described in the next paragraphs.
The CB fields for data and simulations are obtained through the \(\chi^2_{EB}\) minimization, described above. Following , the covariance matrices associated to the data are assumed to be diagonal, and they are computed from the variance of the simulations in each pixel. See App. [\[app:results on simulations\]](#app:results on simulations){reference-type="ref" reference="app:results on simulations"} for more details about this choice. In the computation of the angular power spectra, we marginalize over monopole (for all the cases) and dipole (for the \(\alpha T\), \(\alpha E\) and \(\alpha B\) cases) contributions following the procedure described in. Hence, note that we retain \(L=1\) as the lowest multipole for \(\alpha\alpha\) spectra.
For the CMB temperature field we always use the `Commander` map since it is the one used by the collaboration for the large scale temperature cosmological analysis. We produce 300 signal-only CMB temperature simulations at resolution \(N_{side}=8\) smoothed with a Gaussian beam with 880' FWHM and an analytic pixel window function, to which we sum isotropic white noise realizations with standard deviation of 500 nK for regularization. The `Commander` data map is down-sampled and regularized accordingly. The CMB temperature covariance matrix is analytically calculated from the FFP10 fiducial spectrum and contains the marginalization of monopole and dipole and a diagonal term to take into account the 500 nK regularization noise.
For the CMB polarization fields, instead, we use the `SMICA` polarization maps because it is the only component separation method providing harmonic weights[^5] that allow to correctly compute the pixel space noise covariance matrix, from single frequency noise covariances. We down-sample the 300 PR3 polarization simulations and the data maps from \(N_{side}=2048\) to \(N_{side}=8\), applying, in addition to an analytic pixel window function, a cosine window function defined as: \[b_{\ell} = \left\{\begin{matrix} 1 & \hspace{-28pt} \mbox{for \(\ell < N_{side}\)}\\ \frac{1}{2}\left(1+\sin\left(\frac{\pi}{2}\frac{\ell}{N_{side}}\right)\right) & \hspace{20pt} \mbox{for \(N_{side} \leq \ell < 3\ N_{side}\)}\\ 0 & \hspace{-44.5pt} \mbox{otherwise}\\ \end{matrix}\right..\] We then sum a 20 nK regularization noise. The corresponding covariance matrix contains: the correlated noise term (computed using the `SMICA` weights and the FFP8 noise covariance matrices), the signal term based on a fiducial power spectrum, and the regularization noise term. We limit the estimates of the \(\alpha E\) and \(\alpha B\) spectra at \(L = 23\) by definition of the considered smoothing window.
We show in Fig. [\[fig:CB_spectra_masks\]](#fig:CB_spectra_masks){reference-type="ref" reference="fig:CB_spectra_masks"} the masks used for the extraction of the spectra with the QML estimator. The mask employed with the \(\alpha\) and \(T\) maps has \(\approx\) 74% of non-masked pixels, while the fraction of active pixels for the \(Q\) and \(U\) mask is \(\approx\) 47%. The masks are obtained as the product of the PR3 \(\alpha\) mask, that excludes the grey patches shown in the bottom left panel of Fig. [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"}, and the CMB galactic mask for intensity (\(\alpha\) and \(T\) maps) or polarization (\(Q\) and \(U\) maps).
After having obtained the \(C_L^{\alpha X}\) spectra, where X can be \(\alpha\), T, E or B, we set constraints in terms of the scale-invariant amplitude \(A^{\alpha X}\). To do so, we minimize the following \(\chi^2\left(A^{\alpha X}\right)\): \[\label{eq:chi2_A} \chi^2\left(A^{\alpha X}\right) = \sum_{L,L'}\left(\dfrac{L(L+1)}{2\pi}C_L^{\alpha X}-A^{\alpha X}\right) M_{LL'}^{-1} \left(\dfrac{L'(L'+1)}{2\pi}C_{L'}^{\alpha X}-A^{\alpha X}\right),\] where \(C_L^{\alpha X}\) are the data spectra and \(M_{LL'} = \left< \dfrac{L(L+1)}{2\pi} C_L^{\alpha X} \dfrac{L'(L'+1)}{2\pi}C_{L'}^{\alpha X} \right>\) is calculated from simulations.
# Results {#sec:Results}
In this section we show the results of our analysis, adopting the following color code: red for `Commander`, orange for , green for , and blue for `SMICA`.
In Figure [\[fig:beta_maps_PR3\]](#fig:beta_maps_PR3){reference-type="ref" reference="fig:beta_maps_PR3"} we show the CB maps (left panels) extracted from CMB data, and the associated error maps computed from simulations (right panels) for the PR3 dataset. In Figure [\[fig:beta_maps_NPIPE\]](#fig:beta_maps_NPIPE){reference-type="ref" reference="fig:beta_maps_NPIPE"} the same quantities are displayed for NPIPE. By visually comparing PR3 CB maps, we can recognize similar patterns for all the four component separation methods. The same applies to the NPIPE dataset. The errors associated to NPIPE are generally lower with respect to the PR3 ones. This is simply due to the fact that the NPIPE CMB maps are slightly less noisy than the PR3 ones . Furthermore some NPIPE low resolution patches contain a larger number of observed pixels, with respect to the PR3 dataset, whose mask contains also the contribution of half mission missing pixels (see top panels of Fig. [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"}). Finally, the darker regions in error maps, which represent patches where the error on the CB angle is lower, are well correlated with the *Planck* scanning strategy, that observes more deeply the ecliptic poles.
In Table [1](#tab:monopole){reference-type="ref" reference="tab:monopole"} we report the monopoles of the CB maps. They are obtained as a weighted average of the data maps with weights corresponding to the inverse variance of the simulations. All the monopole values are consistent with zero provided the systematic error is taken into account. Moreover they are well compatible, within the statistical uncertainty, across the different component separation methods. The statistical error contribution is computed from the standard deviation of the simulations, while the systematic error is due to the uncertainty in the orientation of 's polarization-sensitive bolometers and is taken from Ref. . Also here we note that, due to the lower level of the instrumental noise and to the larger number of non-masked pixels, the NPIPE data provide smaller statistical errors. The monopole values found in this work agree with the ones reported in literature, see e.g. .
In Figure [\[fig:results_data_spectra\]](#fig:results_data_spectra){reference-type="ref" reference="fig:results_data_spectra"} we show the auto-correlation of the CB maps and their cross-correlation with CMB[^6]. As expected the scatter of the NPIPE \(\alpha\alpha\) spectra around zero and the associated errors are slightly lower with respect to PR3. Across the spectra, we find a general good compatibility with the null effect. This is assessed by comparing the harmonic \(\chi^2\) of data and simulations, reported in Fig. [\[fig:results_histo\]](#fig:results_histo){reference-type="ref" reference="fig:results_histo"}.
The data \(\chi^2\) for all the spectra lies within the histograms, suggesting that the data are compatible with the simulations, hence with the null effect. However, we note that for the \(\alpha E\) and \(\alpha B\) cases the \(\chi^2\) of the data are closer to the right tail of the corresponding histograms, with the following Probability To Exceed (PTE): 0.9--10.6% for \(\alpha E\) and 1.4--2.5% for \(\alpha B\) (see also Tab. [2](#tab:pte){reference-type="ref" reference="tab:pte"} where all the PTEs are reported).
Despite the overall good compatibility with the null effect, some single multipole fluctuations deserve a closer look. Considering the \(\alpha T\) spectra, \(L=5\) shows a PTE between 1% and 2.7% for the four component separation methods, i.e. \(\sim2.5\sigma\) fluctuation. For the \(\alpha E\) spectra, the multipoles \(L=5, 12, 14, 19\) have the largest deviation from the null effect, if compared with the corresponding empirical distributions. These PTEs are below 1% for \(D_{19}^{\alpha E}\) `Commander` and \(D_{12}^{\alpha E}\) , but still within \(3\sigma\) fluctuation. The same happens for \(D_2^{\alpha B}\) , and `SMICA`, all with PTEs lower than 1%, but within \(3\sigma\). The PTE for \(D_{13}^{\alpha B}\) is between 2% and 2.7%, which corresponds to \(\sim 2.3 \sigma\) fluctuation. All the spectra obtained with the `SMICA` component separation method are reported in Tab. [3](#tab:smica_all_spectra){reference-type="ref" reference="tab:smica_all_spectra"} along with the statistical uncertainty at \(1\sigma\) level for each multipole[^7]. The averaged spectra of the simulations are shown in App. [\[app:results on simulations\]](#app:results on simulations){reference-type="ref" reference="app:results on simulations"}. They all show a nice compatibility with the null-hypothesis, meaning that the residual systematic effects included in the simulations do not have a significant impact on the spectra considered.
In Figure [\[fig:results_posteriors\]](#fig:results_posteriors){reference-type="ref" reference="fig:results_posteriors"} we display the posterior distributions for the scale invariant amplitude, \(A^{\alpha X}\), computed using Eq. ([\[eq:chi2_A\]](#eq:chi2_A){reference-type="ref" reference="eq:chi2_A"}). In Table [4](#tab:results_constraints){reference-type="ref" reference="tab:results_constraints"} we report the upper limits at 95% C.L., for \(A^{\alpha\alpha}\), and the constraints at 68% C.L., for \(A^{\alpha T}\), \(A^{\alpha E}\) and \(A^{\alpha B}\). As before, also for these parameters, we find a good compatibility among the different component separation methods and with the null effect. Furthermore, for \(A^{\alpha\alpha}\), there is good compatibility between the PR3 and the NPIPE results. The NPIPE upper limit using the `Commander` component separation method is looser with respect to the PR3 one because the posterior peak is slightly shifted to the right. Even if a small shift is present also for , the lower width of the posterior of NPIPE results in a tighter constraint on \(A^{\alpha\alpha}\), when compared with PR3. The \(\alpha T\) constraints are compatible with, but tighter than , where they used the same approach adopted in this work but at the lower resolution \(N_{side}=4\). Thus, a higher resolution provides more constraining power for \(\alpha T\).
# Conclusions {#sec:Conclusions}
In this work we build CB maps at angular scales larger than \(\sim 7\) deg, exploiting CMB component separated maps from both PR3 and NPIPE releases. From these CB maps we estimate the monopoles (see Tab. [1](#tab:monopole){reference-type="ref" reference="tab:monopole"}), i.e. the isotropic birefringence angle, finding a very good compatibility with previous results . We also compute their spectra and the cross-correlation with the CMB temperature map. Moreover, we provide for the first time the spectra of the cross-correlation of the CB field with the CMB E and B fields. The data CB angle maps and spectra, for both auto-and cross-correlations with CMB, are made publicly available[^8].
We quantify the compatibility of all the aforementioned spectra with null effect through an harmonic \(\chi^2\) analysis, see Fig. [\[fig:results_histo\]](#fig:results_histo){reference-type="ref" reference="fig:results_histo"}, and through a scale invariant amplitude \(A^{\alpha X}\), with \(X=\alpha,T,E,B\), see Fig. [\[fig:results_posteriors\]](#fig:results_posteriors){reference-type="ref" reference="fig:results_posteriors"}. We find no significant evidence of deviation from the null effect. The constraints on \(A^{\alpha X}\) are summarized in Tab. [4](#tab:results_constraints){reference-type="ref" reference="tab:results_constraints"}. The latter are obtained through a \(\chi^2\) minimization assuming null CB, which is supported by data. We also constrain jointly \(A^{\alpha \alpha}\) and \(A^{\alpha T}\) with a pixel-based likelihood which naturally takes into account the effect of CB-induced cosmic variance, see App. [\[app:2D plots\]](#app:2D plots){reference-type="ref" reference="app:2D plots"}.
In this work we extend to \(L=24\) the multipole range of the \(\alpha\alpha\) and \(\alpha T\) spectra covered in , which adopts the same technique for the construction of the CB maps. Note that for the PR3 \(\alpha T\) case we obtain constraints compatible with, but tighter than,. In addition, we provide for the first time estimates of the *Planck* NPIPE \(\alpha\alpha\) spectra. However, the main novelty of this work is represented by the estimates of the *Planck* PR3 \(\alpha E\) and \(\alpha B\) spectra, which are presented here for the first time. The latter might be fruitfully considered to constrain models of anisotropic birefringence that predict correlations with the CMB fields. We acknowledge the financial support from the INFN InDark initiative and from the COSMOS network (www.cosmosnet.it) through the ASI (Italian Space Agency) Grants 2016-24-H.0 and 2016-24-H.1-2018, as well as 2020-9-HH.0 (participation in LiteBIRD phase A). We acknowledge the use of the `healpy` , `NaMaster` , `numpy`, `matplotlib` software packages, and the use of computing facilities at CINECA.
[^1]: For a detailed summary of previous constraints see e.g. references in.
[^2]: <http://healpix.sourceforge.net>
[^3]: The simulations used in the analysis are actually 399 because one turned out corrupted.
[^4]: <https://namaster.readthedocs.io>
[^5]: <https://wiki.cosmos.esa.int/planck-legacy-archive/index.php/SMICA_propagation_code>
[^6]: All the spectra shown in this paper are given in terms of bandpowers, i.e. \(D_L^{\alpha X} = L(L+1)C_L^{\alpha X}/2\pi\), with \(X=\alpha,T,E,B\).
[^7]: The other CB spectra, along with the CB angle maps, are made publicly available at <https://github.com/marcobortolami/AnisotropicBirefringence_patches.git>.
[^8]: <https://github.com/marcobortolami/AnisotropicBirefringence_patches.git> | {'timestamp': '2022-06-06T02:16:25', 'yymm': '2206', 'arxiv_id': '2206.01635', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01635'} |
# Introduction {#sec:intro}
In this numerical simulation study, we consider the task to optimize stimulation currents in the multi-channel version of transcranial electrical stimulation (tES) which is applied non-invasively for stimulating neuronal activity, treating psychiatric disorders and studying neuronal behavior. In tES, a pattern of electric currents are applied through a set of electrodes attached to subject's head. Part of the generated diffusive current field penetrates through the skull into the brain modulating cortical excitability. The procedure for adjusting the electrode montage delivering the stimulus varies from one method to another, considering various properties such as the number of active electrodes, physical description (e.g., positioning, shape, permittivity and impedance values), applied stimulus waveform (e.g., amplitude, pulse shape, pulse width, and polarity), the number of stimulation sessions, and time interval in-between. Transcranial direct current stimulation (tDCS) delivers constant, low-intensity currents injections of 0.5-4.0 mA over a pair of large saline-soaked 20-35 cm\(^2\) electrode patches, with one patch adhered on the scalp, whereas the second patch can be either cephalic or extra-cephalic. The drawback, however, is the limitation of delivering target specific frequencies and the lack of focality. tDCS is a well-known treatment of neuropsychiatric disorders and brain illnesses, for instance, stroke conditions, epilepsy syndromes, Parkinson's disease, major depression disorder, tinnitus, migraine, and alcoholism. Multi-channel tES generally constitutes a task to select a multi-component current pattern to create a sought field in a given location. Selecting such pattern poses an ill-posed inverse problem, i.e., it does not have a unique solution and a slight change in the selected current pattern can significantly change the resulting current density in the brain. The problem can be considered as over-determined, i.e., the three-dimensional current field inside the brain is likely to have more degrees of freedom than the current pattern. The total dose of the current pattern must be limited to a given value, which typically reaches up to 2 mA or 4 mA. Advanced optimization solutions can be obtained via regularized data fitting methods, or projection approaches which target to maximize the current in a given location. This study aims at finding current patterns that would optimize the L1-norm fit between a given focal (well-localized) target current distribution and, at the same time, minimize the dose given to the subject of stimulation. To achieve this, we introduce an L1-norm fitted and regularized linear programming approach (L1L1) method for finding a focal current distribution since L1-norm based solutions are generally known to be well-localized compared to regularized least squares estimation in inverse modelling. While L1-norm has been previously applied to penalize an objective function and linear programming as a strategy to maximize current density at a given location, our method is one of the first to optimize a global L1-norm fit on this application. We hypothesize that the our method can be advantageous for a configuration where a focal current distribution is sought using a given number of electrodes to deliver the stimulus. Solutions from our L1L1 algorithm are sensitive to parameter selection, that is, a wide range of regularization parameters and optimization tolerances are required to be covered to allow the search algorithm to effectively obtain a solution. Therefore, we apply a two-stage lattice search algorithm which finds a set of candidate solutions to optimize the current distribution in the head given a focal vector field, that is, the target of stimulus. We consider the resulting parameter optimization problem as a metaheuristic task of computational intelligence, wherein the goal is to find the best fitting solution with respect to one or more metacriteria which, in this document, are referred as Case (A) ([2.3.1](#sec:case_a){reference-type="ref" reference="sec:case_a"}) for magnitude of the focused current density at the targeted stimulus location, and Case (B) ([2.3.2](#sec:case_b){reference-type="ref" reference="sec:case_b"}) for ratio between focused and nuisance current intensity. We compare the performance of the proposed L1-norm regularized L1-norm fitting (L1L1) method with L1-norm regularized L2-norm fitting (L1L2) version, a L2-norm equivalent fitted version obtained via semidefinite programming, and the Tikhonov regularized Least Squares (TLS) method. On each method, the level of regularization and the relative nuisance field weight, which is explicit in L1L1 and L1L2 and embedded in the objective function in TLS, constitute the parameters of the candidate solution set. The comparison was performed by coupling the CVX optimization toolbox with MATLAB-based Zeffiro Interface (ZI) code package [^1], which allows creating a lead field matrix for a multi-compartment volume head model ([\[app:Lead_Field\]](#app:Lead_Field){reference-type="ref" reference="app:Lead_Field"}) using the Finite Element Method (FEM) together with Complete Electrode Model (CEM) boundary conditions ([\[sec:BConditions\]](#sec:BConditions){reference-type="ref" reference="sec:BConditions"}). As a test domain, we used a realistic head model obtained from an openly available MRI data set ([2.5](#sec:Test_domain){reference-type="ref" reference="sec:Test_domain"}). The numerical results obtained, over a 8 and 20 active electrode montage, supports our initial hypothesis; compared to L1L2 and TLS, the L1L1 is shown to be advantageous with respect to both metacriteria cases. We propose that the metaheuristic L1L1 optimization approach, presented in this study, provides a potential alternative to determine the stimulation montage. This article is organized as follows: methodological details, including the optimization techniques two-stage metaheuristic search, test domain, and target placement are described in Section [2](#sec:Method){reference-type="ref" reference="sec:Method"}. The results can be found in Section [3](#sec:Results){reference-type="ref" reference="sec:Results"} and discussion in Section [4](#sec:Discussion){reference-type="ref" reference="sec:Discussion"}. The mathematical grounds and principles of tES forward modelling and weighted least-squares are explained in [\[app:forward_model\]](#app:forward_model){reference-type="ref" reference="app:forward_model"}.
# Methods {#sec:Method}
The inverse problem of tES is to find a current pattern \({\bf y} = (y_1, y_2, \ldots, y_{\ell})\) that can generate a discretized current field \({\bf x} = (x_1, x_2, \ldots, x_{N})\) in the brain. The current pattern is required to meet Kirchhoff's current conservation conditions, that is \(\sum_{\ell = 1}^{L} y_\ell = 0\) or \({\bf 1}^T {\bf y} = 0\) with \({\bf 1} = (1,1,\ldots,1)^T\), to yield both a total dose \(\| {\bf y} \|_1 = \sum_{\ell = 1}^{L} |y_\ell|\) smaller than or equal to a given safety current limit \({\bm \mu}\) and to have an entry-wise upper-limit of less or equal to \({\bm \gamma}\), i.e., \({\bf y} \preceq \gamma {\bf 1}\). Here, we assume that \(\gamma = \mu/2\), that is, the maximum absolute total dose can be achieved by a system with two or more electrodes---one bearing positive polarity and the other with negative. The fitting between vectors \({\bf y}\) and \({\bf x}\) is enabled by the matrix equation \[\label{optimization_task} {\bf L} {\bf y} = {\bf x},\] where \({\bf L}\) is a linear mapping following from a discretization of Maxwell's equations ([\[app:forward_model\]](#app:forward_model){reference-type="ref" reference="app:forward_model"}). We consider finding optimized current pattern which, when applied into the head model \({\bf \Omega}\) through a given number of active electrodes attached on the scalp, generates a focused volume current distribution matching a synthetic dipolar current at a given orientation and location within the brain, while the nuisance field component remains suppressed. To enable an even comparison between different optimized current patterns, the total dose of each pattern is equaled to \(\mu=\) 4 mA (\(\gamma =\) 2 mA).
## Optimization {#sec:optimization}
To approximately solve ([\[optimization_task\]](#optimization_task){reference-type="ref" reference="optimization_task"}), we consider a weighted optimization scheme, where the equation ([\[optimization_task\]](#optimization_task){reference-type="ref" reference="optimization_task"}) is split into two different components as \[{\bf L} = \begin{pmatrix} {\bf L}_1 \\ {\bf L}_2 \end{pmatrix} \quad \hbox{and} \quad {\bf x} = \begin{pmatrix} {\bf x}_1 \\ {\bf 0} \end{pmatrix}.\] We call the first solution component \({\bf L}_1 {\bf y}\) the *focused field*, i.e., the part that contains the given stimulus target, and the second one \({\bf L}_2 {\bf y}\) the *nuisance field*, i.e., the remaining part of the field which we aim to suppress. To limit the number of non-zero currents in the current pattern \({\bf y}\), the objective function of the optimization task is regularized (penalized) by the norm of the current pattern \({\bf y}\). We control the magnitude difference between \({\bf L}_1 {\bf y}\) and \({\bf L}_2 {\bf y}\) by varying the weight of the nuisance field which is given explicitly for L1L1 and L1L2, and given as a penalty parameter embedded in the objective function for TLS, where explicit constraints are not applicable.
### L1-norm regularized L1-norm fitting {#sec:L1L1}
We propose solving the following L1-norm regularized L1-norm fitting problem (L1L1) \[\min_{{\bf y}} \{ \, \| {\bf L}_1 {\bf y}-{\bf x}_1 \|_1 + \| {\bf L}_2 {\bf y}\|_1 + {\alpha} \zeta \| {\bf y} \|_1 \}, \label{objective_function}\] subject to \(\| {\bf L}_2 {\bf y}\|_\infty \geq \varepsilon \nu\), \({\bf y} \preceq \gamma {\bf 1}\), \(\| {\bf y}\|_1 \leq \mu\), and \(\sum_{\ell = 1}^L y_\ell = 0\). Here \({\bm \alpha}\) is the regularization parameter, \({\bm \zeta} = \| {\bf L} \|_1\) and \({\bm \nu} = \| {\bf x} \|_\infty\) are scaling factors, and \({\bm \varepsilon}\) is the relative weight (numerical tolerance) of the nuisance field. Problem ([\[objective_function\]](#objective_function){reference-type="ref" reference="objective_function"}) constitutes the following linear programming task: \[\min_{{\bf y}, {\bf t}^{(1)}, {\bf t}^{(2)},{\bf t}^{(3)}} \left( \sum_{k = 1}^N t^{(1)}_k + \sum_{m = 1}^M t^{(2)}_m+\alpha \zeta \sum_{\ell = 1}^L t^{(3)}_\ell \right)\] subject to \[\begin{aligned}
-\begin{pmatrix} {\bf t}^{(1)} \\ {\bf t}^{(2)} \\ \alpha \zeta {\bf t}^{(3)} \end{pmatrix} & \preceq & \begin{pmatrix} {\bf L}_1 \\ {\bf L}_2 \\ {\bf I} \end{pmatrix} {\bf y }-\begin{pmatrix} {\bf x }_1 \\ {\bf 0 } \\ {\bf 0} \end{pmatrix} \preceq \begin{pmatrix} {\bf t}^{(1)} \\ {\bf t}^{(2)} \\ \alpha \zeta {\bf t}^{(3)} \end{pmatrix}, \nonumber\\ \begin{pmatrix} {\bf 0} \\ {\varepsilon} \nu {\bf 1}\end{pmatrix} & \preceq & \begin{pmatrix} {\bf t}^{(1)} \\ {\bf t}^{(2)} \end{pmatrix} \nonumber\\ \begin{pmatrix} {\bf 0} \\ {\bf 0} \end{pmatrix} & \preceq & \begin{pmatrix} {\bf t}^{(3)} \\ {\bf 1}^T {\bf t}^{(3)} \end{pmatrix} \preceq \begin{pmatrix} \gamma {\bf 1} \\ \mu \end{pmatrix} \nonumber\\ {\bf 1}^T {\bf y} & = & 0. \label{linear_programming_1} \end{aligned}\] Here, \({\bf t}^{(1)}\), \({\bf t}^{(2)}\) and \({\bf t}^{(3)}\) constitute auxiliary **N**-by-1, **M**-by-1 and **L**-by-1 vectors, respectively. A numerically implementable form of ([\[linear_programming_1\]](#linear_programming_1){reference-type="ref" reference="linear_programming_1"}) with one inequality and equality constraint can be expressed as follows: \[\min_{{\bf y}, {\bf t}^{(1)}, {\bf t}^{(2)}, {\bf t}^{(3)}} \!\! \begin{pmatrix} {\bf 0} \\ {\bf 1} \\ {\bf 1} \\ {\bf 1} \end{pmatrix}^T \!\! \begin{pmatrix} {\bf y} \\ {\bf t}^{(1)} \\ {\bf t}^{(2)} \\ {\bf t}^{(3)} \end{pmatrix}\] subject to \[\begin{aligned}
\begin{pmatrix} {\bf L}_1 &-{\bf I} & {\bf 0} & {\bf 0} \\ {\bf L}_2 & {\bf 0} &-{\bf I} & {\bf 0} \\ -{\bf I} & {\bf 0} & {\bf 0} &-{\bf I} \\ -{\bf L}_1 &-{\bf I} & {\bf 0}& {\bf 0} \\ -{\bf L}_2 & {\bf 0} &-{\bf I} & {\bf 0} \\ {\bf I} & {\bf 0} & {\bf 0} &-{\bf I} \\ {\bf 0} &-{\bf I} & {\bf 0} & {\bf 0} \\ {\bf 0} & {\bf 0} &-{\bf I} & {\bf 0} \\ {\bf 0} & {\bf 0} & {\bf 0} &-{\bf I} \\ {\bf 0} & {\bf 0} & {\bf 0} & {\bf I} \\ {\bf 0} & {\bf 0} & {\bf 0} & {\bf 1}^T \\ \end{pmatrix} \begin{pmatrix} {\bf y} \\ {\bf t}^{(1)} \\ {\bf t}^{(2)} \\ {\bf t}^{(3)} \end{pmatrix} & \preceq & \begin{pmatrix} {\bf x}_1 \\ {\bf 0} \\ {\bf 0} \\-{\bf x_1} \\ {\bf 0} \\ {\bf 0} \\ {\bf 0} \\-\varepsilon \nu {\bf 1} \\ {\bf 0} \\ \gamma {\bf 1} \\ \mu \\ \end{pmatrix} \nonumber\\ {\bf 1}^T {\bf y} & = & 0. \label{eq:LP} \end{aligned}\] The solution is found via primal-dual interior-point algorithm of the SDPT3 package, accessible via the open CVX toolbox[^2].
### L1-norm regularized L2-norm fitting {#sec:L1L2}
For the following L1-norm regularized L2-norm fitting problem (L1L2), \[\min_{{\bf y}} \{ \, \| {\bf L}_1 {\bf y}-{\bf x}_1 \|_2 + \| {\bf L}_2 {\bf y}\|_2 + {\alpha} \zeta \| {\bf y} \|_1 \},\] the L1-norm fitting in ([\[objective_function\]](#objective_function){reference-type="ref" reference="objective_function"}) has been substituted with L2-norm, while the L1-norm regularization and linear constraints ([\[linear_programming_1\]](#linear_programming_1){reference-type="ref" reference="linear_programming_1"}) are the same as in the L1L1 approach. In CVX, the solution is obtained through semidefinite programming incorporating both linear and quadratic constraints of which the latter follow from L2-norm fitting in a straightforward manner.
### Tikhonov regularized least-squares {#sec:TLS}
In Tikhonov regularized least-squares (TLS) estimation, the optimization problem to be solved is \[\min_{{\bf y}}\{ \| {\bf L}_1 {\bf y}-{\bf x}_1 \|^2_2 + \alpha^2 \delta^2 \| {\bf L}_2 {\bf y} \|^2_2 + \alpha^2 \sigma^2 \| {\bf y}\|^2_2\}, \label{TLS_minimization}\] where \({\bm \sigma} = \| {\bf L} \|_2\). To enforce focality, targeting the stimulus location and other areas of the brain, the nuisance field weight \({\bm \delta} \geq 0\) is considered as a variable parameter. The solution of ([\[TLS_minimization\]](#TLS_minimization){reference-type="ref" reference="TLS_minimization"}) is given by the linear system \[\left( {\bf L}_1^T {\bf L}_1 + \delta^2 \alpha^2 {\bf L}_2^T {\bf L}_2 + \alpha^2 \sigma^2 {\bf I} \right) {\bf y} = {\bf L}_1^T {\bf x}_1\. \label{TLS_formula}\] which one can solve numerically using Matlab's backslash (`‘ `) operator. It can be shown that the focused current density\({\bm \Gamma}\) (unit A/m^2^), \[\label{gamma_def} \Gamma = \frac{{\bf x}_1^T {\bf L}_1 {\bf y}}{\| {\bf x}_1 \|_2}\] in the direction of the targeted brain activity yields its maximum when \(\delta = 0\) and that the ratio, \({\bm \Theta}\) (unitless), \[\label{theta_def} \Theta = \frac{ \Gamma}{ \| {\bf L}_2 {\bf y} \|_2/\sqrt{M}}\.\] between \(\Gamma\) and the average nuisance field magnitude increases along with the value of \(\delta\) when \(\delta\) is a small positive number (further details can be found in [\[effect_of_weighting\]](#effect_of_weighting){reference-type="ref" reference="effect_of_weighting"}). Notice that for \(\delta >0\), the minimization problem ([\[TLS_minimization\]](#TLS_minimization){reference-type="ref" reference="TLS_minimization"}) can be written in the following alternative form \[\min_{\bf y}\{ \kappa \| {\bf L}_1 {\bf y}-{\bf x}_1 \|^2_2 + \| {\bf L}_2 {\bf y} \|^2_2 + (\sigma^2/\delta^2) \| {\bf y}\|^2_2\}\] with \({\bm \kappa} = 1/(\alpha^2 \delta^2)\). Therefore, either focused or nuisance field can be weighted.
## Candidate solution set {#sec:candidate_solution}
The above formulations of L1L1, L1L2 and TLS optimization problem depends on the regularization parameter and the nuisance field weight, i.e., the effect the nuisance field component has on the solution of the optimization problem. In both L1L1 and L1L2 methods, the weight is given explicitly while in TLS the nuisance field component is expressed in a weighted form, leading to a different dependence of the solution on the weight as compared to the previous two methods. To find the optimal case-wise parameter combination we examine a two-dimensional 36 \(\times\) 36 lattice of optimized candidate solutions covering a wide 180 dB dynamical range with 5 dB increments for each parameter value. In L1L1, \(\alpha\) and \(\varepsilon\) are varied between-160 and 20 dB, in L1L2 between-140 and 40 dB, and in TLS, the \(\alpha\) and \(\delta\) varied between-240 and-60 dB and-100 and 80 dB, respectively. Parameter variation is considered necessary to obtain the best possible performance, since the scale of the objective function is affected by both parameters. The lattice resolution was selected so that the total computing time was maintained on an acceptable level. The maximal uncertainty related to the lattice was estimated after obtaining the candidate solutions. To compute the candidate solution, we employ a Dell 5820 workstation equipped with Intel Core i9-10900X processor and 256 GB RAM. The total computing time required to evaluate a full lattice of candidate solutions was 7390, 11134, and 138 seconds with L1L1, L1L2, and TLS, respectively. L1L1 and L1L2 utilized a single thread mode while TLS was automatically parallelized by Matlab's interpreter. A relative solver tolerance of 1E-10 was used as a stopping criterion of L1L1 and L1L2.
## Two-stage metaheuristic search {#sec:two_stage_lattice}
To filter the set of candidate solutions, we perform a two-stage metaheuristic lattice search, where focused current density \(\Gamma\) and current ratio \(\Theta\) are used as metacriteria. Of the following two cases, (A) utilizes both criteria, while (B) constitutes a reference for maximizing \(\Gamma\).
### Case (A) {#sec:case_a}
The first stage sets a threshold condition \(\Gamma \geq 0.11\) A/m^2^, and the second stage maximizes the thresholded set of candidates with respect to \(\Theta\). By using these two criteria, we ensure that the selected candidate will have adequate current magnitude in the targeted position and appropriately suppressed nuisance field component at the same time.
### Case (B) {#sec:case_b}
For comparison, we consider a simpler scheme in which the focused current density \(\Gamma\) alone is maximized over the full candidate set. That is, the search is based on a single criterion and stage.
### Post-optimization with non-fixed vs. fixed montage {#sec:post_opt}
Aiming at the best possible optimization outcome, each search run is performed twice: in the first run all the current channels are present in the optimization process while the second one uses a limited montage which is selected based on the first run; the electrodes \(\ell\) with the greatest current \(|I_{\ell}|\) contribution to the total maximum current value are selected to carry non-zero amplitudes, while the remaining ones are opted out (set to zero) from the second run. In the first run, we apply a cap of 128 electrode positions which are reduced to 8 and 20 active channel montages in the second one. These channel counts are inspired by the commercial state-of-the-art tES systems .
## Synthetic sources and placement {#sec:synthetic_source}
The amplitude of the dipolar target current is related to the corresponding local current density in the brain. As reference, the cortex thickness was set with 4 mm (millimeter) and the activity density with 0.77 nAm/mm^2^ (nanoampere per square millimeter). Three 10 nAm dipoles were placed in the following three left hemispheric Brodmann's areas: postcentral gyrus (red), superior temporal gyrus (cyan), and occipital lobe (blue) (Fig. [\[fig:ZI_Parcellations\]](#fig:ZI_Parcellations){reference-type="ref" reference="fig:ZI_Parcellations"}). Each dipole is oriented normally with respect to the surface of the gray matter tissue to satisfy the normal constraint of the brain activity in the cerebral cortex. The two-stage metaheuristic search and the optimization routines were conducted using a individual dipole throughout the head model resulting in three independent numerical solutions; in this document, the results are categorized as *Somatosensory* for postcentral gyrus, *Auditory* for superior temporal gyrus, and *Visual* for occipital lobe.
## Test domain {#sec:Test_domain}
As a test domain of the numerical experiments, we used a multi-compartment volume conductor head model based on openly available anatomical T1-weighted Magnetic Resonance Imaging (MRI) data[^3] obtained from a real subject. The data were segmented using FreeSurfer Software Suite[^4] which distinguishes different head and brain tissue compartments including scalp, skull, cerebrospinal fluid, gray and white matter as well as subcortical structures such as brain stem, thalamus, amygdala, and ventricles with their own complex geometrical properties. The volume segmentation was obtained using ZI's Finite Element (FE) mesh generator which identifies the compartments obtained from the surface segmentation and creates a smoothed and optimized FE mesh composed of these compartments. To discretize the head mode, we use a finite element mesh resolution of 1 mm to obtain physiologically accurate results. The conductivity distribution influences the accuracy of the forward solution. In our model, the conductivity is constant in each tissue compartment with the values corresponding to the set proposed in. The placement of the 128 EEG/tES electrodes (Fig. [\[fig:white_matter_example\]](#fig:white_matter_example){reference-type="ref" reference="fig:white_matter_example"}) follows the International 10-20, 10-10, or 10-5 EEG hardware system. Electrode impedance was set to be 2 kOhm (kiloohms) uniformly. Impedance modelling was enabled by the incorporation of the complete electrode model into the forward model. The tES lead field matrix \({\bf L}\) was generated as explained in ([\[app:Lead_Field\]](#app:Lead_Field){reference-type="ref" reference="app:Lead_Field"}) for 1000 uniformly randomly selected spatial set of points contained by the gray matter compartment, including three Cartesian degrees of freedom per point.
### Accuracy analysis {#sec:Accuracy_analysis}
The optimization outcome was examined by evaluating the maximum injected current in the optimized current pattern \(\| {\bf y} \|_\infty\), the focused current density \(\Gamma\), the current ratio \(\Theta\), and the Angle Difference (AD) between the focused and the targeted fields, i.e., \[\hbox{AD}(\vec{j}_1,\vec{j}_2) = \arccos\,\left(\frac{\left\langle \vec{j}_1,\vec{j}_2\right\rangle}{\left\|\vec{j}_1\right\|\left\|\vec{j}_2\right\|}\right)\] with \(\vec{j}_1\) representing the volume current distribution at the target location generated by the injected pattern, and \(\vec{j}_2\) the dipolar target current, respectively. The limits for lattice-induced deviation of \(\| {\bf y} \|_\infty\), \(\Gamma\), \(\Theta\), AD, were estimated by forming a second order Taylor's polynomial approximation in the 3-by-3 lattice centered at the selected candidate solution. These limits were obtained as the maximum deviation of the polynomial within a co-centered 3-by-3 environment of a hypothetical lattice with double the resolution compared to the actual one.
# Results {#sec:Results}
Numerical comparison between L1L1, L1L2, and TLS have been included in Table [\[table:method_comparison\]](#table:method_comparison){reference-type="ref" reference="table:method_comparison"}. The outcome of their metaheuristic search process are shown in Figure [\[fig:imagesc_L1L1\]](#fig:imagesc_L1L1){reference-type="ref" reference="fig:imagesc_L1L1"}, [\[fig:imagesc_L1L2\]](#fig:imagesc_L1L2){reference-type="ref" reference="fig:imagesc_L1L2"} and [\[fig:imagesc_TLS\]](#fig:imagesc_TLS){reference-type="ref" reference="fig:imagesc_TLS"}, respectively. Stimulation accuracy in Figure [\[fig:stem_comparison\]](#fig:stem_comparison){reference-type="ref" reference="fig:stem_comparison"}, optimized current patterns (as well as their projections in the brain) for the 20 and 8 channel montage in Figure [\[fig:results_20_channel\]](#fig:results_20_channel){reference-type="ref" reference="fig:results_20_channel"} and [\[fig:results_8\_channel\]](#fig:results_8_channel){reference-type="ref" reference="fig:results_8_channel"}, respectively.
[\[table:method_comparison\]]{#table:method_comparison label="table:method_comparison"}
:::
## Results of the metaheuristic search {#sec:Lattice_Search_Results}
For each optimization method, the range of the metaheuristic lattice search was effective enough to cover those regions wherein the magnitude of the focused current density \(\Gamma\) was close to its maximum. These regions were surrounded by comparably smooth transition zones from high to low values, showing the regularity of the optimization process with respect to parameter variation. The solutions obtained through the metaheuristic search case (A), i.e., the maximum of \(\Theta\) for \(\Gamma \geq 0.11\) A/m^2^, were generally found from these transition zones. In the case (B), the maximum of \(\Gamma\) was found from a regular region, where the variation of \(\Gamma\) stayed on a comparably low level. The charts showing \(\Gamma\), \(\Theta\), AD, and \(\| {\bf y} \|_\infty\) for the lattices of the first metaheuristic search run (with 128 electrode positions) together with the corresponding estimates found in the cases (A) and (B) have been included in Figures [\[fig:imagesc_L1L1\]](#fig:imagesc_L1L1){reference-type="ref" reference="fig:imagesc_L1L1"}, [\[fig:imagesc_L1L2\]](#fig:imagesc_L1L2){reference-type="ref" reference="fig:imagesc_L1L2"} and [\[fig:imagesc_TLS\]](#fig:imagesc_TLS){reference-type="ref" reference="fig:imagesc_TLS"}. The charts illustrated are shown with respect to the 20 most intense electric potential channels in the current pattern. The optimizer found in the case (A), where \(\Theta\) is maximized in the second optimization stage for those lattice points which satisfy \(\Gamma \geq 0.11\) A/m^2^ in the first stage, represented by a purple star, while in the case (B), in which the global maximizer of \(\Gamma\) is found, as yellow star. Differences between the first and second runs, as well as electrode montage setups were minor, other charts are not shown. The quantities \(\Gamma\), \(\Theta\), AD, and \(\| {\bf y} \|_\infty\) and their estimated maximal deviations corresponding to the optimizers found are shown in Table [\[table:method_comparison\]](#table:method_comparison){reference-type="ref" reference="table:method_comparison"} and Figure [\[fig:stem_comparison\]](#fig:stem_comparison){reference-type="ref" reference="fig:stem_comparison"}. Compared to L1L2 and TLS, L1L1 yielded a greater or equal value of \(\Theta\) and \(\Gamma\) in cases (A) and (B), respectively. Agreeing with our initial hypothesis, the extra gain provided by L1L1 as compared to L1L2 was observed to be systematic and the most pronounced in case (A), where the optimized \(\Theta\)-value of L1L1 was 1.6 and 1.4 times that of L1L2. Considering TLS, L1L1 yielded a greater maximum current \(\| {\bf y} \|_\infty\) in all the examined cases, while the optimizers tend to deviate overall somewhat more than with TLS. The greatest difference between the L1L1-and TLS-optimized focal current density \(\Gamma\) was obtained for the auditory target current with both 20-and 8-electrode montage for which this difference was 0.032 and 0.022 A/m^2^, respectively. In most cases, the estimated maximum level of deviation does not exceed the observed differences between the optimized values, confirming that the mutual performance differences between the optimization methods. The results concerning AD are somewhat more obscure than for the other examined quantities due to comparably larger estimates for the maximal deviation. The tendency of TLS to result in a smaller AD compared to L1L1 and L1L2 seems obvious. Notably, in a mutual comparison between the results obtained with the 20-and 8-electrode montage, the latter was observed to result in overall smaller mutual differences between the methods. The regularity of the candidate solutions was found to decrease towards the boundaries of the search lattice, which was reflected as a somewhat elevated deviation estimate, when an optimizer was found close to a boundary. In particular, the L1L1 method did not found an optimizer, when the nuisance field weight was greater than one, which can be observed from the charts in Figure [\[fig:stem_comparison\]](#fig:stem_comparison){reference-type="ref" reference="fig:stem_comparison"}.
## Current pattern and volume density {#sec:Results_ROI}
Based on the results, the dependence of the optimization accuracy and deviation on the spatial position and orientation of the target dipole becomes evident while the current patterns obtained via each applied method maintain their general characteristics regardless of the positioning of the target dipole. L1L1 and L1L2 tend to find a pattern where a large part of the stimulus current is driven through a comparably few electrodes in the current pattern as compared to TLS where the current amplitudes have smoother transitions between the electrodes. Consequently, L1L1-and L1L2-optimized current patterns are also likely to have a greater maximum current \(\| {\bf y} \|_\infty\). Moreover, L1L1 and L1L2 solutions tend to include relatively many low-amplitude currents with close-to-equal amplitudes, which distributes the nuisance field current density over a large area decreasing its amplitude. This can be interpreted as a consequence of the relatively large maximum current and is demonstrated by the results of the 20-electrode montage in Figure [\[fig:results_20_channel\]](#fig:results_20_channel){reference-type="ref" reference="fig:results_20_channel"}, while it is a somewhat less pronounced phenomenon with 8-electrode montages. In L1L1 and L1L2, the anodal and cathodal electrodes tend to be further apart and the whole pattern is likely to be wider than in TLS, especially, in the pattern obtained via optimization strategy (B). Thus, such patterns can be interpreted as beneficial for enhancing the focused current density \(\Gamma\) which is maximized in (B). Finally, L1L1 seems to find the most focal current density with a comparably large threshold for \(\Gamma\), allowing for finding a focal stimulus with a relatively high focal current density. This is reflected by the results obtained via optimization strategy (A), where the current ratio \(\Theta\) was greater for L1L1 than for L1L2 or TLS. The improved focality obtained with L1L1 as compared to TLS was observed with 20-electrode montage for somatosensory and auditory dipole, where the L1L2 and TLS solutions for cases (A) and (B) are mutually similar and less focal than the L1L1 solution for (A). For the 8-electrode montage the difference between L1L1 (A) and TLS (A) is less distinct, while L1L1 (A) yields a greater \(\Theta\) with a larger marginal than the estimated maximal deviation regardless the electrode count.
# Discussion {#sec:Discussion}
This study considered L1-norm data fitting via L1-norm regularized convex optimization (L1L1) as a potential alternative optimization approach for finding stimulus currents for a multi-channel tES exercise in comparison with both L2-norm data fitting counterpart (L1L2), and the weighted Tikhonov regularized least squares method (TLS). L1L1 has been earlier suggested as a tool to maximize the focused current for a given location and a given set of electrodes. We proposed the means to optimize global current density fitting to obtain the best possible localization of the stimulus current distribution and pattern. The present topic is important due to the general tendency of L1-norm fitted solutions to be sparse compared to those obtained via L2-norm methods. In tES, this means a greater focused current density driven in the targeted brain region. We applied L1-norm in both fitting and regularization, i.e., penalization of the non-zero entries in the current pattern, hypothesizing that the resulting volume current distribution and the current pattern are sparse. We considered this approach necessary to obtain the best possible fit for a given, user-defined number of active electrodes, from a montage with a minimal set of two channels, i.e., the standard two-patch tDCS, to a significantly higher version. We apply our study into montages of 8 and 20 active electrodes based on commercial clinically-applied tES systems. We defined an explicit parameter to steer the weight of the nuisance field distribution in L1L1 and L1L2, the reason being that with zero-weighting the maximum current in the targeted location is lower than with an appropriately chosen weight. In TLS, the weight is incorporated as the multiplier of the nuisance field term as shown in [\[effect_of_weighting\]](#effect_of_weighting){reference-type="ref" reference="effect_of_weighting"}. The results suggest that the present L1L1 optimization technique performs appropriately for different target dipole locations and active electrode counts. L1L1 allowed finding a focal current pattern and well-localized stimulus current density in the brain with a focused maximum greater than 0.11 A/m^2^ which can be considered as adequate in tES with the maximum current dose 2 mA. Compared to L1L2 and TLS, an enhanced focused vs. nuisance field current ratio \(\Theta\) was obtained. This result is in accordance with our initial hypothesis on the potential benefit of L1L1 in localizing the stimulus current and the general knowledge that L1-norm optimization is advantageous to enhance contrasts between different solution components. The comparison between L1L1 and L1L2 was particularly important to enlighten the enhanced potential of L1-norm fitting in maximizing the ratio \(\Theta\), i.e., in the suppression of the nuisance currents. The search case (B) suggest that the present L1L1 yields a greater focused current density *per se* compared to TLS, which is in agreement with the earlier observations. The maximum obtained with L1L1 being was also systematically greater than that of L1L2, however, by a small margin. Notably, in this study, the greatest focused current amplitude is found with a non-zero nuisance field weight, which highlights the feasibility and importance of the present convex optimization approach, where both the current density field anywhere in the brain can contribute the solution of the optimization process: fitting the focused field alone will not yield the best possible optimization outcome. Akin to the volume fields in the brain, the current patterns found using L1L1 were more concentrated and had, overall, a greater contrast than those obtained using L1L2 or TLS. In L1L1 and L1L2, the anodal and cathodal electrodes have greater current amplitudes and tend to be further apart from each other and less clustered than in TLS. In particular, L1L1 was shown to find a current pattern with a steep contrast between the anodal and cathodal electrodes while suppressing the nuisance currents in the brain, hence, providing a potential alternative to modulate the effects of the stimulation, e.g., the sensation experienced by the subject. These observations might be significant regarding physiological impediments, hardware constrains, or concentrations of skin irritability that the subject experiences during the stimulation session. Thus, an L1L1-based electrode montage and current pattern might provide a potential alternative, if there is a need to modulate the stimulation configuration due to various effects it may cause. The optimization parameter ranges covered in the metaheuristic optimization process included the neighborhoods of the global maximizers for \(\Gamma\) and \(\Theta\), while the \(36 \times 36\) search lattice allowed finding the full set of candidate optimizers in a relatively short computing time which was 7390, 11134, and 138 seconds with L1L1, L1L2, and TLS, respectively. Notably, our L1L1 implementation ran faster than L1L2, obviously, as L1L1 does not include additional quadratic constraints which are necessary in a semidefinite formulation of a linearly constrained problem. To improve the computational performance of the metaheuristic L1L1 or L1L2 process, one can consider computing multiple candidate solutions simultaneously, as the CVX optimizer is a single-thread process and allows for a straightforward parallelization in a multi-core processing environment. In addition to being a simpler method, TLS is automatically parallelized by Matlab's interpreter, as it includes only full matrices algebraic simple linear algebraic operations, which in part explains the computing time differences to L1L1 and L1L2 optimization. Our estimates for the maximum lattice-based deviation of the optimized quantities suggest that the current computing accuracy is high enough to demonstrate the major differences between the L1L1, L1L2 and TLS method and to verify our initial hypotheses on the performance of L1L1. Evidently, the applied linear programming algorithm itself might also affect the optimization outcome. The primal-dual interior-point algorithm of the SDPT3 solver was found to perform robustly in this study but, for example, MATLAB's built-in dual-simplex algorithm, which was also tested in solving the L1L1 task, was found to result in a less robust outcome. Potential future work would be to compare the present metaheuristic and CVX/SDPT3-based L1L1 and L1L2 implementations with an alternative optimization algorithm, most prominently, the Alternating Direction Method of Multipliers (ADMM). Our metaheuristic optimizers, which are available in ZI, can also potentially be extended to include other non-invasive and invasive brain stimulation modalities to enhance the present electrical stimulation toolbox, for instance, Deep Brain Stimulation (DBS) applications. Further investigations of the relationship between the explicit nuisance field weight \(\varepsilon\) (tolerance) and modelling or other uncertainties is an important topic, as the uncertainty may be expected to limit the maximal obtainable accuracy for the nuisance field. Based on the present results, the current ratio \(\Theta\) is optimized with a considerably lower value of \(\varepsilon\) as compared to the focused current density \(\Gamma\). Thus, the maximal obtainable \(\Theta\) can be expected to decrease along with an increasing uncertainty, while the maximal \(\Gamma\) can be assumed to be less affected by that. Finally, experimental work will obviously be needed to learn about other than the mathematical or computational aspects of L1L1-optimized current patterns in practice. | {'timestamp': '2022-06-06T02:17:09', 'yymm': '2206', 'arxiv_id': '2206.01660', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01660'} |
# Introduction {#introduction .unnumbered}
Ordered dynamical phases of motile organisms are ubiquitous in nature across all scales, from bacterial colonies, to insect swarms and bird flocks. In particular, self-organization into vortex patterns is often observed and has been attributed to some local external attractor, e.g., light or nutrient concentration, together with behavioral rules like collision avoidance and mutual alignment. Pertinent social interactions are commonly thought to be based on perception and the ability to actively control the direction of motion. They are also generally presumed to provide some benefits to the individual and to the collective, as in the case of collision avoidance or predator evasion. However, since such interactions are usually derived only indirectly and approximately from observations, it is arguably useful to coarse grain them, e.g., into simple alignment rules. This strategy has been successful in physics in order to rationalize complex collective effects with the help of simple mechanistic models, in particular with respect to emerging universal traits. It is also supported by the observation that biological many-body assemblies often appear highly susceptible to environmental influences and exhibit a dynamical finite-size scaling reminiscent of critical states in inanimate physical systems. Importantly, the cascades of complex biochemical/biophysical processes needed to transform signal perception into a navigational reaction inevitably result in retarded interactions upon coarse-graining . This generic complication is often dismissed in the analysis, and dedicated models and experiments addressing the role of time delays in active matter are still rare, although these have occasionally been shown to fundamentally alter the collective dynamics and to bring it closer to that found in nature. To a first approximation, delay effects can resemble inertial corrections to an otherwise overdamped biological dynamics. In particular, both have a propensity to give rise to oscillations, and inertia moreover to rotational motion around an attractive center, as familiar from planetary orbits. Experiments that can assess or even deliberately control retarded interactions in living systems turn out to be difficult. But by imposing time delays onto synthetic active particles via computer control, we can create an ideal laboratory system to experimentally emulate such situations. Suitable feedback control techniques for active particles have recently become available through photon nudging . The technique allows to adjust a particle's propulsion speed to acquired real-time information (positions, directions of motion) about the dynamical state of an ensemble. It has previously been employed to rectify the rotational Brownian motion for particle steering and trapping, to explore orientation-density patterns in activity landscapes, and to study information flow between active particles and their emerging critical states. Beyond what related computer simulations accomplish, these experiments additionally incorporate the full real-world complexity arising from actual physical interactions due to hydrodynamic, thermal, or concentration fields. In the following we describe experiments with feedback-controlled active Brownian microswimmers "aiming" to reach a fixed target by a retarded thermophoretic self-propulsion. The resulting systematic navigational "errors" are seen to cause a spontaneous symmetry breaking to a bi-stable dynamical state, in which the swimmers self-organize into a merry-go-round motion that switches transiently between degenerate chiralities.
# Results {#results .unnumbered}
#### Single Particle Retarded Interaction {#single-particle-retarded-interaction .unnumbered}
The basic component of a swarm is a single active particle whose direction of motion depends dynamically on its environment. A perturbation of the particle position leads to an adjustment of the direction of motion, which can inevitably only take place with a delay between detection and reaction to the perturbation. In this work, we study the simplest case, where an active particle moves toward an immobilized target particle of the same size. Assuming that the active particle has a programmed time delay \(\delta t\) in its response, its propulsion direction \(\hat{\mathbf{u}}(t)\) at time \(t\) is determined by its relative position to the target particle it \"sensed\" at time \(t-\delta t\) in the past according to \[\label{eq:target_cohesion} \hat{\mathbf{u}}(t) = \frac{-\mathbf{r}(t-\delta t)}{|\mathbf{r}(t-\delta t)|},\] where \(\mathbf{r}\) is the location of the active particle with respect to the target particle at the origin. This interaction rule is implemented in an experimental feedback system that controls the propulsion of active particles. Here, the active particles are polymer particles with a radius of \(a=1.09\,\si{\micro\metre}\) coated with gold nanoparticles and suspended in a thin film of water. A laser with a wavelength of 532 nm is focused on the active particle with a displacement \(d\) from the center (Fig. [\[fig:figure1\]](#fig:figure1){reference-type="ref" reference="fig:figure1"}B) to drive the particle at a constant velocity \(v_0\) in the direction defined by Eq. [\[eq:target_cohesion\]](#eq:target_cohesion){reference-type="ref" reference="eq:target_cohesion"} via self-thermophoresis. A darkfield microscopy setup is used to image the particles (Fig. [\[fig:figure1\]](#fig:figure1){reference-type="ref" reference="fig:figure1"}C). A feedback loop running on a computer analyzes and records the positions of the particles to control the laser position accordingly via an acousto-optical deflector. We use a separate calibrator particle running on a quadratic trajectory as a reference for the velocity \(v_0\) of a free particle. Further details are described in the Methods section and the Supporting Materials. When the system is started with a programmed time delay \(\delta t=0\,{\rm s}\), the active particle moves towards the target particle until it collides with it. The motion of the active particle is then constrained by the target particle carrying out a diffusive motion around the target's circumference. Its distance to the target particle obeys the barometric distribution. As the delay increases, the amplitude of this diffusive motion grows until a critical delay is reached and the particle begins to rotate around the target (see Supplementary Movies 1--3). We characterize this dynamics with the help of the angle \(\theta\) between the direction of motion in Eq. [\[eq:target_cohesion\]](#eq:target_cohesion){reference-type="ref" reference="eq:target_cohesion"} and the negative radial direction \(-\mathbf{r}(t)\) (see Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A). Note that \(\sin(\theta)\) corresponds to the contribution of a single active particle \(i\) to the rotational order parameter commonly used to describe collective rotational motion, i.e., \(o_{R,i}=(\hat{{\bf r}}_i\times \hat {\bf u}_i)\cdot {\bf e}_z=\sin(\theta_i)\). Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A shows the experimental trajectories of this propulsion angle \(\theta\) for a single active particle with \(v_0=2.16\,\si{\micro\metre\per\second}\) and three different delays. For short delays, \(\theta\) fluctuates with a small amplitude around zero (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A top). The fluctuations increase with the delay and lead to a flat-top probability density of the propulsion angle for \(\delta t \approx 0.87\, {\rm s}\) (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A middle). At larger delays (\(\delta t=1.14\, {\rm s}\)), the propulsion angle fluctuates around a stable nonzero value which intermittently changes between a clockwise and counter-clockwise rotation sense of rotation (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A bottom). In this case, the probability density \(p(\theta)\) becomes bi-modal (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}C). The periods of the swimmer staying in either stably rotating state increase in duration when the delay is further increased. At \(\delta t=1.4\, \si{\second}\), the propulsion angle fluctuates around \(\pm 80 \si{\degree}\). Under these conditions, the cohesion of the particle to the target becomes weak as the particle velocity is almost tangential to the target particle circumference. As a result, the distance \(R\) of the particle to the target starts to fluctuate more strongly, as shown in the position histograms in Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}B. The non-zero propulsion angle is the result of an angular displacement during the period \([t-\delta t,t]\) of the active particle, and we can write it as \[\label{eq:deftheta} \theta(t) = \int^t_{t-\delta t} \omega(t') \,\mathrm{d} t' = \phi(t)-\phi(t-\delta t) = \angle (\hat{\mathbf{u}}(t),-\mathbf{r}(t)).\] Here, \(\phi(t)\) is the polar angle of the active particle in polar coordinates centered in the target particle and \(\omega(t) = \dot \phi(t)\) is the corresponding instantaneous angular velocity (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A). The observed dynamics can be understood by considering the active and target particles in physical contact. Their distance is then constrained to be the sum of their radii (\(R=2a\)). In this case, the active particle moves along the circumference of the target particle with an angular velocity \(\omega(t)\) according to \(\omega(t) =\omega_0\sin( \theta(t))\), where \(\omega_0= v_0/R\) and the angle \(\theta(t)\) is defined in Eq. [\[eq:deftheta\]](#eq:deftheta){reference-type="eqref" reference="eq:deftheta"}. The term \(\omega_0 \sin(\theta(t))\) originates from projecting the propulsion direction in Eq. [\[eq:target_cohesion\]](#eq:target_cohesion){reference-type="eqref" reference="eq:target_cohesion"} parallel to the surface of the target particle at the point of contact. As sketched in Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}A, assuming a constant angular velocity \(\omega\) with \(\theta = \omega\delta t\), the solutions to the equation for \(\theta\) are given by the intersections of a sine function and a linear function, \[\label{eq:simple} (\omega_0 \delta t)^{-1}\theta=\sin(\theta).\] For \(\omega_0\delta t <1\), there is a single intersection at \(\theta\) = 0, indicating a non-rotating state. For \(1<\omega_0\delta t<\pi/2\), one non-rotating and two rotating solutions with opposite rotation senses arise. These rotations are stable under perturbations (see SI), while the non-rotating solution is unstable. For \(\omega_0\delta t >\pi/2\), the rotating solutions correspond to \(|\theta|>\pi/2\), and the radial component of propulsion becomes positive (repulsive), driving the active particle away from the target particle. As a result, the radius of the orbit, \(R\), increases until a new stable orbit with \(R = 2 v_0 \delta t/\pi>2a\) and \(|\theta|=\pi/2\) is reached. For infinitely small particles (\(a\to 0\)), the distance of the swimmer to the target position can thus, in principle, vanish (\(R\to 0\)), and the rotation can occur at infinitely short programmed delays (\(\delta t\to 0\)). The retarded attraction hence always leads to rotation, and the observed transition from the non-rotating to the stable rotating state is caused by the minimum distance to the target. In the experiment, the minimum distance is determined not only by the particle radius \(a\) but also by the Brownian motion of the active particle within the instrumental delay of the feedback loop, and it is nonzero even for \(a>0\). Adding Brownian fluctuations to the deterministic motion of the active particle described by Eq. [\[eq:simple\]](#eq:simple){reference-type="ref" reference="eq:simple"}, results in the non-linear delayed stochastic differential equation \(\dot{\phi}(t) = \omega_0 \sin\left(\phi(t)-\phi(t-\tau)\right) + \sqrt{2D_0/R^2} \, \eta(t)\), where \(D_0 \approx 0.0642\) μm^ -1^ denotes the translational diffusion coefficient of the active particle and \(\eta(t)\) white noise. This equation is very difficult to solve. However, as detailed in the supplementary information, after several approximations it yields the simple overdamped Langevin equation \[\dot{\theta} =\frac{1}{3 \delta t}\left[\theta_{\pm}^2-\theta^2\right]\theta + \sqrt{2D_\theta}\eta \label{eq:domega_approx_overdamped}\] with \[\theta_{\pm} = \pm \sqrt{\frac{6}{\theta_0}(\theta_0-1)} \label{eq:roots}\] and \(\theta_0 \equiv \omega_0 \delta t\). The noise term in Eq. [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} describes the angular Brownian motion of the active particle around the target with an effective diffusion coefficient \(D_\theta\), which is a single free parameter in our theory. Eq. [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} yields the stationary solutions \(0\) and \(\theta_{\pm}\), consistent with our previous simplified model. All the three solutions coincide when \(\theta =\theta_0 = \omega_0\delta t = 1\), implying a transition from a non-rotating to two rotating states. The data points in Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B display the experimentally obtained maxima of the histograms \(p(\theta)\) of the propulsion angle (see Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}C) as a function of \(\omega_0\delta t\). While the transition to the two rotating states is expected at \(\omega_0\delta t\) = 1, the transition points in experiments are located at lower values due to an additional instrumental delay \(\Delta t\) in the feedback loop of the experimental setup. This instrumental delay between the most recent exposure to the camera and the laser positioning affects the motion direction, causing an earlier onset of the transition to a stable rotation. The dashed line in Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B shows the theoretical prediction, which includes the instrumental delay \(\Delta t\) and the programmed delay \(\delta t\), as detailed in the SI (Eq. (17)). The Langevin equation [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} can be interpreted as a dynamical equation for the position \(\theta\) of an overdamped Brownian particle with friction and diffusion coefficients \(\gamma\) and \(D_\theta\) in a quartic potential (see derivation in the SI), \[U(\theta)=\frac{\gamma}{\delta t}\left[ \left(\frac{1}{\theta_0}-1\right )\theta^2 +\frac{1}{12}\theta^4 \right], \label{eq:potential}\] which corresponds to the generic case of a supercritical pitchfork bifurcation. The potential can be extracted from the experimental data (Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B) by fitting the histogram \(p(\theta)\) with the Boltzmann distribution \(\exp(-U(\theta)/ \gamma D_\theta)/Z\) including an effective temperature \(\gamma D_\theta/k_{\rm B}\) and normalization factor \(Z\). The effective temperature thus represents a scaling factor that brings the measured potential of mean force \(-\gamma D_\theta \log p(\theta)\) to the theoretical prediction in Eq. [\[eq:potential\]](#eq:potential){reference-type="eqref" reference="eq:potential"}. The self-generated quartic potential in Eq. [\[eq:potential\]](#eq:potential){reference-type="eqref" reference="eq:potential"} emerging from the delayed response finds its mathematical analogue in the free energy function in Landau's theory of second-order phase transitions , with the control parameter \(\theta_0\). This correspondence suggests that \(\mathcal{T} = (\omega_0\delta t)^{-1}=\theta_0^{-1}\) is the equivalent of a temperature with a critical point \(\mathcal{T}_C\) = 1. Both the activity \(\omega_0\) and the delay \(\delta t\) in the product can lead to the transition. Hence, the active particle velocity and delay are inversely coupled, and large particle velocities need only short delays causing an ordered rotating state. Yet, our system does not correspond to a thermodynamic phase, where the transitions are of collective nature. The bifurcation and the potential are caused by the delayed response of the single active particle to an external signal including the steric repulsion by the target particle. A stability analysis of Eq. [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} (see SI) yields the susceptibility \[\label{eq:susceptibility} \chi\left (\frac{1}{\theta_0} \right) = \begin{cases} \frac{\delta t}{2\left (\frac{1}{\theta_0}-1\right )} & \text{if \(\frac{1}{\theta_0}>1\)}\\ -\frac{\delta t}{4\left (\frac{1}{\theta_0}-1\right )} & \text{if \(\frac{1}{\theta_0}<1\)} \end{cases}\;,\] as familiar from Landau theory (for \(\gamma = 1\) s). The susceptibility corresponds to the relaxation time of the system after a perturbation from a stable state (see SI). It is determined from the experiments by measuring the autocorrelation function \[\label{eq:autocorr} C(\tau)=\frac{\langle \delta \theta(t+\tau) \delta \theta(t)\rangle_t}{\langle\delta \theta(t)^2 \rangle_t}\] of fluctuations of the propulsion angle \(\delta \theta(t)=\theta(t)-\langle\theta(t)\rangle\). Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}C shows the decay time \(t_R\) obtained from the autocorrelation function as \(C(t_R)\) = 1/e (symbols) together with susceptibilities predicted theoretically without (Eq. [\[eq:susceptibility\]](#eq:susceptibility){reference-type="eqref" reference="eq:susceptibility"}, dashed line) and with (see SI, solid line) instrumental delay \(\Delta t\). Close to the transition point (\(\omega_0 \delta t \approx 1\)), the theory suggests a slowing down of the relaxation due to an increasingly flat potential (e.g., the potential plot in the middle of Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B). The experimental results nicely confirm this prediction without any fitting parameter. In addition to calculating the susceptibility from the autocorrelation function, we also performed relaxation measurements after a step-like perturbation of the propulsion angle \(\theta\). The results of the measurements agree well with the theoretical predictions, as shown in Sec. 4 of the SI, together with details on the evaluation of the susceptibilities. While both stable rotation directions can be inferred from a purely deterministic model excluding Brownian motion, the observed spontaneous reversal of the rotation direction is driven by fluctuations in the propulsion angle and thus by the noise in the system. The changes in the rotation direction correspond to transitions between the minima \(\pm\theta_{\pm}\) of the self-generated potential, Eq. [\[eq:potential\]](#eq:potential){reference-type="eqref" reference="eq:potential"}. We may thus apply Kramers' theory to estimate the corresponding transition rate as \[k = \frac{\sqrt{2}}{\pi}\frac{|\theta_0-1|}{\theta_0 \delta t}\exp\left[-\frac{3}{\gamma D_\theta}\frac{(\theta_0-1)^2}{\theta_0^2}\right]. \label{eq:rate}\] The effective temperature \(\gamma D_\theta/k_{\rm B}\) driving the fluctuations in the potential is obtained from the previously mentioned scaling of the theoretical prediction to the measured potential of mean force. Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}D displays the results of experiments measuring the transition rate from mean residence times of \(\theta\) in the two potential wells. The rate is compared to the prediction from the Kramers theory, Eq. [\[eq:rate\]](#eq:rate){reference-type="eqref" reference="eq:rate"}, and shows a good agreement.
## Multiple Particles {#multiple-particles .unnumbered}
As demonstrated in the previous section, the rotation observed in our experiments results from a spontaneous symmetry breaking in the dynamics of a single active agent under a retarded self-propulsion to a target, different from standard models of rotational dynamics in overdamped systems, which assume mutual or \"social\" interactions among the agents. Hence, when more active particles are added to the system, each of them strives to exhibit the same rotation and bifurcation as the single swimmer. However, steric, hydrodynamic, and thermophoretic interactions among the particles synchronize and stabilize their motion so that the system exhibits collective behavior. Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"} summarizes the key results obtained for an ensemble of 15 active particles attracted to the target particle with the same delay. For the considered range of time delays, the active particles form two tightly packed shells around the target particle (Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}A). The typical distance of the inner shell particles to the target is about half that of the outer shell, \(R^\mathrm{out} \approx 2R^\mathrm{in} = 4a\). So the inner shell is expected to enter the rotational phase with a correspondingly shorter delay than the outer shell. However, the interparticle interactions in the compact cluster strongly correlate the particle motion, which quantitatively changes the results obtained for a single particle. Compared to the theoretical prediction, \(\omega_0 \delta t\) = 0.73, we observe that for \(v_0\) = 2.06 \(\si{\micro\metre\per\second}\) the transition to the rotational phase of the inner shell is postponed to \(\omega^{\mathrm{in}}_0 \delta t \equiv v_0\delta t/ R^\mathrm{in}\approx\) 0.83, corresponding to \(\delta t\) = 0.9 s (see the rightmost red data point lying on the horizontal axis in Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}B). Close to the transition, the inner shell exhibits alternating periods of rotation and non-rotation, whereas the non-rotating outer shell compresses the inner shell due to its inwards pointing propulsion direction (Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C, left). Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C displays the velocity fields of the particles averaged over their trajectories with three different delays. The rotation of the outer shell is observed at \(\omega^{\mathrm{out}}_0\delta t \equiv v_0\delta t/ R^\mathrm{out} \approx\) 0.41, corresponding to the same delay \(\delta t\) = 0.9 s when the inner shell starts to rotate (see Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}B and Supplementary Movies 4--6). For delays slightly above the transition, \(0.9<\delta t\) \(<\) 1.41 s, the two shells rotate in opposite directions, as shown in the middle plot of Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C. The simultaneous transition and the counter-rotation of the two shells suggest that the inner shell particles generate forces in the opposite direction to their propulsion, repelling the outer shell particles and causing a bias on their rotation, as depicted in Figs. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}D--F. The bias is presumably caused by the directional hydrodynamic and thermophoretic interactions. The surface temperature gradient across the particle creates a thermo-osmotic surface flow that propels the particle. If the particle motion is (partly) halted by an external force such as the steric force applied by the immobilized target particle, the particle acts as a pump creating a hydrodynamic outflow at the hot side of the swimmer (Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}D and SI). Similarly, thermophoretic interactions arise from temperature gradients across the surface of an active particle caused by its neighboring particles. They are commonly repulsive as found, e.g., for Janus particles in external temperature gradients. We have carried out finite element simulations of the flow field around a mobile and an immobile self-propelling swimmer (see SI). The overall near field hydrodynamic interactions are nevertheless complex due to the presence of the other particles and the substrate surface, and further depend on the propulsion angle \(\theta\). The inner shell propulsion angle increases with the delay, which results in changing direction and magnitude of the inter-shell interaction and thus the outer shell rotation bias, which presumably varies as sketched in Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}F (see SI). For \(\delta t \geq 1.41\) s, the two shells predominantly rotate in the same direction, as shown in Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C, right. The transition from counter-to co-rotation shells corresponds to the sign flip of the bias, occurring at \(\theta^\mathrm{in}\approx\) 67\(^\circ\). While we currently cannot separate thermophoretic and hydrodynamic effects in the experiment, possible hydrodynamic interactions yield an interesting perspective as they are supposed to influence also the collective behavior of fish or birds. At longer retardation, \(\theta^\mathrm{in}\) tends to reach 90\(^\circ\), and thus the inner shell tries to expand against the compressing outer shell. These competing tendencies lead to particle exchange between the two shells.
# Discussion {#discussion .unnumbered}
We have demonstrated above that the motion of an active particle induced by the delayed attraction to a target point can spontaneously undergo a transition from a diffuse isotropic state to a dynamical chiral state, upon increasing the activity and/or the delay time. The transition is well described by a pitchfork bifurcation accompanied by a characteristic critical slowing down of the response. The single-particle dynamics thus already exhibits non-trivial features more commonly associated with (mean-field) phase transitions in strongly interacting passive many-body systems . This can be explained by noting that the deterministic part, \(\dot{\phi}(t) =\omega_0\sin\left (\phi(t)-\phi(t-\delta t)\right)\), of our stochastic delay differential equation can also be understood as the dynamical equation for a single Kuramoto phase oscillator, with vanishing eigenfrequency and coupling strength \(\omega_0\), which is trying to synchronize with its own past state. In the chiral state, the particle orbits around the central obstacle. This orbiting motion is stable against noise, but its chirality is only transiently maintained. This should be contrasted to the chiral states resulting from non-reciprocal coupling in the Kuramoto model as discussed by Vitelli et al., which require many-body cooperativity. Despite the similarity of the two models, we are clearly observing a single-particle effect, instead. The lack of many-body cooperativity is compensated for by the infinite number of relaxation modes encoded in the time-delayed equation of motion. The nonlinear dynamics of our experimental system can be described by an approximate analytical model resulting in an emergent self-generated quartic potential. While such potentials are frequently found in descriptions of phase transitions and collective effects in active particle ensembles, following various behavioral rules, this is not the case, here. Due to the activity and the (programmed) delay, it already occurs for a single active particle aiming at a spatially fixed target. In a whole swarm of particles that are all attracted to a common target, which might be its own perceived center of mass, the single-particle bifurcation is preserved. Inter-particle collisions merely synchronize and stabilize the rotational states of the individual particles. Upon close contact, hydrodynamic and thermophoretic interactions become important and help the swimmers to self-organize into co-and counter-rotating orbits. In biological motile ensembles, from bacteria to fish, similar hydrodynamic mechanisms may be at work, although precise details and scales may differ widely. The corresponding many-body effects can be subtle and may elude coarse grained simulations and theories. This underscores the importance of well controlled experimental model systems that may act as "hybrid simulations", combining computer-controlled active particles with real-world environments. While time delays are an unavoidable outcome of coarse-graining microscopic descriptions of the feedback processes in natural systems, they are often neglected in low-dimensional models of active particle collective effects. In this respect, our model system provides a new perspective, as it takes the unavoidable systematic delays in the dynamics seriously and explores their generic effects. It thereby provides a microscopic underpinning for empirical vision-cone models that have been found to produce rotational dynamics in overdamped flocks or swarms. In overdamped systems, retardation thus plays a similar role as added inertia. Both effects lead to persistence in the particle dynamics, but the effects of the time delay may be much richer . While we considered only a positive delay, i.e., synchronization with the past, sophisticated biological organisms also have predictive capabilities at their disposal that allow them to extrapolate the current state into the future. These can to a first approximation be incorporated in the form of a negative time delay. The inclusion of positive and negative delays may therefore also provide a new perspective on claims that collision and alignment interactions, as well as other complex rules, are the sources of the emerging complex adaptive responses observed in living many-body systems.
# Methods {#methods .unnumbered}
## Sample Preparation
Samples were prepared using two glass coverslips (20 mm \(\times\) 20 mm, 24 mm \(\times\) 24 mm) to confine a thin liquid layer (3 \(\si{\micro\metre}\) thickness) in between. The edges of one coverslip are sealed with a thin layer of PDMS (polydimethylsiloxane) to prevent leakage and evaporation. The liquid film used in the sample is composed of 2.19 \(\si{\micro\metre}\) diameter gold-coated melamine formaldehyde (MF) particles (microParticles GmbH) dispersed in 0.1% Pluronic F-127 solution. The latter prevents the cohesion of the particles and adsorption to the cover slide surface. The surface of the MF particles is uniformly scattered with gold nano-particles of about 8 nm diameter with a total surface coverage of about 10% (Fig. S3A). SiO\(_2\) particles (2.96 \(\si{\micro\metre}\) in diameter, microParticles GmbH) are added into the solution to keep the thickness of the liquid layer at about 3 \(\si{\micro\metre}\). Finally, 0.3 \(\mathrm{\mu l}\) of the mixed particle suspension is pipetted on one of the coverslips, and the other is put on top.
## Experimental Setup
The experimental setup (see SI) consists of an inverted microscope (Olympus, IX71) with a mounted piezo translation stage (Physik Instrumente, P-733.3). The sample is illuminated with an oil-immersion darkfield condenser (Olympus, U-DCW, NA 1.2--1.4) and a white-light LED (Thorlabs, SOLIS-3C). The scattered light is imaged by an objective lens (Olympus, UPlanApo \(\times\)`<!-- -->`{=html}100/1.35, Oil, Iris, NA 0.5--1.35) and a tube lens (250 mm) to an EMCCD (electron-multiplying charge-coupled device) camera (Andor, iXon DV885LC). The variable numerical aperture of the objective was set to a value below the minimum aperture of the darkfield condenser. The microparticles are heated by a focused, continuous-wave laser at a wavelength of 532 nm (CNI, MGL-III-532). The beam diameter is increased by a beam expander and sent to an acousto-optic deflector (AA Opto-Electronic, DTSXY-400-532) and a lens system to steer the laser focus in the sample plane. The deflected beam is directed towards the sample by a dichroic beam splitter (D, Omega Optical, 560DRLP) and focused by an oil-immersion objective (Olympus, UPlanApo \(\times\)`<!-- -->`{=html}100/1.35, Oil, Iris, NA 0.5--1.35) to the sample plane (\(w_0 \approx 0.8\,\si{\micro\metre}\) beam waist in the sample plane). A notch filter (Thorlabs, NF533-17) is used to block any remaining back reflections of the laser from the detection path. The acousto-optic deflector (AOD), as well as the piezo stage, are driven by an AD/DA (analog-digital/digital-analog) converter (Jäger Messtechnik, ADwin-Gold II). A LabVIEW program running on a desktop PC (Intel Core i7 2600 4 \(\times\) 3.40 GHz CPU) is used to record and process the images as well as to control the AOD feedback via the AD/DA converter.
# Author Contributions {#author-contributions .unnumbered}
XW and FC conceived the experiment. XW carried out the experiment. XW, FC, PC, VH analyzed the data. PC and VH developed the theory. XW, PC and FC carried out simulations. All authors discussed the results and wrote the manuscript.
# Introduction {#introduction-1 .unnumbered}
Ordered dynamical phases of motile organisms are ubiquitous in nature across all scales, from bacterial colonies, to insect swarms and bird flocks. In particular, self-organization into vortex patterns is often observed and has been attributed to some local external attractor, e.g., light or nutrient concentration, together with behavioral rules like collision avoidance and mutual alignment. Pertinent social interactions are commonly thought to be based on perception and the ability to actively control the direction of motion. They are also generally presumed to provide some benefits to the individual and to the collective, as in the case of collision avoidance or predator evasion. However, since such interactions are usually derived only indirectly and approximately from observations, it is arguably useful to coarse grain them, e.g., into simple alignment rules. This strategy has been successful in physics in order to rationalize complex collective effects with the help of simple mechanistic models, in particular with respect to emerging universal traits. It is also supported by the observation that biological many-body assemblies often appear highly susceptible to environmental influences and exhibit a dynamical finite-size scaling reminiscent of critical states in inanimate physical systems. Importantly, the cascades of complex biochemical/biophysical processes needed to transform signal perception into a navigational reaction inevitably result in retarded interactions upon coarse-graining . This generic complication is often dismissed in the analysis, and dedicated models and experiments addressing the role of time delays in active matter are still rare, although these have occasionally been shown to fundamentally alter the collective dynamics and to bring it closer to that found in nature. To a first approximation, delay effects can resemble inertial corrections to an otherwise overdamped biological dynamics. In particular, both have a propensity to give rise to oscillations, and inertia moreover to rotational motion around an attractive center, as familiar from planetary orbits. Experiments that can assess or even deliberately control retarded interactions in living systems turn out to be difficult. But by imposing time delays onto synthetic active particles via computer control, we can create an ideal laboratory system to experimentally emulate such situations. Suitable feedback control techniques for active particles have recently become available through photon nudging . The technique allows to adjust a particle's propulsion speed to acquired real-time information (positions, directions of motion) about the dynamical state of an ensemble. It has previously been employed to rectify the rotational Brownian motion for particle steering and trapping, to explore orientation-density patterns in activity landscapes, and to study information flow between active particles and their emerging critical states. Beyond what related computer simulations accomplish, these experiments additionally incorporate the full real-world complexity arising from actual physical interactions due to hydrodynamic, thermal, or concentration fields. In the following we describe experiments with feedback-controlled active Brownian microswimmers "aiming" to reach a fixed target by a retarded thermophoretic self-propulsion. The resulting systematic navigational "errors" are seen to cause a spontaneous symmetry breaking to a bi-stable dynamical state, in which the swimmers self-organize into a merry-go-round motion that switches transiently between degenerate chiralities.
# Results {#results-1 .unnumbered}
#### Single Particle Retarded Interaction {#single-particle-retarded-interaction-1 .unnumbered}
The basic component of a swarm is a single active particle whose direction of motion depends dynamically on its environment. A perturbation of the particle position leads to an adjustment of the direction of motion, which can inevitably only take place with a delay between detection and reaction to the perturbation. In this work, we study the simplest case, where an active particle moves toward an immobilized target particle of the same size. Assuming that the active particle has a programmed time delay \(\delta t\) in its response, its propulsion direction \(\hat{\mathbf{u}}(t)\) at time \(t\) is determined by its relative position to the target particle it \"sensed\" at time \(t-\delta t\) in the past according to \[\label{eq:target_cohesion} \hat{\mathbf{u}}(t) = \frac{-\mathbf{r}(t-\delta t)}{|\mathbf{r}(t-\delta t)|},\] where \(\mathbf{r}\) is the location of the active particle with respect to the target particle at the origin. This interaction rule is implemented in an experimental feedback system that controls the propulsion of active particles. Here, the active particles are polymer particles with a radius of \(a=1.09\,\si{\micro\metre}\) coated with gold nanoparticles and suspended in a thin film of water. A laser with a wavelength of 532 nm is focused on the active particle with a displacement \(d\) from the center (Fig. [\[fig:figure1\]](#fig:figure1){reference-type="ref" reference="fig:figure1"}B) to drive the particle at a constant velocity \(v_0\) in the direction defined by Eq. [\[eq:target_cohesion\]](#eq:target_cohesion){reference-type="ref" reference="eq:target_cohesion"} via self-thermophoresis. A darkfield microscopy setup is used to image the particles (Fig. [\[fig:figure1\]](#fig:figure1){reference-type="ref" reference="fig:figure1"}C). A feedback loop running on a computer analyzes and records the positions of the particles to control the laser position accordingly via an acousto-optical deflector. We use a separate calibrator particle running on a quadratic trajectory as a reference for the velocity \(v_0\) of a free particle. Further details are described in the Methods section and the Supporting Materials. When the system is started with a programmed time delay \(\delta t=0\,{\rm s}\), the active particle moves towards the target particle until it collides with it. The motion of the active particle is then constrained by the target particle carrying out a diffusive motion around the target's circumference. Its distance to the target particle obeys the barometric distribution. As the delay increases, the amplitude of this diffusive motion grows until a critical delay is reached and the particle begins to rotate around the target (see Supplementary Movies 1--3). We characterize this dynamics with the help of the angle \(\theta\) between the direction of motion in Eq. [\[eq:target_cohesion\]](#eq:target_cohesion){reference-type="ref" reference="eq:target_cohesion"} and the negative radial direction \(-\mathbf{r}(t)\) (see Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A). Note that \(\sin(\theta)\) corresponds to the contribution of a single active particle \(i\) to the rotational order parameter commonly used to describe collective rotational motion, i.e., \(o_{R,i}=(\hat{{\bf r}}_i\times \hat {\bf u}_i)\cdot {\bf e}_z=\sin(\theta_i)\). Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A shows the experimental trajectories of this propulsion angle \(\theta\) for a single active particle with \(v_0=2.16\,\si{\micro\metre\per\second}\) and three different delays. For short delays, \(\theta\) fluctuates with a small amplitude around zero (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A top). The fluctuations increase with the delay and lead to a flat-top probability density of the propulsion angle for \(\delta t \approx 0.87\, {\rm s}\) (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A middle). At larger delays (\(\delta t=1.14\, {\rm s}\)), the propulsion angle fluctuates around a stable nonzero value which intermittently changes between a clockwise and counter-clockwise rotation sense of rotation (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A bottom). In this case, the probability density \(p(\theta)\) becomes bi-modal (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}C). The periods of the swimmer staying in either stably rotating state increase in duration when the delay is further increased. At \(\delta t=1.4\, \si{\second}\), the propulsion angle fluctuates around \(\pm 80 \si{\degree}\). Under these conditions, the cohesion of the particle to the target becomes weak as the particle velocity is almost tangential to the target particle circumference. As a result, the distance \(R\) of the particle to the target starts to fluctuate more strongly, as shown in the position histograms in Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}B. The non-zero propulsion angle is the result of an angular displacement during the period \([t-\delta t,t]\) of the active particle, and we can write it as \[\label{eq:deftheta} \theta(t) = \int^t_{t-\delta t} \omega(t') \,\mathrm{d} t' = \phi(t)-\phi(t-\delta t) = \angle (\hat{\mathbf{u}}(t),-\mathbf{r}(t)).\] Here, \(\phi(t)\) is the polar angle of the active particle in polar coordinates centered in the target particle and \(\omega(t) = \dot \phi(t)\) is the corresponding instantaneous angular velocity (Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}A). The observed dynamics can be understood by considering the active and target particles in physical contact. Their distance is then constrained to be the sum of their radii (\(R=2a\)). In this case, the active particle moves along the circumference of the target particle with an angular velocity \(\omega(t)\) according to \(\omega(t) =\omega_0\sin( \theta(t))\), where \(\omega_0= v_0/R\) and the angle \(\theta(t)\) is defined in Eq. [\[eq:deftheta\]](#eq:deftheta){reference-type="eqref" reference="eq:deftheta"}. The term \(\omega_0 \sin(\theta(t))\) originates from projecting the propulsion direction in Eq. [\[eq:target_cohesion\]](#eq:target_cohesion){reference-type="eqref" reference="eq:target_cohesion"} parallel to the surface of the target particle at the point of contact. As sketched in Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}A, assuming a constant angular velocity \(\omega\) with \(\theta = \omega\delta t\), the solutions to the equation for \(\theta\) are given by the intersections of a sine function and a linear function, \[\label{eq:simple} (\omega_0 \delta t)^{-1}\theta=\sin(\theta).\] For \(\omega_0\delta t <1\), there is a single intersection at \(\theta\) = 0, indicating a non-rotating state. For \(1<\omega_0\delta t<\pi/2\), one non-rotating and two rotating solutions with opposite rotation senses arise. These rotations are stable under perturbations (see SI), while the non-rotating solution is unstable. For \(\omega_0\delta t >\pi/2\), the rotating solutions correspond to \(|\theta|>\pi/2\), and the radial component of propulsion becomes positive (repulsive), driving the active particle away from the target particle. As a result, the radius of the orbit, \(R\), increases until a new stable orbit with \(R = 2 v_0 \delta t/\pi>2a\) and \(|\theta|=\pi/2\) is reached. For infinitely small particles (\(a\to 0\)), the distance of the swimmer to the target position can thus, in principle, vanish (\(R\to 0\)), and the rotation can occur at infinitely short programmed delays (\(\delta t\to 0\)). The retarded attraction hence always leads to rotation, and the observed transition from the non-rotating to the stable rotating state is caused by the minimum distance to the target. In the experiment, the minimum distance is determined not only by the particle radius \(a\) but also by the Brownian motion of the active particle within the instrumental delay of the feedback loop, and it is nonzero even for \(a>0\). Adding Brownian fluctuations to the deterministic motion of the active particle described by Eq. [\[eq:simple\]](#eq:simple){reference-type="ref" reference="eq:simple"}, results in the non-linear delayed stochastic differential equation \(\dot{\phi}(t) = \omega_0 \sin\left(\phi(t)-\phi(t-\tau)\right) + \sqrt{2D_0/R^2} \, \eta(t)\), where \(D_0 \approx 0.0642\) μm^ -1^ denotes the translational diffusion coefficient of the active particle and \(\eta(t)\) white noise. This equation is very difficult to solve. However, as detailed in the supplementary information, after several approximations it yields the simple overdamped Langevin equation \[\dot{\theta} =\frac{1}{3 \delta t}\left[\theta_{\pm}^2-\theta^2\right]\theta + \sqrt{2D_\theta}\eta \label{eq:domega_approx_overdamped}\] with \[\theta_{\pm} = \pm \sqrt{\frac{6}{\theta_0}(\theta_0-1)} \label{eq:roots}\] and \(\theta_0 \equiv \omega_0 \delta t\). The noise term in Eq. [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} describes the angular Brownian motion of the active particle around the target with an effective diffusion coefficient \(D_\theta\), which is a single free parameter in our theory. Eq. [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} yields the stationary solutions \(0\) and \(\theta_{\pm}\), consistent with our previous simplified model. All the three solutions coincide when \(\theta =\theta_0 = \omega_0\delta t = 1\), implying a transition from a non-rotating to two rotating states. The data points in Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B display the experimentally obtained maxima of the histograms \(p(\theta)\) of the propulsion angle (see Fig. [\[fig:figure2\]](#fig:figure2){reference-type="ref" reference="fig:figure2"}C) as a function of \(\omega_0\delta t\). While the transition to the two rotating states is expected at \(\omega_0\delta t\) = 1, the transition points in experiments are located at lower values due to an additional instrumental delay \(\Delta t\) in the feedback loop of the experimental setup. This instrumental delay between the most recent exposure to the camera and the laser positioning affects the motion direction, causing an earlier onset of the transition to a stable rotation. The dashed line in Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B shows the theoretical prediction, which includes the instrumental delay \(\Delta t\) and the programmed delay \(\delta t\), as detailed in the SI (Eq. (17)). The Langevin equation [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} can be interpreted as a dynamical equation for the position \(\theta\) of an overdamped Brownian particle with friction and diffusion coefficients \(\gamma\) and \(D_\theta\) in a quartic potential (see derivation in the SI), \[U(\theta)=\frac{\gamma}{\delta t}\left[ \left(\frac{1}{\theta_0}-1\right )\theta^2 +\frac{1}{12}\theta^4 \right], \label{eq:potential}\] which corresponds to the generic case of a supercritical pitchfork bifurcation. The potential can be extracted from the experimental data (Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B) by fitting the histogram \(p(\theta)\) with the Boltzmann distribution \(\exp(-U(\theta)/ \gamma D_\theta)/Z\) including an effective temperature \(\gamma D_\theta/k_{\rm B}\) and normalization factor \(Z\). The effective temperature thus represents a scaling factor that brings the measured potential of mean force \(-\gamma D_\theta \log p(\theta)\) to the theoretical prediction in Eq. [\[eq:potential\]](#eq:potential){reference-type="eqref" reference="eq:potential"}. The self-generated quartic potential in Eq. [\[eq:potential\]](#eq:potential){reference-type="eqref" reference="eq:potential"} emerging from the delayed response finds its mathematical analogue in the free energy function in Landau's theory of second-order phase transitions , with the control parameter \(\theta_0\). This correspondence suggests that \(\mathcal{T} = (\omega_0\delta t)^{-1}=\theta_0^{-1}\) is the equivalent of a temperature with a critical point \(\mathcal{T}_C\) = 1. Both the activity \(\omega_0\) and the delay \(\delta t\) in the product can lead to the transition. Hence, the active particle velocity and delay are inversely coupled, and large particle velocities need only short delays causing an ordered rotating state. Yet, our system does not correspond to a thermodynamic phase, where the transitions are of collective nature. The bifurcation and the potential are caused by the delayed response of the single active particle to an external signal including the steric repulsion by the target particle. A stability analysis of Eq. [\[eq:domega_approx_overdamped\]](#eq:domega_approx_overdamped){reference-type="eqref" reference="eq:domega_approx_overdamped"} (see SI) yields the susceptibility \[\label{eq:susceptibility} \chi\left (\frac{1}{\theta_0} \right) = \begin{cases} \frac{\delta t}{2\left (\frac{1}{\theta_0}-1\right )} & \text{if \(\frac{1}{\theta_0}>1\)}\\ -\frac{\delta t}{4\left (\frac{1}{\theta_0}-1\right )} & \text{if \(\frac{1}{\theta_0}<1\)} \end{cases}\;,\] as familiar from Landau theory (for \(\gamma = 1\) s). The susceptibility corresponds to the relaxation time of the system after a perturbation from a stable state (see SI). It is determined from the experiments by measuring the autocorrelation function \[\label{eq:autocorr} C(\tau)=\frac{\langle \delta \theta(t+\tau) \delta \theta(t)\rangle_t}{\langle\delta \theta(t)^2 \rangle_t}\] of fluctuations of the propulsion angle \(\delta \theta(t)=\theta(t)-\langle\theta(t)\rangle\). Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}C shows the decay time \(t_R\) obtained from the autocorrelation function as \(C(t_R)\) = 1/e (symbols) together with susceptibilities predicted theoretically without (Eq. [\[eq:susceptibility\]](#eq:susceptibility){reference-type="eqref" reference="eq:susceptibility"}, dashed line) and with (see SI, solid line) instrumental delay \(\Delta t\). Close to the transition point (\(\omega_0 \delta t \approx 1\)), the theory suggests a slowing down of the relaxation due to an increasingly flat potential (e.g., the potential plot in the middle of Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}B). The experimental results nicely confirm this prediction without any fitting parameter. In addition to calculating the susceptibility from the autocorrelation function, we also performed relaxation measurements after a step-like perturbation of the propulsion angle \(\theta\). The results of the measurements agree well with the theoretical predictions, as shown in Sec. 4 of the SI, together with details on the evaluation of the susceptibilities. While both stable rotation directions can be inferred from a purely deterministic model excluding Brownian motion, the observed spontaneous reversal of the rotation direction is driven by fluctuations in the propulsion angle and thus by the noise in the system. The changes in the rotation direction correspond to transitions between the minima \(\pm\theta_{\pm}\) of the self-generated potential, Eq. [\[eq:potential\]](#eq:potential){reference-type="eqref" reference="eq:potential"}. We may thus apply Kramers' theory to estimate the corresponding transition rate as \[k = \frac{\sqrt{2}}{\pi}\frac{|\theta_0-1|}{\theta_0 \delta t}\exp\left[-\frac{3}{\gamma D_\theta}\frac{(\theta_0-1)^2}{\theta_0^2}\right]. \label{eq:rate}\] The effective temperature \(\gamma D_\theta/k_{\rm B}\) driving the fluctuations in the potential is obtained from the previously mentioned scaling of the theoretical prediction to the measured potential of mean force. Fig. [\[fig:figure3\]](#fig:figure3){reference-type="ref" reference="fig:figure3"}D displays the results of experiments measuring the transition rate from mean residence times of \(\theta\) in the two potential wells. The rate is compared to the prediction from the Kramers theory, Eq. [\[eq:rate\]](#eq:rate){reference-type="eqref" reference="eq:rate"}, and shows a good agreement.
## Multiple Particles {#multiple-particles-1 .unnumbered}
As demonstrated in the previous section, the rotation observed in our experiments results from a spontaneous symmetry breaking in the dynamics of a single active agent under a retarded self-propulsion to a target, different from standard models of rotational dynamics in overdamped systems, which assume mutual or \"social\" interactions among the agents. Hence, when more active particles are added to the system, each of them strives to exhibit the same rotation and bifurcation as the single swimmer. However, steric, hydrodynamic, and thermophoretic interactions among the particles synchronize and stabilize their motion so that the system exhibits collective behavior. Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"} summarizes the key results obtained for an ensemble of 15 active particles attracted to the target particle with the same delay. For the considered range of time delays, the active particles form two tightly packed shells around the target particle (Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}A). The typical distance of the inner shell particles to the target is about half that of the outer shell, \(R^\mathrm{out} \approx 2R^\mathrm{in} = 4a\). So the inner shell is expected to enter the rotational phase with a correspondingly shorter delay than the outer shell. However, the interparticle interactions in the compact cluster strongly correlate the particle motion, which quantitatively changes the results obtained for a single particle. Compared to the theoretical prediction, \(\omega_0 \delta t\) = 0.73, we observe that for \(v_0\) = 2.06 \(\si{\micro\metre\per\second}\) the transition to the rotational phase of the inner shell is postponed to \(\omega^{\mathrm{in}}_0 \delta t \equiv v_0\delta t/ R^\mathrm{in}\approx\) 0.83, corresponding to \(\delta t\) = 0.9 s (see the rightmost red data point lying on the horizontal axis in Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}B). Close to the transition, the inner shell exhibits alternating periods of rotation and non-rotation, whereas the non-rotating outer shell compresses the inner shell due to its inwards pointing propulsion direction (Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C, left). Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C displays the velocity fields of the particles averaged over their trajectories with three different delays. The rotation of the outer shell is observed at \(\omega^{\mathrm{out}}_0\delta t \equiv v_0\delta t/ R^\mathrm{out} \approx\) 0.41, corresponding to the same delay \(\delta t\) = 0.9 s when the inner shell starts to rotate (see Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}B and Supplementary Movies 4--6). For delays slightly above the transition, \(0.9<\delta t\) \(<\) 1.41 s, the two shells rotate in opposite directions, as shown in the middle plot of Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C. The simultaneous transition and the counter-rotation of the two shells suggest that the inner shell particles generate forces in the opposite direction to their propulsion, repelling the outer shell particles and causing a bias on their rotation, as depicted in Figs. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}D--F. The bias is presumably caused by the directional hydrodynamic and thermophoretic interactions. The surface temperature gradient across the particle creates a thermo-osmotic surface flow that propels the particle. If the particle motion is (partly) halted by an external force such as the steric force applied by the immobilized target particle, the particle acts as a pump creating a hydrodynamic outflow at the hot side of the swimmer (Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}D and SI). Similarly, thermophoretic interactions arise from temperature gradients across the surface of an active particle caused by its neighboring particles. They are commonly repulsive as found, e.g., for Janus particles in external temperature gradients. We have carried out finite element simulations of the flow field around a mobile and an immobile self-propelling swimmer (see SI). The overall near field hydrodynamic interactions are nevertheless complex due to the presence of the other particles and the substrate surface, and further depend on the propulsion angle \(\theta\). The inner shell propulsion angle increases with the delay, which results in changing direction and magnitude of the inter-shell interaction and thus the outer shell rotation bias, which presumably varies as sketched in Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}F (see SI). For \(\delta t \geq 1.41\) s, the two shells predominantly rotate in the same direction, as shown in Fig. [\[fig:figure4\]](#fig:figure4){reference-type="ref" reference="fig:figure4"}C, right. The transition from counter-to co-rotation shells corresponds to the sign flip of the bias, occurring at \(\theta^\mathrm{in}\approx\) 67\(^\circ\). While we currently cannot separate thermophoretic and hydrodynamic effects in the experiment, possible hydrodynamic interactions yield an interesting perspective as they are supposed to influence also the collective behavior of fish or birds. At longer retardation, \(\theta^\mathrm{in}\) tends to reach 90\(^\circ\), and thus the inner shell tries to expand against the compressing outer shell. These competing tendencies lead to particle exchange between the two shells.
# Discussion {#discussion-1 .unnumbered}
We have demonstrated above that the motion of an active particle induced by the delayed attraction to a target point can spontaneously undergo a transition from a diffuse isotropic state to a dynamical chiral state, upon increasing the activity and/or the delay time. The transition is well described by a pitchfork bifurcation accompanied by a characteristic critical slowing down of the response. The single-particle dynamics thus already exhibits non-trivial features more commonly associated with (mean-field) phase transitions in strongly interacting passive many-body systems . This can be explained by noting that the deterministic part, \(\dot{\phi}(t) =\omega_0\sin\left (\phi(t)-\phi(t-\delta t)\right)\), of our stochastic delay differential equation can also be understood as the dynamical equation for a single Kuramoto phase oscillator, with vanishing eigenfrequency and coupling strength \(\omega_0\), which is trying to synchronize with its own past state. In the chiral state, the particle orbits around the central obstacle. This orbiting motion is stable against noise, but its chirality is only transiently maintained. This should be contrasted to the chiral states resulting from non-reciprocal coupling in the Kuramoto model as discussed by Vitelli et al., which require many-body cooperativity. Despite the similarity of the two models, we are clearly observing a single-particle effect, instead. The lack of many-body cooperativity is compensated for by the infinite number of relaxation modes encoded in the time-delayed equation of motion. The nonlinear dynamics of our experimental system can be described by an approximate analytical model resulting in an emergent self-generated quartic potential. While such potentials are frequently found in descriptions of phase transitions and collective effects in active particle ensembles, following various behavioral rules, this is not the case, here. Due to the activity and the (programmed) delay, it already occurs for a single active particle aiming at a spatially fixed target. In a whole swarm of particles that are all attracted to a common target, which might be its own perceived center of mass, the single-particle bifurcation is preserved. Inter-particle collisions merely synchronize and stabilize the rotational states of the individual particles. Upon close contact, hydrodynamic and thermophoretic interactions become important and help the swimmers to self-organize into co-and counter-rotating orbits. In biological motile ensembles, from bacteria to fish, similar hydrodynamic mechanisms may be at work, although precise details and scales may differ widely. The corresponding many-body effects can be subtle and may elude coarse grained simulations and theories. This underscores the importance of well controlled experimental model systems that may act as "hybrid simulations", combining computer-controlled active particles with real-world environments. While time delays are an unavoidable outcome of coarse-graining microscopic descriptions of the feedback processes in natural systems, they are often neglected in low-dimensional models of active particle collective effects. In this respect, our model system provides a new perspective, as it takes the unavoidable systematic delays in the dynamics seriously and explores their generic effects. It thereby provides a microscopic underpinning for empirical vision-cone models that have been found to produce rotational dynamics in overdamped flocks or swarms. In overdamped systems, retardation thus plays a similar role as added inertia. Both effects lead to persistence in the particle dynamics, but the effects of the time delay may be much richer . While we considered only a positive delay, i.e., synchronization with the past, sophisticated biological organisms also have predictive capabilities at their disposal that allow them to extrapolate the current state into the future. These can to a first approximation be incorporated in the form of a negative time delay. The inclusion of positive and negative delays may therefore also provide a new perspective on claims that collision and alignment interactions, as well as other complex rules, are the sources of the emerging complex adaptive responses observed in living many-body systems.
# Methods {#methods-1 .unnumbered}
## Sample Preparation
Samples were prepared using two glass coverslips (20 mm \(\times\) 20 mm, 24 mm \(\times\) 24 mm) to confine a thin liquid layer (3 \(\si{\micro\metre}\) thickness) in between. The edges of one coverslip are sealed with a thin layer of PDMS (polydimethylsiloxane) to prevent leakage and evaporation. The liquid film used in the sample is composed of 2.19 \(\si{\micro\metre}\) diameter gold-coated melamine formaldehyde (MF) particles (microParticles GmbH) dispersed in 0.1% Pluronic F-127 solution. The latter prevents the cohesion of the particles and adsorption to the cover slide surface. The surface of the MF particles is uniformly scattered with gold nano-particles of about 8 nm diameter with a total surface coverage of about 10% (Fig. S3A). SiO\(_2\) particles (2.96 \(\si{\micro\metre}\) in diameter, microParticles GmbH) are added into the solution to keep the thickness of the liquid layer at about 3 \(\si{\micro\metre}\). Finally, 0.3 \(\mathrm{\mu l}\) of the mixed particle suspension is pipetted on one of the coverslips, and the other is put on top.
## Experimental Setup
The experimental setup (see SI) consists of an inverted microscope (Olympus, IX71) with a mounted piezo translation stage (Physik Instrumente, P-733.3). The sample is illuminated with an oil-immersion darkfield condenser (Olympus, U-DCW, NA 1.2--1.4) and a white-light LED (Thorlabs, SOLIS-3C). The scattered light is imaged by an objective lens (Olympus, UPlanApo \(\times\)`<!-- -->`{=html}100/1.35, Oil, Iris, NA 0.5--1.35) and a tube lens (250 mm) to an EMCCD (electron-multiplying charge-coupled device) camera (Andor, iXon DV885LC). The variable numerical aperture of the objective was set to a value below the minimum aperture of the darkfield condenser. The microparticles are heated by a focused, continuous-wave laser at a wavelength of 532 nm (CNI, MGL-III-532). The beam diameter is increased by a beam expander and sent to an acousto-optic deflector (AA Opto-Electronic, DTSXY-400-532) and a lens system to steer the laser focus in the sample plane. The deflected beam is directed towards the sample by a dichroic beam splitter (D, Omega Optical, 560DRLP) and focused by an oil-immersion objective (Olympus, UPlanApo \(\times\)`<!-- -->`{=html}100/1.35, Oil, Iris, NA 0.5--1.35) to the sample plane (\(w_0 \approx 0.8\,\si{\micro\metre}\) beam waist in the sample plane). A notch filter (Thorlabs, NF533-17) is used to block any remaining back reflections of the laser from the detection path. The acousto-optic deflector (AOD), as well as the piezo stage, are driven by an AD/DA (analog-digital/digital-analog) converter (Jäger Messtechnik, ADwin-Gold II). A LabVIEW program running on a desktop PC (Intel Core i7 2600 4 \(\times\) 3.40 GHz CPU) is used to record and process the images as well as to control the AOD feedback via the AD/DA converter.
# Author Contributions {#author-contributions-1 .unnumbered}
XW and FC conceived the experiment. XW carried out the experiment. XW, FC, PC, VH analyzed the data. PC and VH developed the theory. XW, PC and FC carried out simulations. All authors discussed the results and wrote the manuscript. | {'timestamp': '2022-06-06T02:14:43', 'yymm': '2206', 'arxiv_id': '2206.01576', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01576'} |
# Introduction
For a topological space \(X\), we denote by \(C_p(X)\) the family \(C(X)\) of all continuous function from \(X\) to the real line \(\mathbb{R}\) endowed with the topology of pointwise convergence. A real-valued function \(f\) on a space \(X\) is a *Baire-one function* (or a *function of the first Baire class*) if \(f\) is a pointwise limit of a sequence of continuous functions on \(X\). Let \(B_1(X)\) denote the family of all Baire-one real-valued functions on a space \(X\) endowed with the topology of pointwise convergence. We say that a sequence \(\{f_n\}_{n\in\omega}\subseteq Y^X\) *stable (or discretely) converges* to a function \(f\in Y^X\) if for every \(x\in X\) the set \(\{n\in \omega: f_n(x)\neq f(x)\}\) is finite. Let \(B^{st}_1(X)\) be the family of all functions \(f:X\rightarrow \mathbb{R}\) which are limits of sequence \(\{f_n\}_{n\in\omega}\subseteq C_p(X)\) stably converging to \(f\). Functions in the family \(B_1^{st}(X)\) are called the *functions of stable first Baire class (or pointwise stabilizing functions of the first Baire class)*. A space is *meager* (or *of the first Baire category*) if it can be written as a countable union of closed sets with empty interior. A topological space \(X\) is *Baire* if the Baire Category Theorem holds for \(X\), i.e., the intersection of any sequence of open dense subsets of \(X\) is dense in \(X\). Clearly, if \(X\) is Baire, then \(X\) is not meager. Being a Baire space is an important topological property for a space and it is therefore natural to ask when function spaces are Baire. The Baire property for continuous mappings was first considered in . Then a paper appeared, where various aspects of this topic were considered. In , necessary and, in some cases, sufficient conditions on a space \(X\) were obtained under which the space \(C_p(X)\) is Baire. The problem about a characterization of Baireness for \(C_p(X)\) was solved independently by Pytkeev , Tkachuk and van Douwen .
A collection \(\mathcal{G}\) of subsets of \(X\) is *discrete* if each point of \(X\) has a neighborhood meeting at most one element of \(\mathcal{G}\), and is *strongly discrete* if for each \(G\in \mathcal{G}\) there is an open superset \(U_G\) of \(G\) such that \(\{U_G: G\in\mathcal{G}\}\) is discrete. A \(Coz_{\delta}\)-subset of \(X\) containing \(x\) is called a *\(Coz_{\delta}\) neighborhood* of \(x\).
In, we have obtained a characterization when \(B_1(X)\) is Baire.
In, T.Banakh and S.Gabriyelyan considers the following question (Problem 6.10 and Problem 9.11): *Is there a topological space \(X\) such that \(B_1(X)\) is Baire and \(B^{st}_1(X)\) is meager?* In this paper, we get the following theorem which answers the Banakh-Gabriyelyan question.
# Main definitions and notation
Throughout this paper, all spaces are assumed to be Tychonoff. The set of positive integers is denoted by \(\mathbb{N}\) and \(\omega=\mathbb{N}\cup \{0\}\). Let \(\mathbb{R}\) be the real line, we put \(\mathbb{I}=[0,1]\subset \mathbb{R}\), and let \(\mathbb{Q}\) be the rational numbers. Let \(f:X\rightarrow \mathbb{R}\) be a real-valued function, then \(\parallel f \parallel= \sup \{|f(x)|: x\in X\}\), \(S(g,\epsilon)=\{f: \parallel g-f \parallel<\epsilon\}\), \(B(g,\epsilon)=\{f: \parallel g-f \parallel\leq\epsilon\}\), where \(g\) is a real valued function and \(\epsilon>0\). Let \(V=\{f\in \mathbb{R}^X: f(x_i)\in V_i, i=1...,n\}\) where \(x_i\in X\), \(V_i\subseteq \mathbb{R}\) are bounded intervals for \(i=1...,n\), then \(supp V=\{x_1...,x_n\}\) , \(diam V=\max \{diam V_i: 1\leq i \leq n \}\). We recall that a subset of \(X\) that is the complete preimage of zero for a certain function from \(C(X)\) is called a *zero-set*. A subset \(O\subseteq X\) is called a cozero-set (or functionally open) of \(X\) if \(X\setminus O\) is a zero-set of \(X\). It is easy to check that zero sets are preserved by finite unions and countable intersections. Hence cozero sets are preserved by finite intersections and countable unions. Countable unions of zero sets will be denoted by \(Zer_{\sigma}\) (or \(Zer_{\sigma}(X)\)), countable intersection of cozero sets by \(Coz_{\delta}\) (or \(Coz_{\delta}(X)\)). It is easy to check that \(Zer_{\sigma}\)-sets are preserved by countable unions and finite intersections. Note that any zero set is \(Coz_{\delta}\) and any cozero-set is \(Zer_{\sigma}\). It is well known that \(f\) is of the first Baire class if and only if \(f^{-1}(U)\in Zer_{\sigma}\) for every open \(U\subseteq \mathbb{R}\) (see Exercise 3.A.1 in ).
# Proof of Theorem 1.4
Let \(\mathfrak{b}=\min\{|X|:\) \(X\) has a countable pseudocharacter but \(X\) is not a \(\lambda\)-space\(\}\) (see , p.149). By Theorem [\[th23\]](#th23){reference-type="ref" reference="th23"}, Proposition 3.15 and Example 4.8 in, we get the following results: \(\bullet\) If \(X\) is metrizable and \(|X|<\mathfrak{b}\) then \(B^{st}_1(X)\) is Baire. \(\bullet\) (CH) If \(X\) is a Luzin set then \(B^{st}_1(X)\) is meager. \(\bullet\) It is consistent with \(ZFC\) there is a zero-dimensional metrizable separable space \(X\) with \(|X|=\mathfrak{b}\) such that \(B_1^{st}(X)\) is Baire but \(B_1^{st}(X)\) is not Choquet (see the definition of a *Choquet* space in ). | {'timestamp': '2022-06-06T02:15:28', 'yymm': '2206', 'arxiv_id': '2206.01602', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01602'} |
# Introduction {#intro}
Kinetic plasma physics is described, typically, using various formulations or realizations of Boltzmann-like transport equations for the electron and ion species. A special case that is frequently studied is that of dynamics restricted to one spatial dimension and one momentum dimension ('1D-1V') for a single species of particles, given by \[\frac{\partial f}{\partial t} + v\frac{\partial f}{\partial x}-E \frac{\partial f}{\partial v} = \left[\frac{\delta f}{\delta t}\right]_{coll} \label{eq:vfp},\] where the electric field is \(E(x,t) = \int f ~dv-1\), and the particle distribution function is \(f = f(t,x,v)\). The left-hand side of this equation describes the evolution of the particle distribution in 'macroscopic' field \(E\) that arises self-consistently from the one-particle distribution \(f\). The right-hand side provides a description of two-particle and higher correlations (i.e. collisions). Equation [\[eq:vfp\]](#eq:vfp){reference-type="ref" reference="eq:vfp"}, along with Gauss's Law along with a particular form of the collision operator for two-particle interactions, is often termed the Vlasov-Poisson-Fokker-Planck (VPFP) equation set. Solving the VPFP set is often analytically intractable, even in 1D-1V. This is because the left-hand-side has a stiff linear transport term, has a non-linear term in \(E \partial f/\partial v\), and can sustain wave propagation and other hyperbolic partial-differential-equation (PDE) behavior. Additionally, the right hand side is typically represented by a hyperbolic, advection-diffusion, partial-differential-equation. Making progress on kinetic plasma physics requires computational simulation tools. Numerical solutions to the 1D-1V VPFP equation set have been applied in research on laser-plasma interactions in the context of inertial fusion, for example in plasma-based accelerators (), space physics (), fundamental plasma physics (), and inertial fusion (). Such numerical simulations may be used to explore initial conditions and forcing functions to understand the behavior of a physical effect in response to input parameters. Multi-dimensional 'brute-force' scans are, however, inefficient and costly and it is therefore beneficial to seek a more guided approach. Differentiable simulations have been used in a variety of contexts for such guided searches, for example learning parameters for molecular dynamics (), learning differencing stencils in PDEs (), and controlling PDEs (). In particular, there have been several recent applications of gradient-descent to fusion plasma physics. Analytic approaches have resulted in the development of adjoint methods for shape derivatives of functions that depend on magnetohydrodynamics (MHD) equilibria (). These methods have been used to perform optimization of stellarator design (). In other work, by using gradients obtained from analytic () and automatic differentiation (AD) (), the FOCUS and FOCUSADD codes optimize coil shape. These advances are founded on the concept of performing a sensitivity analysis towards device design. Here, we apply AD towards learning physical relationships and discovering novel phenomena in the VPFP dynamical system by training neural networks through differentiable simulations that solve Eq. [\[eq:vfp\]](#eq:vfp){reference-type="ref" reference="eq:vfp"}. In this work, we extend previous findings of the etching of nonlinear plasma wavepackets by localized damping at the rear of the packet in the context of inertial fusion to study the effect of the hot electrons from one wavepacket on upstream wavepackets using a gradient-based approach. We train a neural network that provides control parameters to the PDE solver. By choosing physical parameters as inputs and control parameters as outputs of the neural network, we enable the neural network to learn a function that describes the physical relationship between the plasma parameters and the forcing function parameters e.g. the resonance frequency. We train the neural network in an unsupervised fashion using a cost function based on the maximum entropy principle. This enables us to create self-learning plasma physics simulations, where the optimization process provides a physically interpretable function that can enable physics discovery.
# Physics Discovery using Differentiable Simulations
In this section, we provide a step-by-step description of how a traditional simulation-based computational physics workflow may be modified to perform closed-loop optimization.
## Open Loop: Manual Workflow
Figure [\[fig:basic_sim\]](#fig:basic_sim){reference-type="ref" reference="fig:basic_sim"} depicts a typical workflow of a computational scientist represented as a cyclic graph. The scientist defines the parametric inputs that create the state vector \(\mathbf{x}\). This can contain any parameters that are used to define the simulation e.g. the grid size, the number of solver steps, etc. For didactic purposes, the physical parameters to the simulation may be separated from \(\mathbf{x}\) into a different vector of inputs \(\mathbf{p_d}\) e.g. the forcing function parameters, the viscosity coefficient etc. Each of \(\mathbf{x}\) and \(\mathbf{p_d}\) is passed to the algorithm that solves the PDE which is represented by the function, \(\mathcal{V}\). The output of these simulations is stored in the final state vector \(\mathbf{x_f}\). The final state is postprocessed using a domain-specific set of algorithms devised by the scientist or otherwise. The results of the postprocessing are interpreted by the scientist who then determines the next set of inputs and parameters.
## Closed Loop: Brute Force Parameter Scan
Figure [\[fig:param_scan\]](#fig:param_scan){reference-type="ref" reference="fig:param_scan"} shows a more automated workflow. We replace the gray-box postprocessing step with the calculation of a scalar quantity \(S\) using a Cost Function \(\mathcal{C}\) on the final state \(\mathbf{x_f}\). This reduces the complexity of the interpretation of the postprocessing and enables a more rapid search in parameter space. The decrease in required human effort for completing one cycle enables the scientist to execute this loop as a brute force parameter scan over a pre-defined parameter space. At the end, the scientist can look up the minimum/maximum of the scalar cost function, and find the parameters which provide that minimum. The parameter scan approach scales with the number of different unique parameters and the number of values of each parameter. e.g. a 2-D search in \(x\) and \(y\) requires \(N_x \times N_y\) calculations. Therefore, the parameter scan approach quickly becomes inefficient when there are many parameters to scan, or when the required resolution in parameter space is very high. To search this parameter space efficiently, and to escape the linear scaling with each parameter, we can use gradient descent.
## Gradient-Descent-Based Parameter Learning {#sec:gbpl}
Figure [\[fig:param\]](#fig:param){reference-type="ref" reference="fig:param"} includes two modifications. The scientist/parameter search graybox has been replaced with a gradient-descent-based optimization algorithm. This algorithm provides the updated parameters, e.g. \(\omega_G\), a guess for the resonant frequency of the system, for the next iteration of the loop. The gradient-descent algorithm requires the calculation of an accurate gradient. Symbolic differentation is out of the question here as we do not have an analytical form for our system. In the appendix, we compare the performance of Finite Differencing to acquire the gradient and confirm that Automatic Differentiation is a superior method for this purpose. Therefore, by writing our PDE solver \(\mathcal{V}\) and the cost function \(\mathcal{C}\) using a numerical framework that supports Automatic Differentiation, we are able to perform gradient-descent. Since \[\mathcal{S} = \mathcal{C}(\mathbf{x_f}) = \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathbf{p_d})),\] the gradient for the update-step is given by \[\frac{\partial \mathcal{S}}{\partial \mathbf{p_d}} = \frac{\partial \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathbf{p_d}))}{\partial \mathbf{p_d}} = \frac{\partial \mathcal{S}}{\partial \mathcal{V}} \frac{\partial \mathcal{V}}{\partial \mathbf{p_d}}.\] For example, if we wish to learn the resonant frequency, \(\omega\), that optimizes for the scalar, \(\mathcal{S}\), we compute \[\begin{aligned}
\frac{\partial \mathcal{S}}{\partial \mathbf{\omega}} = \frac{\partial \mathcal{S}}{\partial \mathcal{V}} \frac{\partial \mathcal{V}}{\partial \mathbf{\omega}}. \label{eq:gbpl-grad} \end{aligned}\] Assuming a well-behaved solution manifold, performing gradient-descent tends to reduce the number of iterations required to find the minimum in comparison to a evenly-spaced parameter scan, especially when the required resolution is unknown (). Put another way, gradient-descent effectively provides an adaptive stepping mechanism in-lieu of the pre-defined values that represent a parameter scan.
## Gradient-Descent-Based Function Learning
In the final step, we can replace the lookup-like capability of the parameter optimization and choose to learn a blackbox function that can do the same. Through that process, we also gain the ability to interpolate and extrapolate with respect to the input space. Here, we choose to use neural networks, with a parameter vector \(\theta\), to represent the black box function. This allows us to extend the gradient-descent based methodology and leverage existing numerical software to implement this differentiable programming loop. Now, \[\mathcal{S} = \mathcal{C}(\mathbf{x_f}) = \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathbf{p_d})) = \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathcal{G}(\mathbf{x}, \mathbf{p_d}; \theta)),\] where \(\mathcal{G}\) is a function that generates the desired forcing function parameter given a parameter vector \(\mathbf{\theta}\). To extend the example from sec. [6.3](#sec:gbpl){reference-type="ref" reference="sec:gbpl"}, \(\omega\) is now a function given by \(\omega = \mathcal{G}(\mathbf{x}, \mathbf{p_d}; \theta)\). We compute the same gradient as in eq. [\[eq:gbpl-grad\]](#eq:gbpl-grad){reference-type="ref" reference="eq:gbpl-grad"} and add a correction factor that arises because the parameter (vector) is now \(\theta\), rather than \(\omega\). The necessary gradient for the gradient update is now given by \[\begin{aligned}
\frac{\partial \mathcal{S}}{\partial \mathbf{\theta}} = \left[\frac{\partial \mathcal{S}}{\partial \mathcal{V}} \frac{\partial \mathcal{V}}{\partial \mathcal{G}} \right] \frac{\partial \mathcal{G}(\mathbf{x}, \mathbf{p_d}; \mathbf{\theta})}{\partial \mathbf{\theta}}. \label{eq:nnloss} \end{aligned}\]
# Discovery of Long-lived Nonlinear Plasma Wavepackets
When electrostatic waves are driven to large amplitude, electrons can become trapped in the large potential (). Simulations of Stimulated Raman Scattering (SRS) in inertial confinement fusion (ICF) scenarios show that similar large-amplitude waves of finite extent are generated in the laser-plasma interaction, and that particle trapping is correlated with the transition to the high-reflectivity burst regime of SRS (). Simulating wavepackets, similar to those generated in SRS, but in isolation, has illuminated kinetic dynamics where trapped electrons, with velocity \(\approx v_{ph}\), transit the wavepacket which is moving at the slower group velocity \(v_g\). The transit of the trapped electrons from the back of the wavepacket to the front results in the resumption of Landau damping at the back and the wavepacket is then damped away (). Recent work () modeled the interaction of multiple speckles with a magnetic field acting as a control parameter. Since the effect of the magnetic field is to rotate the distribution in velocity space, the field strength serves as a parameter by which the authors control scattered particle propagation. Using this, along with carefully placed laser speckles, they show that scattered light and particles can serve as the trigger for SRS. Here, we ask *What happens when a non-linear electron plasma wavepacket is driven on top of another?* To answer this question, we reframe it as an optimization problem and ask, *What is the best way to excite a wavepacket that interacts with a pre-existing wavepacket?* We start with a large-amplitude, finite-length electrostatic wavepacket driven by a forcing function with parameters given by \[\begin{aligned}
\mathbf{p_0} &= \left[x_{0}, \omega_{0},t_{0},k_{0}\right], \end{aligned}\] where \(x_{i}\) is the location of excitation, \(\omega_i\) is the frequency, \(t_i\) is the time of excitation, and \(k_i\) is the wavenumber, of the \(i^\text{th}\) wavepacket. Since we seek to excite a second wavepacket that can interact with the detrapped electrons, we set the wavenumber \(k_1 = k_0\). We reparameterize the resonant frequency, \(\omega_1\), with a frequency shift, \(\Delta \omega_1\) and the linear resonant frequency \(\omega_0\) such that \(\omega_1 = \omega_0 + \Delta \omega_1\). We use the time of excitation of the second wavepacket, \(t_1\), as an independent variable along with \(k_0\). For each \(t_1\) and \(k_0\), we seek to learn functions that produce \(x_1\) and \(\Delta \omega_1\) i.e. we seek to learn \(x_1(t_1, k_0)\) and \(\Delta \omega_1( t_1, k_0)\). The entire parameter vector for the second wavepacket is given by \[\begin{aligned}
\mathbf{p_0} &= \left[ x_{1}(t_1, k_0), \Delta \omega_{1}(t_1, k_0),t_{1}, k_{0} \right]. \end{aligned}\] This framing is also illustrated in fig. [\[fig:wavepacketdiscovery\]](#fig:wavepacketdiscovery){reference-type="ref" reference="fig:wavepacketdiscovery"} where given \(k_0\) and \(t_1\), we seek functions for \(\omega_1\) and \(x_1\). We reparameterize \(\Delta \omega_1\) and \(x_1\) with a neural network with a parameter vector, \(\theta^*\), that maximizes the electrostatic energy (minimizes the free energy) and maximizes the kinetic entropy i.e. \[\begin{aligned}
x_1 &= x_1(t_1,k_0; \theta^*), \\ \Delta \omega_1 &= \Delta \omega_1(t_1, k_0; \theta^*). \end{aligned}\] where, \[\begin{aligned}
\theta^* &= argmin ~~\left[U_\text{es}(\textbf{p}; \theta)-\Delta \mathcal{KE}(\textbf{p}; \theta)\right], \end{aligned}\] and \[\begin{aligned}
U_\text{es} &= \sum_{t_{i}}^{t_f} \Delta t \sum_x \Delta x ~ E^2 \label{eq:esloss}\\ \Delta \mathcal{KE} &= \sum_{t_{i}}^{t_f} \Delta t \sum_x \Delta x \sum_v \Delta v ~ (f \log(f)-f_\text{MX} \log(f_\text{MX})) \label{eq:keloss}, \end{aligned}\] are the electrostatic energy, and entropy, terms in the loss function, respectively. \(f_\text{MX} = f_\text{MX}(n, T)\), \(n = \int f dv\), \(T = \int f v^2 dv /n\), where \(f_\text{MX}\) is the Maxwell-Boltzmann distribution. We vary the independent variables such that \[\begin{aligned}
k_0 &\in [0.26, 0.27,..., 0.32], \\ t_1 &\in [400, 500,..., 800], \end{aligned}\] giving an input space of 35 samples from which we seek to learn these functions. We use a neural network with 2 hidden linear layers with 8 nodes activated with a ReLU function. The final layer is activated with a \(\tanh\) function. The output is normalized such that \(p_1 = p_{norm} \times p_{out} + p_{shift}\) where \(p_{out}\) is the output from the \(\tanh\) function. We normalize the inputs between 0 and 1, and outputs with reasonable windows. For \(x_1\), we allow the entire domain, and \(\Delta \omega_1 \in [-0.06, 0.06]\). We use the ADAM optimizer with a learning rate of 0.05. The training simulations are performed with \(N_x = 6656, N_v = 160, N_t = 1200, t_{max} = 1100 \omega_p^{-1}, x_{max} = 6600 \lambda_D\). The amplitude of both drivers is \(0.05\). The spatial width is \(400 \lambda_D\) and temporal width is \(50 \omega_p^{-1}\). We use a sixth order integrator in time () with operator splitting. Both operators are computed with exponential integrators with spectral discretizations in phase space (). We use a spectral solver for Gauss's Law. In order to more closely match realistic plasma conditions, we also implement a Fokker-Planck collision operator and use a non-zero collision frequency \(\nu_{ee} = 10^{-4}\) for these simulations. More information on the solver is provided in the Appendix. We implement absorbing boundaries by increasing the collision frequency of the Krook operator in a small localized region at the boundary (). Figure [\[fig:loss\]](#fig:loss){reference-type="ref" reference="fig:loss"} shows that the loss value is reduced over the duration of the training process, roughly 30 hours, which runs for 60 epochs. The convergence in the loss metric suggests that we were able to train a overparameterized neural network with 35 samples of data in 60 epochs. We attribute this to the effect of having so-called 'physical' gradients from training through a PDE solver (). Figure [\[fig:llived\]](#fig:llived){reference-type="ref" reference="fig:llived"} shows the electric field profile for three different simulations. In fig. [\[fig:fld1\]](#fig:fld1){reference-type="ref" reference="fig:fld1"}, only the first wavepacket is excited, and in fig. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"}, only the second wavepacket is excited. In fig. [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}, both wavepackets are excited. Early in time, \(t=400 \omega_p^{-1}\) (green), when only the first wavepacket has been excited figs. [\[fig:fld1\]](#fig:fld1){reference-type="ref" reference="fig:fld1"} and [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"} agree perfectly. The second wavepacket is excited at \(t=500 \omega_p^{-1}\). At \(t=900 \omega_p^{-1}\), once some time has passed after the excitation of the second wavepacket, the first wavepacket has not fully damped away. It is visible as small bumps in figs. [\[fig:fld1\]](#fig:fld1){reference-type="ref" reference="fig:fld1"} and [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}. The second wavepacket is also present at this time and easily seen in fig. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"}. A larger amplitude wavepacket is seen in fig. [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}. Late in time, the difference in amplitude between the second wavepacket in figs. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"} and [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"} is obvious. The second wavepacket has nearly damped away in fig. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"}. In fig. [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}, the second wavepacket continues to persist, at nearly the same energy as it was at \(t=900 \omega_p^{-1}\). We observe this superadditive behavior, where \(f(x) + f(y) \leq f(x+y)\), for all wavenumbers we model. To determine the mechanism behind this phenomenon, we turn to the phase-space dynamics. In fig. [\[fig:f\]](#fig:f){reference-type="ref" reference="fig:f"}, (i) is a space-time plot of the electric field. The two dashed-dot red lines at the front and back of the wavepacket are parallel and indicate the velocity of the wavefront. In fig. [\[fig:etch-f\]](#fig:etch-f){reference-type="ref" reference="fig:etch-f"}(i), the front of the wavepacket propagates at a seemingly faster rate than the rear. This is due to the etching effect (). In fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"}, the wave survives for a much longer time, as was also illustrated in fig. [\[fig:llived\]](#fig:llived){reference-type="ref" reference="fig:llived"}. Figure [\[fig:f\]](#fig:f){reference-type="ref" reference="fig:f"} (ii) and (iii) are phase-space plots with their center indicated by the intersection of the horizontal timestamp line, and the dashed-dot line at the rear wavepacket red line at the back of the wavepacket. (iv) and (v) correspond to the intersection with the dashed-dot line at the front. (ii) and (iv) show the phase space within a window in x, while (iii) and (v) are the spatially averaged distribution function. (iii) and (v) serve as a proxy for approximating the propensity of Landau damping in that region. In fig. [\[fig:etch-f\]](#fig:etch-f){reference-type="ref" reference="fig:etch-f"}(ii) and (iii), we see that the rear of the wavepacket is Maxwellian. As previously shown, this is why the rear of the wavepacket damps faster than the front as in fig. [\[fig:etch-f\]](#fig:etch-f){reference-type="ref" reference="fig:etch-f"}(i) (). In the simulations described here, fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"} show that the distribution function at the back of the wavepacket has trapped particle activity (fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"}(ii)) and near zero slope at the phase velocity of the wave (fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"}(iii)). Both plots show that the slope is negligible because of the arrival of streaming detrapped particles from the first wavepacket. Due to this effect, the remergence of Landau damping that occurs due to the loss of trapped particles in isolated wavepackets no longer occurs here. This results in a reduction of the etching and the wavepacket propagates freely for some time while the particles from the first wavepacket propagate and arrive at the rear of the second wavepacket. Figure [\[fig:spatial\]](#fig:spatial){reference-type="ref" reference="fig:spatial"} shows the results of the optimization process for the resonant spatial location. For \(t_1 < 500 \omega_p^{-1}\), the resonant location decreases as a function of wavenumber. From analyzing phase space, we have determined that the spatial location is related to the resonant electron transport i.e. \(x_1 \approx v_{ph} t_1\). Waves with larger wavenumbers have smaller phase velocities. Because of this, the resonant spatial location decreases as a function of wavenumber. For a fixed wavenumber, fig. [\[fig:locus\]](#fig:locus){reference-type="ref" reference="fig:locus"} shows a locus of points in space-time where long-lived wavepackets can be excited in the presence of a pre-existing wavepacket. This suggests the possibility of a critical space-time radius within which collisional relaxation has yet to occur, and non-linear effects can be exploited. Likewise, we also learn the dependence of the optimum frequency shift as a function of \(k_0\) and \(t_c\), as shown in fig. [\[fig:dw-surf\]](#fig:dw-surf){reference-type="ref" reference="fig:dw-surf"}. The learned frequency shift, a few percent here, is similar in magnitude as to that observed in previous work related to SRS (). Furthermore, the frequency shift increases in magnitude as a function of wavenumber. As before, waves with larger wavenumbers have smaller phase velocities, and therefore, interact with more particles. Because of this, waves with larger wavenumbers have a larger non-linear frequency shift associated with them, as we see here ().
# Conclusion
We show how one may be able to discover novel physics using differentiable simulations by posing a physical question as an optimization problem. This required domain expertise in determining which functional dependencies to learn using neural networks. In fig. [\[fig:workflows\]](#fig:workflows){reference-type="ref" reference="fig:workflows"}, we show how one may adapt an existing computational science workflow to the autodidactic process described here. In the work performed here, this process enabled the discovery of parameters in a 4D search space with known bounds but an unknown resolution requirement. We trained the model over a coarse grid in \(k_0\) and \(t_1\), and learned functions for \(x_1\) and \(\omega_1\). Using gradient descent here allows an escape from the curse of dimensionality and reduces the problem from a 4D search to a 2D search + 2D gradient descent. This discovery process is not limited to differentiable simulations. While in fig. [\[fig:workflows\]](#fig:workflows){reference-type="ref" reference="fig:workflows"}, \(\mathcal{V}\) represents a PDE solve, it only needs to be a AD-enabled function that is a model for a physical system. For example, rather than a PDE solve, \(\mathcal{V}\) could represent a pre-trained neural-network-based emulator for experimental data. In such a scenario, one may be able to learn forcing function parameters for an experiment using the proposed workflow. Finally, in neural network literature, the gradient required for the update is \(\partial \mathcal{S} / \partial \theta = \partial \mathcal{S} / \partial \mathcal{G} \times \partial \mathcal{G} / \partial \theta\). We see that this is the same as eq. [\[eq:nnloss\]](#eq:nnloss){reference-type="ref" reference="eq:nnloss"} after the addition of one more node in the computational graph for \(\mathcal{V}\), the function that models the physical system. This allows the neural network training process to become unsupervised and data-efficient. In future work, we aim to dive deeper into the discovered phenomenon to better understand the mechanism and derive some scalings. Reduced models of wavepacket dynamics in SRS remain useful for the development of inertial confinement fusion schemes where laser-plasma instabilities occur.
# Introduction {#intro}
Kinetic plasma physics is described, typically, using various formulations or realizations of Boltzmann-like transport equations for the electron and ion species. A special case that is frequently studied is that of dynamics restricted to one spatial dimension and one momentum dimension ('1D-1V') for a single species of particles, given by \[\frac{\partial f}{\partial t} + v\frac{\partial f}{\partial x}-E \frac{\partial f}{\partial v} = \left[\frac{\delta f}{\delta t}\right]_{coll} \label{eq:vfp},\] where the electric field is \(E(x,t) = \int f ~dv-1\), and the particle distribution function is \(f = f(t,x,v)\). The left-hand side of this equation describes the evolution of the particle distribution in 'macroscopic' field \(E\) that arises self-consistently from the one-particle distribution \(f\). The right-hand side provides a description of two-particle and higher correlations (i.e. collisions). Equation [\[eq:vfp\]](#eq:vfp){reference-type="ref" reference="eq:vfp"}, along with Gauss's Law along with a particular form of the collision operator for two-particle interactions, is often termed the Vlasov-Poisson-Fokker-Planck (VPFP) equation set. Solving the VPFP set is often analytically intractable, even in 1D-1V. This is because the left-hand-side has a stiff linear transport term, has a non-linear term in \(E \partial f/\partial v\), and can sustain wave propagation and other hyperbolic partial-differential-equation (PDE) behavior. Additionally, the right hand side is typically represented by a hyperbolic, advection-diffusion, partial-differential-equation. Making progress on kinetic plasma physics requires computational simulation tools. Numerical solutions to the 1D-1V VPFP equation set have been applied in research on laser-plasma interactions in the context of inertial fusion, for example in plasma-based accelerators (), space physics (), fundamental plasma physics (), and inertial fusion (). Such numerical simulations may be used to explore initial conditions and forcing functions to understand the behavior of a physical effect in response to input parameters. Multi-dimensional 'brute-force' scans are, however, inefficient and costly and it is therefore beneficial to seek a more guided approach. Differentiable simulations have been used in a variety of contexts for such guided searches, for example learning parameters for molecular dynamics (), learning differencing stencils in PDEs (), and controlling PDEs (). In particular, there have been several recent applications of gradient-descent to fusion plasma physics. Analytic approaches have resulted in the development of adjoint methods for shape derivatives of functions that depend on magnetohydrodynamics (MHD) equilibria (). These methods have been used to perform optimization of stellarator design (). In other work, by using gradients obtained from analytic () and automatic differentiation (AD) (), the FOCUS and FOCUSADD codes optimize coil shape. These advances are founded on the concept of performing a sensitivity analysis towards device design. Here, we apply AD towards learning physical relationships and discovering novel phenomena in the VPFP dynamical system by training neural networks through differentiable simulations that solve Eq. [\[eq:vfp\]](#eq:vfp){reference-type="ref" reference="eq:vfp"}. In this work, we extend previous findings of the etching of nonlinear plasma wavepackets by localized damping at the rear of the packet in the context of inertial fusion to study the effect of the hot electrons from one wavepacket on upstream wavepackets using a gradient-based approach. We train a neural network that provides control parameters to the PDE solver. By choosing physical parameters as inputs and control parameters as outputs of the neural network, we enable the neural network to learn a function that describes the physical relationship between the plasma parameters and the forcing function parameters e.g. the resonance frequency. We train the neural network in an unsupervised fashion using a cost function based on the maximum entropy principle. This enables us to create self-learning plasma physics simulations, where the optimization process provides a physically interpretable function that can enable physics discovery.
# Physics Discovery using Differentiable Simulations
In this section, we provide a step-by-step description of how a traditional simulation-based computational physics workflow may be modified to perform closed-loop optimization.
## Open Loop: Manual Workflow
Figure [\[fig:basic_sim\]](#fig:basic_sim){reference-type="ref" reference="fig:basic_sim"} depicts a typical workflow of a computational scientist represented as a cyclic graph. The scientist defines the parametric inputs that create the state vector \(\mathbf{x}\). This can contain any parameters that are used to define the simulation e.g. the grid size, the number of solver steps, etc. For didactic purposes, the physical parameters to the simulation may be separated from \(\mathbf{x}\) into a different vector of inputs \(\mathbf{p_d}\) e.g. the forcing function parameters, the viscosity coefficient etc. Each of \(\mathbf{x}\) and \(\mathbf{p_d}\) is passed to the algorithm that solves the PDE which is represented by the function, \(\mathcal{V}\). The output of these simulations is stored in the final state vector \(\mathbf{x_f}\). The final state is postprocessed using a domain-specific set of algorithms devised by the scientist or otherwise. The results of the postprocessing are interpreted by the scientist who then determines the next set of inputs and parameters.
## Closed Loop: Brute Force Parameter Scan
Figure [\[fig:param_scan\]](#fig:param_scan){reference-type="ref" reference="fig:param_scan"} shows a more automated workflow. We replace the gray-box postprocessing step with the calculation of a scalar quantity \(S\) using a Cost Function \(\mathcal{C}\) on the final state \(\mathbf{x_f}\). This reduces the complexity of the interpretation of the postprocessing and enables a more rapid search in parameter space. The decrease in required human effort for completing one cycle enables the scientist to execute this loop as a brute force parameter scan over a pre-defined parameter space. At the end, the scientist can look up the minimum/maximum of the scalar cost function, and find the parameters which provide that minimum. The parameter scan approach scales with the number of different unique parameters and the number of values of each parameter. e.g. a 2-D search in \(x\) and \(y\) requires \(N_x \times N_y\) calculations. Therefore, the parameter scan approach quickly becomes inefficient when there are many parameters to scan, or when the required resolution in parameter space is very high. To search this parameter space efficiently, and to escape the linear scaling with each parameter, we can use gradient descent.
## Gradient-Descent-Based Parameter Learning {#sec:gbpl}
Figure [\[fig:param\]](#fig:param){reference-type="ref" reference="fig:param"} includes two modifications. The scientist/parameter search graybox has been replaced with a gradient-descent-based optimization algorithm. This algorithm provides the updated parameters, e.g. \(\omega_G\), a guess for the resonant frequency of the system, for the next iteration of the loop. The gradient-descent algorithm requires the calculation of an accurate gradient. Symbolic differentation is out of the question here as we do not have an analytical form for our system. In the appendix, we compare the performance of Finite Differencing to acquire the gradient and confirm that Automatic Differentiation is a superior method for this purpose. Therefore, by writing our PDE solver \(\mathcal{V}\) and the cost function \(\mathcal{C}\) using a numerical framework that supports Automatic Differentiation, we are able to perform gradient-descent. Since \[\mathcal{S} = \mathcal{C}(\mathbf{x_f}) = \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathbf{p_d})),\] the gradient for the update-step is given by \[\frac{\partial \mathcal{S}}{\partial \mathbf{p_d}} = \frac{\partial \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathbf{p_d}))}{\partial \mathbf{p_d}} = \frac{\partial \mathcal{S}}{\partial \mathcal{V}} \frac{\partial \mathcal{V}}{\partial \mathbf{p_d}}.\] For example, if we wish to learn the resonant frequency, \(\omega\), that optimizes for the scalar, \(\mathcal{S}\), we compute \[\begin{aligned}
\frac{\partial \mathcal{S}}{\partial \mathbf{\omega}} = \frac{\partial \mathcal{S}}{\partial \mathcal{V}} \frac{\partial \mathcal{V}}{\partial \mathbf{\omega}}. \label{eq:gbpl-grad} \end{aligned}\] Assuming a well-behaved solution manifold, performing gradient-descent tends to reduce the number of iterations required to find the minimum in comparison to a evenly-spaced parameter scan, especially when the required resolution is unknown (). Put another way, gradient-descent effectively provides an adaptive stepping mechanism in-lieu of the pre-defined values that represent a parameter scan.
## Gradient-Descent-Based Function Learning
In the final step, we can replace the lookup-like capability of the parameter optimization and choose to learn a blackbox function that can do the same. Through that process, we also gain the ability to interpolate and extrapolate with respect to the input space. Here, we choose to use neural networks, with a parameter vector \(\theta\), to represent the black box function. This allows us to extend the gradient-descent based methodology and leverage existing numerical software to implement this differentiable programming loop. Now, \[\mathcal{S} = \mathcal{C}(\mathbf{x_f}) = \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathbf{p_d})) = \mathcal{C}(\mathcal{V}(\mathbf{x}, \mathcal{G}(\mathbf{x}, \mathbf{p_d}; \theta)),\] where \(\mathcal{G}\) is a function that generates the desired forcing function parameter given a parameter vector \(\mathbf{\theta}\). To extend the example from sec. [6.3](#sec:gbpl){reference-type="ref" reference="sec:gbpl"}, \(\omega\) is now a function given by \(\omega = \mathcal{G}(\mathbf{x}, \mathbf{p_d}; \theta)\). We compute the same gradient as in eq. [\[eq:gbpl-grad\]](#eq:gbpl-grad){reference-type="ref" reference="eq:gbpl-grad"} and add a correction factor that arises because the parameter (vector) is now \(\theta\), rather than \(\omega\). The necessary gradient for the gradient update is now given by \[\begin{aligned}
\frac{\partial \mathcal{S}}{\partial \mathbf{\theta}} = \left[\frac{\partial \mathcal{S}}{\partial \mathcal{V}} \frac{\partial \mathcal{V}}{\partial \mathcal{G}} \right] \frac{\partial \mathcal{G}(\mathbf{x}, \mathbf{p_d}; \mathbf{\theta})}{\partial \mathbf{\theta}}. \label{eq:nnloss} \end{aligned}\]
# Discovery of Long-lived Nonlinear Plasma Wavepackets
When electrostatic waves are driven to large amplitude, electrons can become trapped in the large potential (). Simulations of Stimulated Raman Scattering (SRS) in inertial confinement fusion (ICF) scenarios show that similar large-amplitude waves of finite extent are generated in the laser-plasma interaction, and that particle trapping is correlated with the transition to the high-reflectivity burst regime of SRS (). Simulating wavepackets, similar to those generated in SRS, but in isolation, has illuminated kinetic dynamics where trapped electrons, with velocity \(\approx v_{ph}\), transit the wavepacket which is moving at the slower group velocity \(v_g\). The transit of the trapped electrons from the back of the wavepacket to the front results in the resumption of Landau damping at the back and the wavepacket is then damped away (). Recent work () modeled the interaction of multiple speckles with a magnetic field acting as a control parameter. Since the effect of the magnetic field is to rotate the distribution in velocity space, the field strength serves as a parameter by which the authors control scattered particle propagation. Using this, along with carefully placed laser speckles, they show that scattered light and particles can serve as the trigger for SRS. Here, we ask *What happens when a non-linear electron plasma wavepacket is driven on top of another?* To answer this question, we reframe it as an optimization problem and ask, *What is the best way to excite a wavepacket that interacts with a pre-existing wavepacket?* We start with a large-amplitude, finite-length electrostatic wavepacket driven by a forcing function with parameters given by \[\begin{aligned}
\mathbf{p_0} &= \left[x_{0}, \omega_{0},t_{0},k_{0}\right], \end{aligned}\] where \(x_{i}\) is the location of excitation, \(\omega_i\) is the frequency, \(t_i\) is the time of excitation, and \(k_i\) is the wavenumber, of the \(i^\text{th}\) wavepacket. Since we seek to excite a second wavepacket that can interact with the detrapped electrons, we set the wavenumber \(k_1 = k_0\). We reparameterize the resonant frequency, \(\omega_1\), with a frequency shift, \(\Delta \omega_1\) and the linear resonant frequency \(\omega_0\) such that \(\omega_1 = \omega_0 + \Delta \omega_1\). We use the time of excitation of the second wavepacket, \(t_1\), as an independent variable along with \(k_0\). For each \(t_1\) and \(k_0\), we seek to learn functions that produce \(x_1\) and \(\Delta \omega_1\) i.e. we seek to learn \(x_1(t_1, k_0)\) and \(\Delta \omega_1( t_1, k_0)\). The entire parameter vector for the second wavepacket is given by \[\begin{aligned}
\mathbf{p_0} &= \left[ x_{1}(t_1, k_0), \Delta \omega_{1}(t_1, k_0),t_{1}, k_{0} \right]. \end{aligned}\] This framing is also illustrated in fig. [\[fig:wavepacketdiscovery\]](#fig:wavepacketdiscovery){reference-type="ref" reference="fig:wavepacketdiscovery"} where given \(k_0\) and \(t_1\), we seek functions for \(\omega_1\) and \(x_1\). We reparameterize \(\Delta \omega_1\) and \(x_1\) with a neural network with a parameter vector, \(\theta^*\), that maximizes the electrostatic energy (minimizes the free energy) and maximizes the kinetic entropy i.e. \[\begin{aligned}
x_1 &= x_1(t_1,k_0; \theta^*), \\ \Delta \omega_1 &= \Delta \omega_1(t_1, k_0; \theta^*). \end{aligned}\] where, \[\begin{aligned}
\theta^* &= argmin ~~\left[U_\text{es}(\textbf{p}; \theta)-\Delta \mathcal{KE}(\textbf{p}; \theta)\right], \end{aligned}\] and \[\begin{aligned}
U_\text{es} &= \sum_{t_{i}}^{t_f} \Delta t \sum_x \Delta x ~ E^2 \label{eq:esloss}\\ \Delta \mathcal{KE} &= \sum_{t_{i}}^{t_f} \Delta t \sum_x \Delta x \sum_v \Delta v ~ (f \log(f)-f_\text{MX} \log(f_\text{MX})) \label{eq:keloss}, \end{aligned}\] are the electrostatic energy, and entropy, terms in the loss function, respectively. \(f_\text{MX} = f_\text{MX}(n, T)\), \(n = \int f dv\), \(T = \int f v^2 dv /n\), where \(f_\text{MX}\) is the Maxwell-Boltzmann distribution. We vary the independent variables such that \[\begin{aligned}
k_0 &\in [0.26, 0.27,..., 0.32], \\ t_1 &\in [400, 500,..., 800], \end{aligned}\] giving an input space of 35 samples from which we seek to learn these functions. We use a neural network with 2 hidden linear layers with 8 nodes activated with a ReLU function. The final layer is activated with a \(\tanh\) function. The output is normalized such that \(p_1 = p_{norm} \times p_{out} + p_{shift}\) where \(p_{out}\) is the output from the \(\tanh\) function. We normalize the inputs between 0 and 1, and outputs with reasonable windows. For \(x_1\), we allow the entire domain, and \(\Delta \omega_1 \in [-0.06, 0.06]\). We use the ADAM optimizer with a learning rate of 0.05. The training simulations are performed with \(N_x = 6656, N_v = 160, N_t = 1200, t_{max} = 1100 \omega_p^{-1}, x_{max} = 6600 \lambda_D\). The amplitude of both drivers is \(0.05\). The spatial width is \(400 \lambda_D\) and temporal width is \(50 \omega_p^{-1}\). We use a sixth order integrator in time () with operator splitting. Both operators are computed with exponential integrators with spectral discretizations in phase space (). We use a spectral solver for Gauss's Law. In order to more closely match realistic plasma conditions, we also implement a Fokker-Planck collision operator and use a non-zero collision frequency \(\nu_{ee} = 10^{-4}\) for these simulations. More information on the solver is provided in the Appendix. We implement absorbing boundaries by increasing the collision frequency of the Krook operator in a small localized region at the boundary (). Figure [\[fig:loss\]](#fig:loss){reference-type="ref" reference="fig:loss"} shows that the loss value is reduced over the duration of the training process, roughly 30 hours, which runs for 60 epochs. The convergence in the loss metric suggests that we were able to train a overparameterized neural network with 35 samples of data in 60 epochs. We attribute this to the effect of having so-called 'physical' gradients from training through a PDE solver (). Figure [\[fig:llived\]](#fig:llived){reference-type="ref" reference="fig:llived"} shows the electric field profile for three different simulations. In fig. [\[fig:fld1\]](#fig:fld1){reference-type="ref" reference="fig:fld1"}, only the first wavepacket is excited, and in fig. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"}, only the second wavepacket is excited. In fig. [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}, both wavepackets are excited. Early in time, \(t=400 \omega_p^{-1}\) (green), when only the first wavepacket has been excited figs. [\[fig:fld1\]](#fig:fld1){reference-type="ref" reference="fig:fld1"} and [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"} agree perfectly. The second wavepacket is excited at \(t=500 \omega_p^{-1}\). At \(t=900 \omega_p^{-1}\), once some time has passed after the excitation of the second wavepacket, the first wavepacket has not fully damped away. It is visible as small bumps in figs. [\[fig:fld1\]](#fig:fld1){reference-type="ref" reference="fig:fld1"} and [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}. The second wavepacket is also present at this time and easily seen in fig. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"}. A larger amplitude wavepacket is seen in fig. [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}. Late in time, the difference in amplitude between the second wavepacket in figs. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"} and [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"} is obvious. The second wavepacket has nearly damped away in fig. [\[fig:fld2\]](#fig:fld2){reference-type="ref" reference="fig:fld2"}. In fig. [\[fig:fld3\]](#fig:fld3){reference-type="ref" reference="fig:fld3"}, the second wavepacket continues to persist, at nearly the same energy as it was at \(t=900 \omega_p^{-1}\). We observe this superadditive behavior, where \(f(x) + f(y) \leq f(x+y)\), for all wavenumbers we model. To determine the mechanism behind this phenomenon, we turn to the phase-space dynamics. In fig. [\[fig:f\]](#fig:f){reference-type="ref" reference="fig:f"}, (i) is a space-time plot of the electric field. The two dashed-dot red lines at the front and back of the wavepacket are parallel and indicate the velocity of the wavefront. In fig. [\[fig:etch-f\]](#fig:etch-f){reference-type="ref" reference="fig:etch-f"}(i), the front of the wavepacket propagates at a seemingly faster rate than the rear. This is due to the etching effect (). In fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"}, the wave survives for a much longer time, as was also illustrated in fig. [\[fig:llived\]](#fig:llived){reference-type="ref" reference="fig:llived"}. Figure [\[fig:f\]](#fig:f){reference-type="ref" reference="fig:f"} (ii) and (iii) are phase-space plots with their center indicated by the intersection of the horizontal timestamp line, and the dashed-dot line at the rear wavepacket red line at the back of the wavepacket. (iv) and (v) correspond to the intersection with the dashed-dot line at the front. (ii) and (iv) show the phase space within a window in x, while (iii) and (v) are the spatially averaged distribution function. (iii) and (v) serve as a proxy for approximating the propensity of Landau damping in that region. In fig. [\[fig:etch-f\]](#fig:etch-f){reference-type="ref" reference="fig:etch-f"}(ii) and (iii), we see that the rear of the wavepacket is Maxwellian. As previously shown, this is why the rear of the wavepacket damps faster than the front as in fig. [\[fig:etch-f\]](#fig:etch-f){reference-type="ref" reference="fig:etch-f"}(i) (). In the simulations described here, fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"} show that the distribution function at the back of the wavepacket has trapped particle activity (fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"}(ii)) and near zero slope at the phase velocity of the wave (fig. [\[fig:llived-f\]](#fig:llived-f){reference-type="ref" reference="fig:llived-f"}(iii)). Both plots show that the slope is negligible because of the arrival of streaming detrapped particles from the first wavepacket. Due to this effect, the remergence of Landau damping that occurs due to the loss of trapped particles in isolated wavepackets no longer occurs here. This results in a reduction of the etching and the wavepacket propagates freely for some time while the particles from the first wavepacket propagate and arrive at the rear of the second wavepacket. Figure [\[fig:spatial\]](#fig:spatial){reference-type="ref" reference="fig:spatial"} shows the results of the optimization process for the resonant spatial location. For \(t_1 < 500 \omega_p^{-1}\), the resonant location decreases as a function of wavenumber. From analyzing phase space, we have determined that the spatial location is related to the resonant electron transport i.e. \(x_1 \approx v_{ph} t_1\). Waves with larger wavenumbers have smaller phase velocities. Because of this, the resonant spatial location decreases as a function of wavenumber. For a fixed wavenumber, fig. [\[fig:locus\]](#fig:locus){reference-type="ref" reference="fig:locus"} shows a locus of points in space-time where long-lived wavepackets can be excited in the presence of a pre-existing wavepacket. This suggests the possibility of a critical space-time radius within which collisional relaxation has yet to occur, and non-linear effects can be exploited. Likewise, we also learn the dependence of the optimum frequency shift as a function of \(k_0\) and \(t_c\), as shown in fig. [\[fig:dw-surf\]](#fig:dw-surf){reference-type="ref" reference="fig:dw-surf"}. The learned frequency shift, a few percent here, is similar in magnitude as to that observed in previous work related to SRS (). Furthermore, the frequency shift increases in magnitude as a function of wavenumber. As before, waves with larger wavenumbers have smaller phase velocities, and therefore, interact with more particles. Because of this, waves with larger wavenumbers have a larger non-linear frequency shift associated with them, as we see here ().
# Conclusion
We show how one may be able to discover novel physics using differentiable simulations by posing a physical question as an optimization problem. This required domain expertise in determining which functional dependencies to learn using neural networks. In fig. [\[fig:workflows\]](#fig:workflows){reference-type="ref" reference="fig:workflows"}, we show how one may adapt an existing computational science workflow to the autodidactic process described here. In the work performed here, this process enabled the discovery of parameters in a 4D search space with known bounds but an unknown resolution requirement. We trained the model over a coarse grid in \(k_0\) and \(t_1\), and learned functions for \(x_1\) and \(\omega_1\). Using gradient descent here allows an escape from the curse of dimensionality and reduces the problem from a 4D search to a 2D search + 2D gradient descent. This discovery process is not limited to differentiable simulations. While in fig. [\[fig:workflows\]](#fig:workflows){reference-type="ref" reference="fig:workflows"}, \(\mathcal{V}\) represents a PDE solve, it only needs to be a AD-enabled function that is a model for a physical system. For example, rather than a PDE solve, \(\mathcal{V}\) could represent a pre-trained neural-network-based emulator for experimental data. In such a scenario, one may be able to learn forcing function parameters for an experiment using the proposed workflow. Finally, in neural network literature, the gradient required for the update is \(\partial \mathcal{S} / \partial \theta = \partial \mathcal{S} / \partial \mathcal{G} \times \partial \mathcal{G} / \partial \theta\). We see that this is the same as eq. [\[eq:nnloss\]](#eq:nnloss){reference-type="ref" reference="eq:nnloss"} after the addition of one more node in the computational graph for \(\mathcal{V}\), the function that models the physical system. This allows the neural network training process to become unsupervised and data-efficient. In future work, we aim to dive deeper into the discovered phenomenon to better understand the mechanism and derive some scalings. Reduced models of wavepacket dynamics in SRS remain useful for the development of inertial confinement fusion schemes where laser-plasma instabilities occur. | {'timestamp': '2022-06-06T02:16:29', 'yymm': '2206', 'arxiv_id': '2206.01637', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01637'} |
null | null |
# Introduction
Vertebrate locomotion is generally driven by central pattern generators (CPG), distributed networks of locomotor neurons that have evolved to generate oscillatory patterns of movements---or gaits---that suit an animal's biomechanics and its environment. Typically, signals from the brain stem can modulate the period of these oscillations in order to adjust walking or running speed. A more complex behaviour is the entrainment of movement to an external stimulus, such as playing or dancing to music, or walking in time with a companion. At least some rhythmic behaviour has been reproduced by continuous-time dynamical systems models of neurons. Two mutually inhibitory neural populations (a "half-center'' model) have a period that is easily adjusted by tonic input. Larger numbers of neural populations can trigger the same gait transitions and bistabilities seen in animals. Robots, compared to neurobiological models, generally employ simpler oscillators with fixed frequencies for building CPGs, although adaptive-frequency variations of these have been developed. One approach to frequency adaptation is continuous control, in which the frequency is controlled by a time-dependent variable that is continually adjusted according to some error signal. Buchli et al. used oscillators with phase errors explicitly fed back into the frequency, so that a robot's gait frequency approached the natural resonance of its passive joints. Iwasaki and Zheng used reciprocal coupling between a half-center CPG and a pendulum to allow synchronization between them. More recently, Egger et al. developed a suitable spiking neuron half-center model where the difference between input and output pulses are fed back to the tonic input . These methods, however, do not fully capture how humans entrain to external rhythms. Evidence from neuroscience suggests that brain oscillations mediate between rhythmic stimulus and the motor system in a top-down fashion, and relax to their normal frequencies after the stimulus ends. On the practical side, while the control theory approach can be applied to simple isochronous pulses, it is unclear how to obtain an error function for more complex periodic inputs. In neuroscience, open-loop models for entrainment have been developed that could in principle be applied to robotics. One is using large recurrent networks, in which time is encoded in the high-dimensional network state. Another approach is the gradient frequency network of Large et al., in which an oscillator with matching frequency or harmonic, becomes resonant out of a collection of several. It is, however, not clear how to integrate these relatively complex systems with a CPG. In this paper, we show that a CPG consisting of a small number of non-linear oscillators can rapidly entrain to a complex periodic signal through simple forcing. We demonstrate this idea with a quadruped model based on a network of Matsuoka oscillators that are modified to have input-dependent frequency. The network is composed of a CPG as well an intermediate cortical network that filters the external signal, with unidirectional coupling between the two parts. Parameters are optimized using multi-objective genetic algorithms, generating systems ranging from maximally flexible to maximally stable. We examine some limits of entrainment as a function of signal complexity, signal amplitude and the number of oscillators in the intermediate network.
# Neural Model
There exists a wide spectrum of mathematical models of oscillators that are commonly applied to central pattern generators. On one end we have the complexity and flexibility of spiking biological models, generally based on variations of the Wilson-Cowan model. This has recently found great success in closely reproducing the gait transitions of the mouse CPG, including regions of bistability. On the other end we have the predictable yet inflexible simple harmonic oscillator, which has been widely used in robotics due to its stability and ease of including feedback. In between there are many levels of variation in non-linearity and degrees of freedom, such as the Hopf oscillator, the Fitzhugh-Nagumo model, Amari-Hopfield networks and the Rowat-Selverston model. The Matsuoka neuron is a biologically motivated yet abstract two-variable model: \[\begin{aligned}
t_0 \frac{dx_i}{dt} &=-x_i-ay_i + I_i(t) \\ t_0 \frac{dy_i}{dt} &=-\gamma y_i + bh(x_i) \end{aligned}\] where \(h(x)\) is a rectified linear unit: \(h(x)=0\) for \(x\leq0\) and \(h(x)=x\) for \(x> 0\). Like most biological models, there is a fast "spiking" variable (\(x\)) and a slow "recovery" variable (\(y\)). Below a critical value of \(a\), a single neuron will not oscillate. However, when two or more neurons are given interconnections using \[I_i(t) = \sum_j w_{ij}h(x_j-\theta_{ij})\] where \(w_{ij}\) and \(\theta_{ij}\) are weights and output thresholds, respectively, it is easy to produce oscillatory CPG-like patterns via mutual inhibition or excitation. The property of a rectified output simplifies the analysis of the phase space. In addition, compared to the Amari-Hopfield model (also sometimes called a continuous-time RNN), the region of phase space containing limit cycles is much larger. However, unlike biological neurons, where the level of constant input can be used to control the firing rate and hence the oscillation frequency, the firing rate of the Matsuoka model is insensitive to tonic input. Hence, we introduce a sigmoidal activation function \(S(x)\) akin to those used to model persistent sodium currents in motor neurons: \[\begin{aligned}
t_0 \frac{dx_i}{dt} &=-x_i-aS(\kappa[x_i-x_0]) y_i + c_i + d_i I_{DC} + I_{ACi}(t) \\ t_0 \frac{dy_i}{dt} &=-\gamma y_i + bh(x_i) \end{aligned}\] where \(S(x)=1/(1+\exp(x))\). Here \(d_i\) is the coefficient for the tonic brain stem drive \(I_{DC}\) used to tune the gait period. When the coefficients satisfy: \[c_i + d_i I_{DC} > x_0 + \frac{2}{k} \;,\] this reproduces the ubiquitous cubic-like shape of the fast variable's nullcline. In this case, each neuron may self-oscillate for a certain parameter range.
# CPG Model
We assembled a quadruped CPG as a modular system of limb controllers connected by interneurons, in a similar fashion to Beer and Ijspeert. For an overview, see Fig. [\[fig:diagram\]](#fig:diagram){reference-type="ref" reference="fig:diagram"}. Each limb controller contains a single interneuron and two motor neurons ('A' and 'B'), the latter of which can be used for a single extensor-flexor pair, or for a pair of single-variable joints. Each of the three neurons has its own bias \(c_i\) and drive coefficient \(d_i\), which are identical for all modules. Thresholds between all CPG neurons \(\theta_{ij}\) are zero, and connection weights \(w_{ij}\) are zero between modules, apart from when \(i\) and \(j\) are both interneurons. The CPG has lateral symmetry, so that connection weights are equal for equivalent connections between the left and right side of the body.
## CPG Optimization
The CPG was implemented in Python[^1] and optimized using the NSGA3 genetic algorithm included in the DEAP toolbox. Each parameter that is not set according to the above constraints is encoded by an integer between 1 and 10, which determines its value within the range shown in Table [1](#paramtable){reference-type="ref" reference="paramtable"}. Note that \(c_i\) and \(w_{ij}\) (for connections that exist according to the schematic) cannot be zero. In addition, connections between interneurons are constrained to be inhibitory (\(w_{ij}<0\)). A given CPG was evaluated by iterating the brainstem drive \(I_{DC}\) from 0 to 1 in steps of \(0.1\). For each drive (indexed by \(k\)), the system was given random initial conditions, followed by a burn-in period, after which the time series of a flexor-extensor-type output (difference between rectified A and B neuron outputs) was analysed for each of the four limb modules. First, the periods were measured using the maximum of the autocorrelation function, and reduced to the mean oscillation period \(T_k\) and coefficient of variation \(\mathrm{CV}_{Tk}\) over the four limbs. A measured correlation peak at a time less than \(0.01\)s was considered non-oscillating and hence invalid. A period shift of \(|T_{k+1}-T_k|/(T_{k+1}+T_k)>0.15\) was also considered invalid in order to filter out large discontinuities. In addition, the mean oscillation amplitude \(A_k\) and coefficient of variation \(\mathrm{CV}_{Ak}\) were measured using a peak finding algorithm on these same time series. Mean amplitudes less than \(0.1\) or greater than 10 were also considered invalid in order to keep all CPG outputs within a comparable range. Finally, "duty functions" \(D_{Ak}\) and \(D_{Bk}\) were measured to penalise unbalanced gaits in which three or more limbs are activated simultaneously: \[D_{A} = E_t\left[\sum_i^4 I\left(\left|\frac{d}{dt} h(x_{Ai})\right|>\epsilon\right) < 3\right] \;,\] and similarly for \(D_B\), where \(E_t\) is the time domain expectation value, \(I\) is the binary indicator function, and \(\epsilon=0.001\) in this study. If there is no oscillation for a given \(k\) then both \(D_{Ak}\) and \(D_{Bk}\) are given a value of zero. The use of the derivative was to allow the possibility of consistent flat output, while penalizing simultaneous spiking of equivalent neurons in three or more limbs. Three fitness functions were to be maximized by the multi-objective optimization: \[\begin{aligned}
F_1 &= \left|\sum_{k=1}^{N-1} V_k V_{k+1}\frac{T_{k+1}-T_k}{T_\mathrm{max}}\right| \\ F_2 &= \frac{1}{N}\sum_k^N \frac{1}{1 + \mathrm{CV}_{Tk} + \mathrm{CV}_{Ak}} \\ F_3 &= \frac{1}{2N}\sum_k^N D_{Ak} + D_{Bk}\\ \end{aligned}\] where \(V_k\) is the validity of the time series during drive \(k\), averaged over the four outputs, with each validity being 0 or 1. NSGA3 generates a Pareto front containing non-dominated individuals according to these fitnesses, thus selecting for CPGs with some combination of large monotonic variation in period as a function of \(I_{DC}\), low variation between limbs, and two or more limbs consistently in the "stance" segment of the cycle. We used a population of 48 individuals that evolved using two-point crossover and mutation for 100 generations. Each individual was evaluated again 5 times in order to use medians as accurate final fitnesses, and from these a final Pareto front of CPGs was generated. Four solutions were selected in a way that balanced variety and average overall performance. First the highest overall fitness was selected using the sum of fitnesses. Then, three more individuals were selected using the maxima of \[F_m^* = zF_m + \sum_{k=1}^3 F_k\] where \(z\) was incremented in intervals of one until the maximum of each \(F_m^*\) was unique.
# Filter Network
For each of the three selected CPGs, a filter was evolved on top. The purpose of the filter is to pre-process the input and distribute the signal among the CPG modules as neuron-like pulses. The filter network was a single non-lateralized layer of \(n\) neurons. The coupling to the input was then governed by \(n\) coefficients. In order to not disturb the CPG in the absence of input, the neuron parameters were set to be below the spontaneous bursting threshold, and all interconnections were made to be inhibitory (\(w_{ij} \leq 0\)). However, this does not entirely preclude oscillation of the network, and so further measures were taken in the fitness function below. From the input, \(n\) coefficients \(G_i\) govern the coupling to the neurons, while 4\(n\) coefficients governed the coupling from the filter to the four CPG interneurons (forming the matrix M in Fig. [\[fig:diagram\]](#fig:diagram){reference-type="ref" reference="fig:diagram"}). For these connections, the offset \(\theta_{ij}\) was set to a constant value \(\theta_0\) to offset the equilibrium output of the neurons.
## Filter Network Optimization
The input consisted of spikes at a regular time interval with period \(\tau\), which was then low-pass filtered using an exponentially decaying impulse response with decay constant \(\Gamma / t_0\). The periods \(\tau_k\) used were \(2/3\), \(1\) and \(3/2\) multiplied by \(T_{0.5}\), where \(T_{0.5}\) is the median period for a tonic input of \(I_{DC} = 0.5\). Together these present a range of more than a factor of two in input period. Each neuron in the filter network received the input multiplied by its own coefficient \(F_i\). The network was evolved to minimise the mean difference between the input period and the periods of the motor neurons. Although this is a relatively simple objective function, NSGA3 was used again to avoid early convergence to a local optimum. The period \(T_{ik}\) of the four CPG motor neurons was measured as in the previous section, and the NSGA3 algorithm was run using three fitness functions, one for each \(\tau_k\): \[F_{fk} = \frac{V_k}{1 + \sigma_0 / \sigma_t + \sqrt{\sum_i{(T_{ik}-\tau_k)^2}/4}}\] where the validity \(V_k\) is defined as in the previous section, \(\sigma_0\) is the mean standard deviation of the filter output with no input and \(\sigma_t\) is a scaling threshold, set to \(0.1\) for the current study. A population of 68 was evolved for 50 generations for filters with \(n=4\) neurons, and for 25 generations for \(n=2\) due to the much smaller number of parameters. The filter with the highest overall sum of fitnesses was then chosen for analysis. Finally, for the overall best CPG, an additional filter was evolved to test the ability to entrain to more complex signals. This was done in the same way as the original filter, but with every fourth pulse missing from the input.
# Results
The three CPG fitnesses reached stable levels within 100 generations, with the final Pareto front shown in Fig. [\[fig:cpgevo\]](#fig:cpgevo){reference-type="ref" reference="fig:cpgevo"}. Results for CPGs maximizing the four weighted averages are shown in Table [\[cpgresults\]](#cpgresults){reference-type="ref" reference="cpgresults"}. The overall highest mean fitness was achieved by a bounding gait, with front and hind limb pairs moving in synchrony, while a variety of walking gaits maximized the fitnesses weighted towards individual components. The four CPGs encompassed a wide range of periods, from \(0.47\) to \(1.8\) seconds. The CPG with highest overall fitness was also the most successful at entrainment, with fitness close to the maximum of one for both 2-neuron and 4-neuron filter configurations, as shown in Fig. [\[fig:filtevo\]](#fig:filtevo){reference-type="ref" reference="fig:filtevo"}. The full CPG0 system was able to generalize beyond the input periods and below the amplitude used for training, as shown in Figure [\[fig:periods\]](#fig:periods){reference-type="ref" reference="fig:periods"}. Notably, the CPG with the least flexible period was least able to evolve a filter to entrain the CPG at other drive periods. The walking gaits in general had more difficulty generalizing entrainment to arbitrary periods, sometimes entraining at a multiple of the input period but often having periods not associated with the input, suggesting highly nonlinear behaviour. Period doubling was also seen in the filter of CPG2, which is an expected outcome of driving the systems at high frequencies beyond those used for training. The amplitude profiles as a function of input period were also mostly flat as shown in Figure [\[fig:amps\]](#fig:amps){reference-type="ref" reference="fig:amps"}, implying very wide transfer functions. The output amplitude curves, however, show substantial jumps in three of the CPGs when the input amplitude reaches a certain threshold. This can be driven either by the filter or the CPG dynamics. Filters were then evolved for CPG0 using a non-isochronous input. The resulting fitness (median \(F_f=0.61\) for \(n=4\), \(F_f=0.39\) for \(n=2\)) were lower than for the filter evolved for isochronous input. However, these filter performs reasonably well on isochronous input (median \(F_f=0.64\) for \(n=4\), \(F_f=0.91\) for \(n=2\)), and better than the original CPG0 filters on non-isochronous input (median \(F_f=0.30\) for \(n=4\), \(F_f=0.33\) for \(n=2\)). As shown in Figure [\[fig:timeseries\]](#fig:timeseries){reference-type="ref" reference="fig:timeseries"}, in both cases the \(n=4\) system adjusts its period rapidly to the stimulus, retaining its original gait pattern, and relaxing back to its original period after the stimulus ends.
# Discussion
To our knowledge, this is the first demonstration of open-loop rhythmic entrainment in a central pattern generator model. This was achieved via a novel modification of a well-used neural model, with evolutionary optimization not only of connection weights but also of several system parameters that determine the degree of nonlinearity. In this architecture, wide tunability of the oscillation period with tonic input appears to be a precondition for synchronization to a wide range of periodic inputs. The neural model we develop captures this important property while remaining simple enough for use in robotic applications. Our results demonstrate that only a small total number of neurons (16) are required for robust entrainment compared to the number of neuron populations in biological quadruped CPGs (with recent modelling studies involving several dozen ). However, the difference in results between two-neuron and four-neuron filters suggests that more neurons will add robustness and flexibility for more complex gaits (such as quadruped walking), transitioning gaits, and more complex inputs. Although communication between the brain and body is clearly two-way in animals, this work presents a complementary approach to closed-loop control methods for temporal prediction. The synchronization-based approach circumvents the problem of determining a suitable error function for complex temporal patterns. For entrainment, feedback may take on a fine-tuning role, as is hinted at by brain research that differentiates processes occurring on long and short time-scales. Our conceptual model has applications both within neurophysiological research and in the design of intelligent systems. For the former, our framework can be further developed to investigate general principles behind rhythmic entrainment and embodied cognition. Practical applications include beat tracking and adaptive and social robotics . In the future, this system will be tested in simulated and physical robots. The addition of sensory and balancing feedback is expected to improve the overall stability of the system. The real-time low-level adaptive behaviour that we demonstrate also opens up the possibility of realistic human-robot and robot-robot interaction. To this end, studies with multiple, mutually interacting agents will allow the study of emerging collective behaviors.
[^1]: Genotypes of individuals in this paper and source code to generate all results are available at <https://github.com/aszorko/COROBOREES/tree/Paper1>. | {'timestamp': '2022-06-06T02:16:29', 'yymm': '2206', 'arxiv_id': '2206.01638', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01638'} |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction {#sec:intro}
Recent studies of rotation of old Sun-like stars suggest a change in stellar activity of middle-aged main-sequence stars. There has been accumulating evidence that the rotation periods of Sun-like stars cease to scale as the square-root of age in the latter halves of their lives. A similar transition has also been noted in chromospheric activities. These may suggest a corresponding change in the mechanism of magnetic field generation that occurs once the rotation period \(P_\mathrm{rot}\) becomes comparable to the convective turnover timescale \(\tau_\mathrm{c}\). Quasi-periodic brightness modulation of stars in broad-band photometry is well-suited for statistical studies of rotational evolution. In particular, light curves from the prime *Kepler* mission have been used to derive rotation periods up to months for tens of thousands of Sun-like stars. Previous investigations of the \(P_\mathrm{rot}\) distribution of *Kepler* stars using the sample of have shown that the observed distribution is truncated roughly around the solar Rossby number \(\mathrm{Ro}=P_\mathrm{rot}/\tau_\mathrm{c}\) and also exhibits a pile-up around slightly shorter \(P_\mathrm{rot}\). While these features are in qualitative agreement with the stalled spin down scenario, the effects of rotational evolution and detection bias have not been clearly disentangled. In general, longer-period tail of the \(P_\mathrm{rot}\) distribution is most prone to the detection bias, and so it requires good understanding of the bias to correctly interpret the observed distribution. In this work, we attempt to better understand the detection bias in the \(P_\mathrm{rot}\) sample constructed by to aid statistical interpretation of the sample. To do so, we first investigate the generic relation between spot-modulation amplitudes \(R_\mathrm{per}\) and rotation periods \(P_\mathrm{rot}\) for Sun-like main-sequence stars in the sample (as defined in Section [2](#sec:sample){reference-type="ref" reference="sec:sample"}), and derive a relation that predicts the modulation amplitude \(R_\mathrm{per}\) given \(P_\mathrm{rot}\) and \(T_\mathrm{eff}\) (Section [3](#sec:analysis){reference-type="ref" reference="sec:analysis"}). Then we clarify how this dependence is combined with the magnitude-dependent detection threshold for rotational modulation to sculpt the observed distribution of \(P_\mathrm{rot}\) as a function of \(T_\mathrm{eff}\) (Section [4](#sec:detection){reference-type="ref" reference="sec:detection"}). We also show that the general pattern derived here is consistent with the sample from that include more detections of rotation periods, and that the latter catalog is subject to a different detection function. In Section [5](#sec:discussion){reference-type="ref" reference="sec:discussion"}, we discuss our finding in connection with coronal and chromospheric activity indicators and the weakened magnetic braking hypothesis, and propose a test to check the veracity of our view on the detection bias further.
# The Sample {#sec:sample}
performed a homogeneous search for quasi-periodic brightness modulation associated with stellar rotation in *Kepler* light curves, and reported detections of robust rotational modulation for \(\approx 34,000\) stars. All these stars are assigned the rotation period \(P_\mathrm{rot}\) as determined from the auto-correlation analysis, along with the average amplitude of variability within one period in units of parts-per-million (ppm), \(R_\mathrm{per}\), defined as the median of the differences between 95th and 5th percentiles of normalized flux in each rotation period cycle. The Large Sky Area Multi-Object Fibre Spectroscopic Telescope (LAMOST) project, on the other hand, provided spectroscopic parameters for \(\approx 60,000\) *Kepler* stars in their sixth data release (DR6). We work on the overlap of the two samples, for which \(P_\mathrm{rot}\), \(R_\mathrm{per}\), and \(T_\mathrm{eff}\) have been derived homogeneously. One concern is the presence of unresolved binaries. The contaminating flux from the secondary affects both the inferred modulation amplitude and stellar classification. Tidal interactions with a close-in companion also affect the rotation period, although such close-in companions would occur in \(\lesssim 10\%\) for Sun-like stars. The analysis may also be complicated by evolved stars whose rotation may have changed due to evolution of internal structure rather than magnetic braking. We use the information on absolute magnitudes made available by *Gaia* to remove such objects as possible. We start with 8,772 unique stars, for which robust periods are detected in and LAMOST DR6 data are publicly available.[^1] For LAMOST stars with multi-epoch observations, the mean of \(T_\mathrm{eff}\) was adopted.[^2] We then used the cross-match service of the Centre de Données astronomiques de Strasbourg (CDS) to find closest *Gaia* EDR3 sources within 5 arcsec and with `parallax_over_error` greater than 10. We find 8,309 matches. The difference between the \(G\)-band and *Kepler*-band magnitudes, latter taken from, has the mean of \(-0.02\) and standard deviation of 0.09, indicating correct matches. We then placed these stars on the absolute *Gaia* magnitude--LAMOST \(T_\mathrm{eff}\) diagram focusing on stars with \(4,000\,\mathrm{K}<T_\mathrm{eff}<6500\,\mathrm{K}\), defined the main sequence by fitting a 5th-order polynomial iteratively clipping \(1\sigma\) and \(3\sigma\) outliers below and above the sequence respectively, and removed stars deviating by more than 0.5 magnitudes at a given \(T_\mathrm{eff}\) from the derived sequence. This removes bright sources that may be either evolved stars or unresolved binaries, where the threshold of 0.5 is chosen to remove the sequence of equal-brightness binaries that are brighter than single stars by 0.75 magnitudes at a given \(T_\mathrm{eff}\). This cut left us with 5,022 stars. We also removed stars with multiple LAMOST measurements in which maximum and minimum radial velocities differ by more than \(20\,\mathrm{km/s}\), above which the distribution of velocity differences exhibits a clear excess from a Gaussian distribution that appears to represent measurement uncertainties. This velocity cut left us with 4,977 stars. We also removed stars with LAMOST \(\log g<4\) in the remaining sample because they shared the same locations in the \(\log g\)--\(T_\mathrm{eff}\) plane as the stars removed in the above cuts, and obtained the final sample of 4,968 stars. The remaining discussion relies on (subsets of) these 4,968 stars with LAMOST \(T_\mathrm{eff}=4,000\)--\(6,500\,\mathrm{K}\). The selection is visualized in the absolute *Gaia* magnitude--\(T_\mathrm{eff}\) plane in Figure [\[fig:hr\]](#fig:hr){reference-type="ref" reference="fig:hr"}.
# Evolution of Modulation Amplitudes {#sec:analysis}
Here we investigate how the photometric modulation amplitude \(R_\mathrm{per}\) evolves as a function of \(P_\mathrm{rot}\). As has been shown in previous works, the relation between \(R_\mathrm{per}\) and \(P_\mathrm{rot}\) for roughly solar-mass stars is concisely summarized in terms of the Rossby number \(\mathrm{Ro}=P_\mathrm{rot}/\tau_\mathrm{c}\). We revisit such a relation for our sample stars and use it for the discussion of detectability in Section [4](#sec:detection){reference-type="ref" reference="sec:detection"}.
## Spot-Modulation Amplitude vs Rotation Period {#ssec:amp_prot}
In Figure [\[fig:r_prot_example\]](#fig:r_prot_example){reference-type="ref" reference="fig:r_prot_example"} and Figure [\[fig:r_prot_all\]](#fig:r_prot_all){reference-type="ref" reference="fig:r_prot_all"} in Appendix, we show \(R_\mathrm{per}\) and \(P_\mathrm{rot}\) for our sample stars separated into 150\(\,\mathrm{K}\) bins ranging from 4000 to 6400 K. In Figure [\[fig:r_prot_example\]](#fig:r_prot_example){reference-type="ref" reference="fig:r_prot_example"}, we show two \(T_\mathrm{eff}\) bins separated by \(\sim 1,000\,\mathrm{K}\) to illustrate two typical behaviors: (i) \(R_\mathrm{per}\) is roughly constant at shorter \(P_\mathrm{rot}\) and exhibits a power-law decay at longer \(P_\mathrm{rot}\), and (ii) the transition period, which we denote by \(P_\mathrm{break}\), is shorter for hotter stars (bottom panel). The data for other \(T_\mathrm{eff}\) ranges in Figure [\[fig:r_prot_all\]](#fig:r_prot_all){reference-type="ref" reference="fig:r_prot_all"} show that the same trend holds continuously over the most \(T_\mathrm{eff}\) range, except for the coolest and hottest stars in the sample (see below). To quantify this visual trend, we model the data with the following broken power-law function: \[\label{eq:plmodel} R_\mathrm{per}(P_\mathrm{rot}; \bm{\theta}) = \begin{cases} R_\mathrm{break}\left(P_\mathrm{rot} \over P_\mathrm{break}\right)^{\beta_\mathrm{sat}} \quad &\mathrm{for}\quad P_\mathrm{rot} < P_\mathrm{break}\\ R_\mathrm{break}\left(P_\mathrm{rot} \over P_\mathrm{break}\right)^{\beta_\mathrm{unsat}}&\mathrm{for}\quad P_\mathrm{rot} > P_\mathrm{break} \end{cases}\] and infer \(\bm{\theta}\equiv (R_\mathrm{break}, P_\mathrm{break}, {\beta_\mathrm{sat}}, {\beta_\mathrm{unsat}})\) for stars in each \(T_\mathrm{eff}\) bin. The subscripts "sat\" and "unsat\" stand for saturated and unsaturated regimes, respectively, following the existing nomenclature---although both regimes here fall within the so-called "unsaturated\" regimes of other activity indicators (see Section [5.1](#ssec:comparison){reference-type="ref" reference="ssec:comparison"}). We also model the measurement uncertainties as well as intrinsic scatters around this deterministic relation, assuming that the measured values of \(\ln R_\mathrm{per}^{\mathrm{obs}}\) and \(\ln P_\mathrm{rot}^{\mathrm{obs}}\) for each star follow independent Gaussian distributions around the model, with common standard deviations \(\sigma_{\lnR_\mathrm{per}}\) and \(\sigma_{\lnP_\mathrm{rot}}\). We also infer these parameters as well as the "true\" value of \(P_\mathrm{rot}\) of each star, bringing the total number of parameters to be five plus the number of stars in each \(T_\mathrm{eff}\) subsample. The likelihood function is therefore \[\begin{aligned}
\notag &p(\{\ln R_\mathrm{per}^{\mathrm{obs},j}\}, \{\ln P_\mathrm{rot}^{\mathrm{obs},j}\}|\{P_\mathrm{rot}^j\}, \bm{\theta}, \sigma_{\lnR_\mathrm{per}}, \sigma_{\lnP_\mathrm{rot}}) \\ \notag &= \prod_j \left[\mathcal{N}(\lnR_\mathrm{per}^{\mathrm{obs},j}; \lnR_\mathrm{per}(P_\mathrm{rot}^j; \bm{\theta}), \ln\sigma_{R_\mathrm{per}}) \right.\\ & \left. \qquad \quad \times \ \mathcal{N}(\lnP_\mathrm{rot}^{\mathrm{obs},j}; \lnP_\mathrm{rot}^j, \ln\sigma_{P_\mathrm{rot}})\right] \end{aligned}\] where \(j\) is the label for stars in each subsample, \(\{x^j\}\) denotes the set of \(x\) in the subsample, and \(\mathcal{N}(x; \mu, \sigma)\) is the Gaussian distribution for \(x\) with mean \(\mu\) and standard deviation \(\sigma\). We consider this broken-power law model as a simple mathematical tool that is useful to quantify how the \(R_\mathrm{per}\)--\(P_\mathrm{rot}\) (or \(\mathrm{Ro}\)) relation depends on \(T_\mathrm{eff}\), and do not claim that this function provides the correct description of the relation (nor do we attempt to identify it). Indeed, we see a hint of more detailed structures than described by this model, as will be discussed below. The inference was performed in a Bayesian manner. We adopt independent prior probability density functions (PDFs) for \(\bm{\theta}\), \(\sigma_{\lnR_\mathrm{per}}\), \(\sigma_{\lnP_\mathrm{rot}}\), and \(\{P_\mathrm{rot}^{\mathrm{obs},j}\}\) as summarized in Table [\[tab:priors\]](#tab:priors){reference-type="ref" reference="tab:priors"}, and infer the joint posterior PDF for these parameters given the data \((\{R_\mathrm{per}^{\mathrm{obs},j}\}, \{P_\mathrm{rot}^{\mathrm{obs},j}\})\) by drawing samples from the posterior PDF using Hamiltonian Monte Carlo with No-U-Turn sampler as implemented in `NumPyro`. In each subsample, we only model the stars whose rotation periods fall between their 5th and 95th percentiles. In Figure [\[fig:r_prot_example\]](#fig:r_prot_example){reference-type="ref" reference="fig:r_prot_example"} and [\[fig:r_prot_all\]](#fig:r_prot_all){reference-type="ref" reference="fig:r_prot_all"}, those stars omitted from fitting are shown as blue open squares. They are both more sensitive to detection bias against weaker modulation in fainter stars that has not been taken into account in our model: the few rapid rotators tend to be rarer and thus fainter, and the slowest rotators tend to have smaller \(R_\mathrm{per}\). The lack of detection model may be considered as a limitation of our analysis and may introduce systematic errors in the inferred parameters; see Section [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"} for further discussion.
The solid orange line and shaded region in Figure [\[fig:r_prot_example\]](#fig:r_prot_example){reference-type="ref" reference="fig:r_prot_example"} shows the mean and 5th-95th percentile of the prediction by the broken power-law model in Equation [\[eq:plmodel\]](#eq:plmodel){reference-type="ref" reference="eq:plmodel"}. The inferred \(P_\mathrm{break}\) (mean and 5th-95th percentile of the marginal posterior PDF) are also shown with vertical dotted line and shade. Our model fitting locates \(P_\mathrm{break}\) that is seen visually, which decreases with increasing \(T_\mathrm{eff}\). A similar pattern was not found robustly for the lowest \(T_\mathrm{eff}\) bin, presumably due to a small number of data points. The break is implied but the pattern is different for stars with \(T_\mathrm{eff}>6,250\,\mathrm{K}\); this is not surprising either because they are the stars above the Kraft break that follow different \(P_\mathrm{rot}\) evolution from cooler stars with convective envelopes. Below we omit these bins and consider stars with \(T_\mathrm{eff}=4,250\,\mathrm{K}\)--\(6,250\,\mathrm{K}\), unless otherwise noted. Figure [\[fig:params_teff\]](#fig:params_teff){reference-type="ref" reference="fig:params_teff"} shows the values of \(R_\mathrm{break}\), \(P_\mathrm{break}\), \({\beta_\mathrm{unsat}}\), \({\beta_\mathrm{sat}}\) inferred in each \(T_\mathrm{eff}\) bin. The values of \(R_\mathrm{break}\), \({\beta_\mathrm{unsat}}\), and \({\beta_\mathrm{sat}}\) are largely insensitive to \(T_\mathrm{eff}\), although the presence of possible trends is not excluded for \(R_\mathrm{break}\) and \({\beta_\mathrm{unsat}}\) (see also Section [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"}). A clear \(T_\mathrm{eff}\) dependence is seen for \(P_\mathrm{break}\), which shortens quickly with increasing \(T_\mathrm{eff}\) in agreement with the visual appearance of Figure [\[fig:r_prot_example\]](#fig:r_prot_example){reference-type="ref" reference="fig:r_prot_example"} and Figure [\[fig:r_prot_all\]](#fig:r_prot_all){reference-type="ref" reference="fig:r_prot_all"}. This suggests that \(P_\mathrm{break}\) scales with \(\tau_\mathrm{c}\), and motivates the analysis in the next subsection. The strong curvature in the \(P_\mathrm{break}\)--\(T_\mathrm{eff}\) relation does not favor power-law scaling with parameters that depend more weakly on \(T_\mathrm{eff}\), such as stellar mass and radius.
## Spot-Modulation Amplitude vs Rossby Number {#ssec:amp_ro}
The exact value of \(\tau_\mathrm{c}\) depends on how it is estimated. The formula by has widely been used, which is based on theoretical evaluation of local turnover timescale near the bottom of the convective envelope as described in and has been calibrated to minimize the scatter in the \(\log R'_\mathrm{HK}\)--\(\mathrm{Ro}\) relation. Other scales have been proposed based on up-to-date stellar models and direct inference of the thickness of convective envelopes from asteroseismology . These works generally yield \(\tau_\mathrm{c}\) values that are larger by a factor of a few than for solar-mass main-sequence stars and have different dependence on \(T_\mathrm{eff}\). In this work, we use the formula in based on theoretical models of, simply because it relates \(\tau_\mathrm{c}\) directly to \(T_\mathrm{eff}\) which we work on, and because it is close to the traditional scale by for dwarf stars and allows for easier comparisons with other works. We find that the typical \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation in our sample remains unchanged for the \(\tau_\mathrm{c}\) prescriptions in and.[^3] Thus the following discussion in Sections [4](#sec:detection){reference-type="ref" reference="sec:detection"} and [5](#sec:discussion){reference-type="ref" reference="sec:discussion"} is insensitive to which of these prescriptions is adopted, as long as \(\mathrm{Ro}\) is scaled adequately; see Appendix [\[ssec:tauc_amp\]](#ssec:tauc_amp){reference-type="ref" reference="ssec:tauc_amp"} for details of these analyses. Here we repeat almost the same analysis as in Section [3.1](#ssec:amp_prot){reference-type="ref" reference="ssec:amp_prot"}, but replacing \(P_\mathrm{rot}\) with the Rossby number \(\mathrm{Ro} = P_\mathrm{rot}/\tau_\mathrm{c}\) evaluated for each star; consequently, the model parameters \(P_\mathrm{break}\), \(\sigma_{\lnP_\mathrm{rot}}\), and \(\ln P_\mathrm{rot}^j\) in Equation [\[eq:plmodel\]](#eq:plmodel){reference-type="ref" reference="eq:plmodel"} are also replaced with \(\ro_\mathrm{break}\), \(\sigma_{\ln\mathrm{Ro}}\), and \(\ln \mathrm{Ro}^j\), respectively (bottom part of Table [\[tab:priors\]](#tab:priors){reference-type="ref" reference="tab:priors"}). As noted above, the \(\tau_\mathrm{c}\) value was calculated using the \(\tau_\mathrm{c}\)--\(T_\mathrm{eff}\) relation ( Equation 36) given in. Figure [\[fig:r_ro_example\]](#fig:r_ro_example){reference-type="ref" reference="fig:r_ro_example"} is analogous to Figure [\[fig:r_prot_example\]](#fig:r_prot_example){reference-type="ref" reference="fig:r_prot_example"}; note that the \(x\)-axis is now \(\mathrm{Ro}\). We see a similar broken power-law pattern as seen in the \(R_\mathrm{per}\)--\(P_\mathrm{rot}\) plane, but now the break occurs at similar \(\mathrm{Ro}\) in different \(T_\mathrm{eff}\) bins. The same is also true in other \(T_\mathrm{eff}\) bins as shown in Figure [\[fig:r_ro_all\]](#fig:r_ro_all){reference-type="ref" reference="fig:r_ro_all"}, which is analogous to Figure [\[fig:r_prot_all\]](#fig:r_prot_all){reference-type="ref" reference="fig:r_prot_all"}. Figure [\[fig:params_teff_ro\]](#fig:params_teff_ro){reference-type="ref" reference="fig:params_teff_ro"} is analogous to Figure [\[fig:params_teff\]](#fig:params_teff){reference-type="ref" reference="fig:params_teff"}, where now the break period \(P_\mathrm{break}\) is replaced with the break Rossby number \(\ro_\mathrm{break}\). The break location \(\ro_\mathrm{break}\) now depends much less on \(T_\mathrm{eff}\) than \(P_\mathrm{break}\) did, while the other parameters remain similar to those in Section [3.1](#ssec:amp_prot){reference-type="ref" reference="ssec:amp_prot"}; this is reasonable because \(\tau_\mathrm{c}\) in each narrow temperature bin is almost the same, and so the transition from \(P_\mathrm{rot}\) to \(\mathrm{Ro}\) shift the whole data by almost the same amount in the \(x\)-direction. On the other hand, we also see some correlated pattern; for example, \(|{\beta_\mathrm{unsat}}|\) and \(\ro_\mathrm{break}\) may be systematically larger at higher \(T_\mathrm{eff}\).[^4] This could be of astrophysical origin, may be an artifact due to our imperfect knowledge of \(\tau_\mathrm{c}(T_\mathrm{eff})\) (see Appendix [\[ssec:tauc_amp\]](#ssec:tauc_amp){reference-type="ref" reference="ssec:tauc_amp"}), may be due to detection bias (see Section [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"}), or a combination of these effects. It is beyond the scope of this work to account for this possible dependence. Motivated by the (roughly) \(T_\mathrm{eff}\)-independent nature of the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation, in Figure [\[fig:r_ro\]](#fig:r_ro){reference-type="ref" reference="fig:r_ro"} we show all stars with \(T_\mathrm{eff}=4,000\,\mathrm{K}\)--\(6,250\,\mathrm{K}\) in the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) plane, and fit a single broken power-law relation \(R_\mathrm{per}(\mathrm{Ro})\) to the entire data (orange solid line and shade). We find \(R_\mathrm{per}=(1.08\pm0.04)\times10^4\,\mathrm{ppm}\), \(\ro_\mathrm{break}=0.84\pm0.02\), \({\beta_\mathrm{sat}}=-0.04\pm0.09\), and \({\beta_\mathrm{unsat}}=-2.6\pm0.1\) (mean and 90%). These values are also shown in Figure [\[fig:params_teff_ro\]](#fig:params_teff_ro){reference-type="ref" reference="fig:params_teff_ro"} with horizontal orange dashed lines, which broadly agree with the values derived in separate \(T_\mathrm{eff}\) bins. Here we see that the result from the entire sample tend to be closer to those of hotter stars, simply because they are more numerous in the sample. We discuss possible systematic errors in the inferred parameters due to detection bias against weak modulation further in Section [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"}. The kink in the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation has been noted in other works. The value of \(\mathrm{Ro}\) corresponding to the kink has been found to be \(\sim 0.4\) in who estimated \(\tau_\mathrm{c}\) using stellar models by; 0.23 in for their \(\tau_\mathrm{c}\) calibration using seismic stars; and 0.82 when adopted the prescription by. When scaled by \(\mathrm{Ro}_\odot\), all these values roughly agree with what we found, \(\mathrm{Ro}/\mathrm{Ro}_\odot \sim 0.4\) (since \(\mathrm{Ro}_\odot\approx 2\) in our scale). We also note a wiggle for stars with \(\mathrm{Ro}\sim 0.4\)--\(0.8\) in our scale (or \(0.2\)--\(0.4\,\mathrm{Ro}_\odot\)), which has also been noted by and is seen in our analyses using other \(\tau_\mathrm{c}\) prescriptions (see Figure [\[fig:tauc\]](#fig:tauc){reference-type="ref" reference="fig:tauc"} in Appendix [\[ssec:tauc_amp\]](#ssec:tauc_amp){reference-type="ref" reference="ssec:tauc_amp"}). While we do not understand its origin, we see a hint of a similar structure in the X-ray data, suggesting that this may not be an artifact related to the calibration of \(\tau_\mathrm{c}\). See Section [5.1](#ssec:comparison){reference-type="ref" reference="ssec:comparison"} for further discussion.
## On the Impact of Detection Bias {#ssec:r_ro_bias}
Figure [\[fig:r_ro\]](#fig:r_ro){reference-type="ref" reference="fig:r_ro"} shows that there exists some dispersion in \(R_\mathrm{per}\) at a fixed value of \(\mathrm{Ro}\): the dispersion is inferred to be \(\approx 0.24\,\mathrm{dex}\) from our modeling in Section [3.2](#ssec:amp_ro){reference-type="ref" reference="ssec:amp_ro"}. Some of the dispersion may be due to difference in spin-axis inclinations and/or activity cycles. Any systematics in the adopted \(\tau_\mathrm{c}\)--\(T_\mathrm{eff}\) relation can also affect the scatter. In particular, we did not take into account its possible dependence on \[Fe/H\]. Indeed, we find that the residuals \(\Delta\log_{10}R_\mathrm{per}\) of the fit in Figure [\[fig:r_ro\]](#fig:r_ro){reference-type="ref" reference="fig:r_ro"} is correlated with the LAMOST \[Fe/H\], with the Pearson R coefficient being \(\approx 0.4\) or \(\Delta\log_{10}R_\mathrm{per} \approx 0.5\mathrm{[Fe/H]}\). This is qualitatively consistent with the finding of that metal-rich stars have enhanced activities, although the dependence may be weaker than was found to be typical by these authors. As will be discussed in Section [4](#sec:detection){reference-type="ref" reference="sec:detection"}, we find evidence that rotational modulation of the fainter stars has been missed due to their larger photometric noise. In the presence of such a threshold, the dispersion in the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation---regardless of its origin---makes modulation with smaller \(R_\mathrm{per}\) more likely to be missed, and thus makes the observed slope to appear shallower. Correspondingly, \(\ro_\mathrm{break}\) is inferred to be smaller. To fully understand the impact of the detection bias, we need the complete knowledge of detection function *as well as \(\mathrm{Ro}\) values of all the observed stars with and without \(P_\mathrm{rot}\) detections*, which is impractical. Instead, here we repeat the same analysis as in Section [3.2](#ssec:amp_ro){reference-type="ref" reference="ssec:amp_ro"} for the brightest stars with \(T_\mathrm{eff}=4,000\)--\(6,250\,\mathrm{K}\), using only the stars with the *Kepler* magnitudes \(K_p<12\) for which detection bias appears to be minimal (see also Section [4](#sec:detection){reference-type="ref" reference="sec:detection"}). Due to the correlation between \(K_p\) and \(T_\mathrm{eff}\), the resulting sample is mostly limited to stars with \(T_\mathrm{eff}>5,500\,\mathrm{K}\). The result is shown in Figure [\[fig:r_ro_kp12\]](#fig:r_ro_kp12){reference-type="ref" reference="fig:r_ro_kp12"}; note that the data typically extend down to lower \(R_\mathrm{per}\) than in Figure [\[fig:r_ro\]](#fig:r_ro){reference-type="ref" reference="fig:r_ro"}. From this analysis, we find \(\ro_\mathrm{break} = 0.99^{+0.07}_{-0.08}\) and \({\beta_\mathrm{unsat}}=-4.6\pm0.9\) (mean and \(90\%\) interval). These values differ by \(\sim 0.1\) and \(\sim 1\) from those inferred for stars with \(T_\mathrm{eff}\gtrsim 5,500\,\mathrm{K}\) (Figure [\[fig:params_teff_ro\]](#fig:params_teff_ro){reference-type="ref" reference="fig:params_teff_ro"}), in the directions consistent with what we expect from the detection bias as discussed above. This should be considered as systematics unaccounted for in our analysis in Section [3.1](#ssec:amp_prot){reference-type="ref" reference="ssec:amp_prot"} and Section [3.2](#ssec:amp_ro){reference-type="ref" reference="ssec:amp_ro"}. As will be discussed in Section [4](#sec:detection){reference-type="ref" reference="sec:detection"} further, cooler *Kepler* stars tend to have larger apparent magnitudes than the hotter ones due to their lower intrinsic luminosities. This causes cooler dwarfs to have higher \(R_\mathrm{per}\) thresholds, as is evident in Figure [\[fig:r_prot_all\]](#fig:r_prot_all){reference-type="ref" reference="fig:r_prot_all"} (and Figure [\[fig:pr_teff\]](#fig:pr_teff){reference-type="ref" reference="fig:pr_teff"}). Therefore the above bias is more severe for cooler stars, and the trend we see in Figure [\[fig:params_teff_ro\]](#fig:params_teff_ro){reference-type="ref" reference="fig:params_teff_ro"} for \({\beta_\mathrm{unsat}}\) and \(\ro_\mathrm{break}\) may in part be explained by this \(T_\mathrm{eff}\)-dependent bias. As expected, on the other hand, the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation inferred from the brightest stars remains unchanged at \(\mathrm{Ro} < \ro_\mathrm{break}\). Thus the \(T_\mathrm{eff}\) dependence of \(R_\mathrm{break}\) seems real. This could be due to smaller spot coverage fractions in cooler stars, or smaller spot contrasts, or both.
## The Santos Sample {#ssec:santos}
While this paper mainly focuses on the sample, it is useful to consult other \(P_\mathrm{rot}\) catalogs to better understand the applicability and limitations of the results based on this specific catalog. We thus analyze the catalog by that provided a larger number of \(P_\mathrm{rot}\) measurements than. We applied the same selection as described in Section [2](#sec:sample){reference-type="ref" reference="sec:sample"} to the stars in and found 8,713 (7,621) stars with \(4,000\,\mathrm{K}<T_\mathrm{eff}<6,500\ (6,250)\,\mathrm{K}\). Figure [\[fig:santos\]](#fig:santos){reference-type="ref" reference="fig:santos"} compares the photometric modulation amplitudes and Rossby numbers of the stars in this Santos sample (orange dots) against those in the McQuillan sample (gray circles), where the latter sample is thinned by a factor of 10 to improve the visibility while showing the main trend. Because the modulation amplitudes in the Santos catalog are given using the proxy \(S_\mathrm{ph}\), here the values of \(S_\mathrm{ph}\) are scaled uniformly by 3.6, which is the median of \(R_\mathrm{per}/S_\mathrm{ph}\) for stars in both samples. Despite the simpleness of the conversion, the figure shows that the distributions of modulation amplitudes and \(\mathrm{Ro}\) in the two samples are very similar, except that the Santos sample reports \(P_\mathrm{rot}\) for more stars with smaller amplitudes than in the McQuillan sample. Quantitatively, the median-filtered \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation in the Santos sample (thick orange line) follows more closely to the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) model derived from the \(K_p<12\) stars in the McQuillan sample (black solid line), at least at larger \(\mathrm{Ro}\). This supports our argument on the detection bias in Section [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"}: we argued that the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation based on all the stars (black dashed line) is shallower than that derived from the brightest stars (black solid line) because the former is biased against stars with weaker modulation, and here we do see that the Santos catalog that is apparently less biased against stars with weaker modulation follows the steeper relation. Interestingly, the amplitude in the Santos sample appears to plateau again at \(\mathrm{Ro}\gtrsim \mathrm{Ro}_\odot\). This hints that the modulation amplitude may not keep decreasing in the same way as in \(\mathrm{Ro} < \mathrm{Ro}_\odot\). We note, however, that a more careful assessment of the detection function is required to confirm whether this is a typical behavior or not. As will be discussed in detail below, the measured photometric amplitudes in this region are close to the photometric noise level of *Kepler*. It is thus conceivable that only the highest variability stars at given \(\mathrm{Ro}\) are seen here and/or that the measured amplitudes may be sensitive to how one corrects for the photon noise; although the latter is taken into account in the analysis of, the authors also comment on difficulties associated with small-amplitude modulation. We leave the detailed study of the amplitude--\(\mathrm{Ro}\) relation in this region for a future work. The following discussion is not affected by this ambiguity, because we will show that those stars mostly fall below the detection limit in the sample anyway---unless the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation turns up at larger \(\mathrm{Ro}\).
# Detection Edge in the McQuillan Sample {#sec:detection}
Now we attempt to clarify how the \(R_\mathrm{per}\)--\(P_\mathrm{rot}\) (\(\mathrm{Ro}\)) relation discussed in Section [3](#sec:analysis){reference-type="ref" reference="sec:analysis"}, when combined with the detection bias, sculpts the longer-period edge of the \(P_\mathrm{rot}\)--\(T_\mathrm{eff}\) distribution in the *Kepler* sample. The \(P_\mathrm{rot}\)--\(T_\mathrm{eff}\) distribution of *Kepler* stars has been known to exhibit a rather well-defined upper edge. One perplexing aspect of this upper edge is that it does not correspond to a constant variability amplitude, i.e., the *lower* edge of the \(R_\mathrm{per}\)--\(T_\mathrm{eff}\) distribution is not flat. These features are also apparent in our sample (Figure [\[fig:pr_teff\]](#fig:pr_teff){reference-type="ref" reference="fig:pr_teff"}). pointed out that the upper \(P_\mathrm{rot}\)--\(T_\mathrm{eff}\) edge lies roughly around a gyrochrone of the solar age. pointed out that the upper edge is around \(\mathrm{Ro}\sim 2\) and discussed the possibility that this is related to detection bias. Another explanation they proposed is that the edge is due to stalled spin down: the stars stop spinning down once they reach \(\mathrm{Ro}\sim2\) and stay around the edge. The presence of stalled spin down has now been supported by multiple studies as mentioned in Section [1](#sec:intro){reference-type="ref" reference="sec:intro"}; more recently, the pile-up in the \(P_\mathrm{rot}\)-\(T_\mathrm{eff}\) distribution around its upper edge (also apparent in the top panel of Figure [\[fig:pr_teff\]](#fig:pr_teff){reference-type="ref" reference="fig:pr_teff"}) has been reported and argued to provide a further support for the stalled spin down scenario. Nevertheless, these arguments for the presence of stalled spin down do not necessarily exclude the possibility that the edge in the sample is shaped by detection bias.
## Evidence for the Noise-Dependent Cutoff
Here, we argue that this edge results from the detection threshold set by photometric precision of *Kepler* that depends on apparent magnitudes of stars in the *Kepler* band, \(K_p\). We first note, in Figure [\[fig:rper_kepmag\]](#fig:rper_kepmag){reference-type="ref" reference="fig:rper_kepmag"}, that the distribution of \(R_\mathrm{per}\) and \(K_p\) for all the stars with \(P_\mathrm{rot}\) detection in has a sharp lower edge with a positive slope, whose value at \(K_p\gtrsim 13\) is not far from the scaling for pure photon noise: \(R_\mathrm{per} \propto 10^{K_p/5}\). This indicates that the detectability is limited by photometric precision for those fainter stars, which comprise the majority of the sample. In the top panel of Figure [\[fig:rnorm_teff\]](#fig:rnorm_teff){reference-type="ref" reference="fig:rnorm_teff"}, we show \(R_\mathrm{per}\) normalized by the photometric precision \(\sigma_\mathrm{Kep}\) for long-cadence (29.4 min) exposure of *Kepler* for each star against \(T_\mathrm{eff}\). Here \(\sigma_\mathrm{Kep}\) was evaluated using the photometric precision estimated by the *Kepler* team as a function of \(K_p\),[^5] which takes into account noise sources other than the photon noise and is applicable to \(K_p\gtrsim 12\).[^6] We do not use the Combined Differential Photometric Precision commonly used for evaluating noise levels relevant to planet search, because here we need to evaluate the noise that does not include intrinsic stellar variabilities. In this plane, the lower edge of the \(R_\mathrm{per}/\sigma_\mathrm{Kep}\) distribution is flat across \(T_\mathrm{eff}\), again indicating that the sample is limited by photometric precision: the lower edge in the \(R_\mathrm{per}\)--\(T_\mathrm{eff}\) distribution (Figure [\[fig:pr_teff\]](#fig:pr_teff){reference-type="ref" reference="fig:pr_teff"}, bottom panel) is higher for cooler stars because they tend to be apparently (and intrinsically) fainter than the hotter ones. The lower edge of \(R_\mathrm{per}/\sigma_\mathrm{Kep}\) is also flat as a function of \(P_\mathrm{rot}\) and \(\mathrm{Ro}\), as shown in the middle and bottom panels of Figure [\[fig:rnorm_teff\]](#fig:rnorm_teff){reference-type="ref" reference="fig:rnorm_teff"}. The histogram of \(R_\mathrm{per}/\sigma_\mathrm{Kep}\) (Figure [\[fig:rnorm_hist\]](#fig:rnorm_hist){reference-type="ref" reference="fig:rnorm_hist"}) shows a sharp cutoff around 3 (bottom \(\approx2\%\)), which we adopt as an empirical detection threshold of the sample to guide the present discussion. This value, shown as the tan horizontal dashed line, agrees visually with the lower edge of the distributions in Figure [\[fig:rnorm_teff\]](#fig:rnorm_teff){reference-type="ref" reference="fig:rnorm_teff"}, and also well explains the difference between the \(R_\mathrm{per}\) distribution of stars with \(T_\mathrm{eff}=4,000\)--\(6,500\,\mathrm{K}\) in the sample and of stars with \(K_p<12\). We note that this threshold value is specific to the sample, as well as to the timescale for which photometric precision is defined. We suspect that this rather steep cutoff is associated with the threshold on the weight parameter \(w\) that was used by to distinguish between periodic and false detections. This parameter is related to the local peak height of the autocorrelation function that would explicitly depend on the noise level. We also note again that Figure [\[fig:rper_kepmag\]](#fig:rper_kepmag){reference-type="ref" reference="fig:rper_kepmag"} shows *all* the stars with significant \(P_\mathrm{rot}\) detections in; thus the lower edge is not due to our sample selection.
## Rossby Number Cutoff
Given the presence of the detection edge, the next question is what value of \(\mathrm{Ro}\) (or \(P_\mathrm{rot}\)) this edge corresponds to---and we find the detection edge should correspond to \(\mathrm{Ro}\sim 1\)--\(2\) in the scale, or \(\mathrm{Ro} \sim 0.5\)--\(1\,\mathrm{Ro}_\odot\). This value is derived by equating the roughly \(T_\mathrm{eff}\)-independent \(R_\mathrm{per}(\mathrm{Ro})\) derived in Sections [3.2](#ssec:amp_ro){reference-type="ref" reference="ssec:amp_ro"} and [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"} with \(3\sigma_\mathrm{Kep}(K_p)\) of each star and by solving for \(\mathrm{Ro}_\mathrm{threshold}(K_p)\): rotational modulation is detectable for a given star with the magnitude \(K_p\) if its \(\mathrm{Ro}\) is lower than \(\mathrm{Ro}_\mathrm{threshold}(K_p)\). The distribution of \(\mathrm{Ro}_\mathrm{threshold}(K_p)\) computed in this way for all the stars for which rotational modulation has been *searched* by (i.e., stars with *and without* \(P_\mathrm{rot}\) detection) is shown in the top panel of Figure [\[fig:roth_hist\]](#fig:roth_hist){reference-type="ref" reference="fig:roth_hist"}. The solid histogram shows the result based on the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation derived in Section [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"} using the brightest (\(K_p<12\)) stars with \(T_\mathrm{eff}=4,000\)--\(6,250\,\mathrm{K}\), which is likely less affected by detection bias and more representative (see also Section [3.4](#ssec:santos){reference-type="ref" reference="ssec:santos"}); the dashed histogram shows the result based on the relation derived using all the sample stars (Section [3.2](#ssec:amp_ro){reference-type="ref" reference="ssec:amp_ro"}), which may be more appropriate for cooler stars. Both distributions are sharply peaked around \(\mathrm{Ro}\approx 1.2 \sim 0.6\,\mathrm{Ro}_\odot\), which corresponds to the faintest (and hence most abundant) stars with \(K_p\sim16\). The peak is also narrow because of the strong \(P_\mathrm{rot}\) dependence of \(R_\mathrm{per}\): for \(R_\mathrm{per} \sim \mathrm{Ro}^{-3}\)--\(\mathrm{Ro}^{-4}\), \(\mathrm{Ro}_\mathrm{threshold}(K_p)\) increases only by a factor of two for the \(K_p\) difference of five. By definition of \(\mathrm{Ro}_\mathrm{threshold}\), its normalized inverse cumulative distribution function, shown in the bottom panel of Figure [\[fig:roth_hist\]](#fig:roth_hist){reference-type="ref" reference="fig:roth_hist"}, provides the detectability function \(p_\mathrm{det}(\mathrm{Ro})\), the fraction of stars in the searched sample for which rotational modulation of a star with a given value of \(\mathrm{Ro}\) would have been reported as a robust detection by. We see that faint stars start to be missed at \(\mathrm{Ro}\gtrsim1.2\), and that the detection becomes impossible for almost all stars at \(\mathrm{Ro}\sim 2 \approx \mathrm{Ro}_\odot\) for the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation derived from the brightest stars (solid line), which is likely more representative than that from all stars (dashed line). This is how the combination of the rapid drop of \(R_\mathrm{per}\) with increasing \(\mathrm{Ro}\) and roughly magnitude-limited nature of the *Kepler* sample imprints the detection edge around \(\mathrm{Ro}\sim 2 \approx \mathrm{Ro}_\odot\); the sample becomes roughly \(\mathrm{Ro}\)-limited, and the longest detected \(P_\mathrm{rot}\) increases with decreasing \(T_\mathrm{eff}\) roughly as \(\tau_\mathrm{c}\). This agrees with what is observed, that the upper edge is close to a curve of \(\mathrm{Ro}=1.7\) (blue dashed line in the top panel of Figure [\[fig:pr_teff\]](#fig:pr_teff){reference-type="ref" reference="fig:pr_teff"}) in our \(\tau_\mathrm{c}\) scale. It is beyond the scope of this work to understand this value more quantitatively: the threshold depends on the unknown distribution of \(\mathrm{Ro}\) in all the searched stars (i.e., stars with *and without* detected \(P_\mathrm{rot}\)) as well as on exact dependence of the detectability on the signal to noise, both of which need to be modeled. It is also very sensitive to the steepness of the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation, as shown in Figure [\[fig:roth_hist\]](#fig:roth_hist){reference-type="ref" reference="fig:roth_hist"}, as well as on its possible dependence on \(T_\mathrm{eff}\) which is difficult to assess in the current sample (Section [3.3](#ssec:r_ro_bias){reference-type="ref" reference="ssec:r_ro_bias"} and Appendix [\[ssec:tauc_amp\]](#ssec:tauc_amp){reference-type="ref" reference="ssec:tauc_amp"}). The above argument, along with external evidence for stalled spin down, suggests that both effects discussed by are important in understanding the \(P_\mathrm{rot}\)--\(T_\mathrm{eff}\) distribution around its upper edge. Around the time when typical Sun-like stars observed by *Kepler* cease to spin down at \(\mathrm{Ro}\sim \mathrm{Ro}_\odot\), their rotational modulation signals have already started to be buried under photometric noise and to be missed from the sample with \(P_\mathrm{rot}\) detection. Therefore, the pile-up we see in the observed \(P_\mathrm{rot}\)--\(T_\mathrm{eff}\) distribution may be just a tip of the iceberg: the true pile-up may be located at longer periods, but have been capped due to the detection edge. This interpretation is qualitatively consistent with the finding of and, who worked on the \(P_\mathrm{rot}\) sample much less biased against slower rotators, and found that most stars in the sample are around *or above* the upper edge defined by stars with \(P_\mathrm{rot}\) from rotational modulation.
## Impact on the Observed \(\mathrm{Ro}\) Distribution {#ssec:romodels}
To demonstrate the impact of the detection edge further, we use simple models to show how the \(p_\mathrm{det}(\mathrm{Ro})\) in the bottom panel of Figure [\[fig:roth_hist\]](#fig:roth_hist){reference-type="ref" reference="fig:roth_hist"} works to bias the observed \(\mathrm{Ro}\) distribution. In the top panel of Figure [\[fig:detmodel\]](#fig:detmodel){reference-type="ref" reference="fig:detmodel"}, we adopt \(p_\mathrm{det}(\mathrm{Ro})\) derived from all the stars (dashed line in Figure [\[fig:roth_hist\]](#fig:roth_hist){reference-type="ref" reference="fig:roth_hist"}) and simulate how this modifies the true \(\mathrm{Ro}\) distribution, here chosen to be a Gaussian with the mean of 2 and scale of \(0.5\) (thick gray line): the result is the orange histogram. The location of the observed peak (\(\mathrm{Ro}\sim 1.5\)) has shifted from the truth due to a rapid decrease of \(p_\mathrm{det}\) at \(\mathrm{Ro} \gtrsim 1\). The same is true but the bias is more severe when we adopt \(p_\mathrm{det}(\mathrm{Ro})\) derived from the stars with \(K_p<12\) (solid line in Figure [\[fig:roth_hist\]](#fig:roth_hist){reference-type="ref" reference="fig:roth_hist"}), as shown in the middle panel. The difference from the top panel also illustrates how the resulting distribution is sensitive to the steepness of the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation above \(\ro_\mathrm{break}\). In the bottom panel, \(p_\mathrm{det}(\mathrm{Ro})\) is again from the stars with \(K_p<12\), and the true \(\mathrm{Ro}\) distribution is constructed by sampling age \(t\) from the uniform distribution between \(0\) and \(10\,\mathrm{Gyr}\), translating \(t\) into \(P_\mathrm{rot} = 25\,\mathrm{days}\,(t/4.6\,\mathrm{Gyr})^{1/2}\) and to \(\mathrm{Ro}=P_\mathrm{rot}/\tau_\mathrm{c}(5,777\,\mathrm{K})\) using the formula in: this simulates a collection of Sun-like stars that (i) have a uniform age distribution, (ii) have the same \(K_p\) distribution as the *Kepler* stars, and (iii) keep spinning down following Skumanich's law during their entire main-sequence life. This is merely another toy model but provides two useful insights. First, the sharp decrease of \(p_\mathrm{det}\) at \(\mathrm{Ro}\gtrsim 1\) produces a peak in the observed distribution (orange thin histogram) as long as the true \(\mathrm{Ro}\) distribution keeps increasing across the threshold \(\mathrm{Ro}\), even without stalled spin down. Second, the fraction of stars with detectable rotational modulation (i.e., mean value of \(p_\mathrm{det}(\mathrm{Ro})\) in the sample) is computed to be 0.28, which is close to the observed value: a combination of the sharp detection edge and the top-heavy \(\mathrm{Ro}\) distribution provides a reasonable explanation for why \(P_\mathrm{rot}\) has not been detected for the majority of solar-mass stars. This experiment suggests that stalled spin down, if real, should start operating at \(\mathrm{Ro} \gtrsim 2\) so that a significant fraction of solar-mass stars evade detection of rotaitonal modulation. It may also explain why the pile-up found by corresponds to a lower \(\mathrm{Ro}\sim 1.5\) than that inferred from asteroseismology or \(v\sin i\) (\(\mathrm{Ro}\sim 2\)). The last model discussed here might even suggest that the pile-up of \(\mathrm{Ro}\) alone does not serve as a conclusive evidence for the stalled spin down, as it shows that the observed distribution is not very sensitive to the \(\mathrm{Ro}\) distribution above the detection edge. We note that this argument does not deny the importance of possible systematic offsets in \(T_\mathrm{eff}\) as discussed by. Nevertheless, these toy models demonstrate the importance of considering detection bias in interpreting the observed \(P_\mathrm{rot}\) distribution.
## Comparison with the Santos Sample {#ssec:santos_noise}
We saw in Section [3.4](#ssec:santos){reference-type="ref" reference="ssec:santos"} that the \(P_\mathrm{rot}\) catalog by includes more stars with smaller modulation amplitudes than in. This suggests that the Santos sample is subject to a very different detection function from McQuillan's. This situation is shown in Figure [\[fig:santos_noise\]](#fig:santos_noise){reference-type="ref" reference="fig:santos_noise"}; here we reproduce Figure [\[fig:rnorm_hist\]](#fig:rnorm_hist){reference-type="ref" reference="fig:rnorm_hist"} and Figure [\[fig:rper_kepmag\]](#fig:rper_kepmag){reference-type="ref" reference="fig:rper_kepmag"} for the Santos sample, where \(S_\mathrm{ph}\) is used instead of \(R_\mathrm{per}\), and the empirical threshold in the *McQuillan* sample is shown in the scale of \(S_\mathrm{ph}\) (see Section [3.4](#ssec:santos){reference-type="ref" reference="ssec:santos"}). Although the top panel does show the decrease in the detection rate at a signal-to-noise corresponding to the McQuillan edge, we do not find such a sharply defined threshold as seen in the McQuillan sample. Correspondingly, we do not see a well-defined lower edge in the amplitude--magnitude plane in the bottom panel. We also find that the number of stars "leaking\" below the McQuillan threshold increase toward higher \(T_\mathrm{eff}\). This comparison illustrates the importance of considering detection functions in a sample-specific manner. While the Santos sample does include more period detections than in McQuillan's, the detection function might also be more difficult to quantify. Although it is beyond the scope of this work to fully assess the impact of detection bias in the Santos sample, the following arguments suggest that it is likely significant in the Santos sample too, at least for nearly solar-mass stars. First, the detection fraction of \(P_\mathrm{rot}\) is still \(\approx 30\%\) for G stars even in. Second, the steep \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation implies that the longest detectable \(P_\mathrm{rot}\) (or largest detectable \(\mathrm{Ro}\)) is not drastically changed by improving the detection threshold: for \(R_\mathrm{per} \sim \mathrm{Ro}^{-4.6}\) derived from the brightest stars, a factor of 10 improvement in the detection threshold results in an only \(\approx 60\%\) increase in the threshold \(\mathrm{Ro}\). This might explain why the detection fraction did not drastically increase even in the updated analysis of. Therefore, similar features seen in both samples with different detection functions do not necessarily indicate their astrophysical origins.
# Discussion {#sec:discussion}
## Comparison with Other Activity Indicators {#ssec:comparison}
The \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation presented in Section [3](#sec:analysis){reference-type="ref" reference="sec:analysis"} is reminiscent of the relation known for X-ray luminosities normalized by the bolometric values \(L_X/L_\mathrm{bol}\): \(L_X/L_{\rm bol}\) plateaus at \(\mathrm{Ro}\lesssim 0.1\), and decays as \(L_X/L_{\rm bol} \sim \mathrm{Ro}^{-2.7\pm 0.13}\) at least up to \(\mathrm{Ro} \sim 2\). The analysis of is based on the \(\tau_\mathrm{c}\) scale from that is close to what we have adopted, and so the saturation of spot-modulation amplitude at \(\mathrm{Ro} \sim 0.8\) occurs within the so-called unsaturated regime of \(L_X/L_\mathrm{bol}\), where it exhibits a power-law decay. Does this mean the X-ray activity and spot-modulation amplitude evolve differently as a function of \(\mathrm{Ro}\) despite their presumably common origin? The analysis of assumes a two-piece power law and has captured a transition at \(\mathrm{Ro} \sim 0.1\), which therefore is insensitive to finer structures at larger \(\mathrm{Ro}\). Thus here we seek for evidence of another transition in the "unsaturated\" X-ray regime. Figure [\[fig:comparison\]](#fig:comparison){reference-type="ref" reference="fig:comparison"} shows the \(\log L_X/L_\mathrm{bol}\)--\(\mathrm{Ro}\) data (gray open circles) from along with \(\logR_\mathrm{per}\)--\(\mathrm{Ro}\) data in our sample (orange dots), where the scale of \(R_\mathrm{per}\) is shifted arbitrarily but that of \(\mathrm{Ro}\) is not. Here we recomputed \(\mathrm{Ro}\) in the sample using their \(T_\mathrm{eff}\) and the relation so that the comparison can be made using the same \(\tau_\mathrm{c}\) scale. We see that the two data in fact follow the same pattern at \(\mathrm{Ro}\gtrsim0.3\) including a possible wiggle mentioned in Section [3](#sec:analysis){reference-type="ref" reference="sec:analysis"}: there is a hint of a "shoulder\" in the \(L_X/L_\mathrm{bol}\) data beginning around \(\ro_\mathrm{break}\) inferred from photometric modulation, which is also apparent in Figure 7 of. The structure is more clearly seen in the median-filtered data with the width of 0.06 dex (solid gray line), i.e., a representation that does not assume a single power-law relation in this \(\mathrm{Ro}\) range. This reinforces the physical connection between surface spots and coronal X-ray emission, and suggests that the structure in the unsaturated X-ray regime is not an artifact; remember that \(R_\mathrm{per}\)--\(P_\mathrm{rot}\) relation shows a kink regardless of the prescription to compute \(\tau_\mathrm{c}\). A break at a similar value of \(\mathrm{Ro}\) in the chromospheric activities has been noted, which attributed to a transition of dominant dynamo regimes. Although reported a break at a lower value of \(\mathrm{Ro}\) than \(\ro_\mathrm{break}\), the location agrees with what has been inferred from \(R_\mathrm{per}\) (and X-rays) if the common \(\tau_\mathrm{c}\) scale is adopted, as is also confirmed by their conversion that the threshold is \(\mathrm{Ro}=0.91\) in the Noyes scale. The data for main-sequence stars from, *Kepler* asteroseismic stars from, along with the overlapping sample from and for which both \(P_\mathrm{rot}\) and \(R'_\mathrm{HK}\) are readily available, are plotted with green open squares in Figure [\[fig:comparison\]](#fig:comparison){reference-type="ref" reference="fig:comparison"}, which shows a kink at \(\log R'_\mathrm{HK}\sim-4.5\) reported by. Here again we use the formula to recompute \(\tau_\mathrm{c}\) for those stars, where \(T_\mathrm{eff}\) is estimated from \(B-V\) using the table from. We also note that the similarity between the chromospheric and X-ray fluxes has been noted by. A transition at a similar \(\mathrm{Ro}\) might also been seen, though less clearly, in the photospheric filling factor \(f_*\) of the magnetic flux. Note again that here we are focusing on the region around \(\mathrm{Ro} \sim 1\), rather than the saturation similar to that in X-ray around \(\mathrm{Ro} \sim 0.1\). The measurements for GKM stars presented in ---along with the Sun---shows that \(f_*\) decreases by roughly two orders of magnitudes between \(\mathrm{Ro}\sim0.2\) and \(2\). Although the data do not densely cover \(\mathrm{Ro}\) around \(\ro_\mathrm{break}\) seen in the spot-modulation amplitude, the empirical scaling they found, \(f_*\propto \mathrm{Ro}^{-2.5}\) to \(f_*\propto \mathrm{Ro}^{-3.4}\), is roughly in agreement with what we found for \(R_\mathrm{per}\). also found a hint of a similar trend in the H\(\alpha\) luminosity of M dwarfs studied by. These data are not shown in Figure [\[fig:comparison\]](#fig:comparison){reference-type="ref" reference="fig:comparison"}, because the transition features are visually less clear. In summary, coronal and chromospheric fluxes (and perhaps magnetic and H\(\alpha\) fluxes as well) show transitions at \(\mathrm{Ro}\) similar to \(\ro_\mathrm{break}\sim 0.4\,\mathrm{Ro}_\odot\) found for the photometric modulation amplitude, thus suggesting that they share the same physical origin. We also confirmed that the same remains to be the case when the \(\tau_\mathrm{c}\) prescription from is adopted instead; see Figure [\[fig:comparison_l21\]](#fig:comparison_l21){reference-type="ref" reference="fig:comparison_l21"} in Appendix [\[ssec:tauc_comp\]](#ssec:tauc_comp){reference-type="ref" reference="ssec:tauc_comp"}. Although our sample does not constrain the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation at \(\mathrm{Ro}\lesssim 0.2\,\mathrm{Ro}_\odot\), other ground-and space-based photometry works generally show even larger spot-modulation amplitudes up to \(\sim 10\%\) for those younger stars with shorter rotation periods. Thus the evolution at lower \(\mathrm{Ro}\) may also be similar to the X-ray and chromospheric fluxes.
## Implications for Weakened Magnetic Braking {#ssec:wmb}
proposed that magnetic braking ceases at a critical Rossby number of \(\mathrm{Ro}_\mathrm{WMB} \sim \mathrm{Ro}_\odot\) based on comparison between their spin evolution models and the age/rotation measurements for \(\sim 20\) stars. As we saw in Section [5.1](#ssec:comparison){reference-type="ref" reference="ssec:comparison"}, the information on how various activity indicators evolve around \(\mathrm{Ro}\sim\mathrm{Ro}_\odot\) is in general limited (Figure [\[fig:comparison\]](#fig:comparison){reference-type="ref" reference="fig:comparison"}), but some indicators may be showing hints of corresponding changes. A transition at \(\mathrm{Ro}\sim\mathrm{Ro}_\odot\) has been suggested in the chrmospheric fluxes. A small number of measurements in the Santos sample hints that \(R_\mathrm{per}\) might also follow a similar pattern at \(\mathrm{Ro} \gtrsim \mathrm{Ro}_\odot\) (Section [3.4](#ssec:santos){reference-type="ref" reference="ssec:santos"}), although a more careful analysis would be required to confirm weather this is a typical behavior or not, because the detection bias is significant here (Section [4](#sec:detection){reference-type="ref" reference="sec:detection"}). On the other hand, all the indicators show that the activity pattern changes in a continuous but non-monotonic manner up to \(\mathrm{Ro}_\odot\). Thus it also seems conceivable that the departure from the standard spin evolution starts earlier than \(\mathrm{Ro}_\odot\) and proceeds gradually. In particular, the decrease in \(R_\mathrm{per}\) at \(\mathrm{Ro}\gtrsim 0.5\,\mathrm{Ro}_\odot\) may indicate that large spots suddenly start dissolving into smaller pieces. If so, this seems qualitatively consistent with a scenario that the concentration of the magnetic fields into smaller spatial scales and the associated reduction of angular momentum loss is responsible for weakened magnetic braking. We note, though, that the relation between the photometric light curves and spot distribution is generally very complicated and that the modulation amplitude may not be readily translated into the largest spot size. More in-depth analyses of the light curve morphology as a function of \(\mathrm{Ro}\) may bring this hypothesis into sharper focus. Regardless of whether or not the pattern in the \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation we derived is physically related to weakened magnetic braking, our finding has important implications for studies of weakened magnetic braking using rotation periods from photometric modulation. We presented evidence that detection bias becomes particularly important in the relevant \(\mathrm{Ro}\) range (Section [4](#sec:detection){reference-type="ref" reference="sec:detection"}). The subtlety arises from the fact that the \(\mathrm{Ro}\) corresponding to detection edge, \(\mathrm{Ro}_\mathrm{edge}\), happens to be close to \(\mathrm{Ro}_\mathrm{WMB}\). This is in part a coincidence, because \(\mathrm{Ro}_\mathrm{edge}\) is determined by photometric precision of *Kepler*. On the other hand, it is also true that the strong dependence makes \(\mathrm{Ro}_\mathrm{edge}\) insensitive to photometric precision---so if \(\ro_\mathrm{break} \sim \mathrm{Ro}_\mathrm{WMB}\) due to physics, it is inevitable that \(\mathrm{Ro}_\mathrm{edge}\sim \mathrm{Ro}_\mathrm{WMB}\). Given \(R_\mathrm{per} \sim \mathrm{Ro}^{-4}\) found for solar-mass stars, the photometric precision needs to be improved by an order of magnitude to push \(\mathrm{Ro}_\mathrm{edge}\) up by a factor of two. This argument explains why it has not been easy to find the signature of the weakened magnetic braking in the photometric sample, and indicates that it is crucial to consider detection bias when interpreting the sample quantitatively in terms of weakened magnetic braking scenario.
## Further Test of the View
We argued in Section [4](#sec:detection){reference-type="ref" reference="sec:detection"} that the detectability of photometric rotational modulation in *Kepler* stars is determined by the combination of the (roughly) \(T_\mathrm{eff}\)-independent steep \(R_\mathrm{per}\)--\(\mathrm{Ro}\) relation and magnitude-dependent detection threshold. If this is correct, photometric rotational modulation should have been detected for Sun-like main-sequence stars *if and only if* a star is younger than a certain \(T_\mathrm{eff}\)-dependent threshold age. In the other extreme case where the upper edge in the \(P_\mathrm{rot}\)--\(T_\mathrm{eff}\) distribution is solely due to stalled spin down, on the other hand, stars with photometrically detected \(P_\mathrm{rot}\) around the edge should contain many stars older than the age corresponding to the onset of the stalled spin down. We investigate this hypothesis in a companion paper (Masuda 2022, in prep.) using the isochronal age estimates for a sample of *Kepler* stars with *and without* detected rotational modulation.
# {#section .unnumbered}
The data and the code underlying this article are available through GitHub.[^7]
The author thanks Shinsuke Takasao for helpful conversations on the subject of this paper. This work was supported by JSPS KAKENHI Grant Number 21K13980. This work has made use of data from the LAMOST project. Guoshoujing Telescope (the Large Sky Area Multi-Object Fiber Spectroscopic Telescope LAMOST) is a National Major Scientific Project built by the Chinese Academy of Sciences. Funding for the project has been provided by the National Development and Reform Commission. LAMOST is operated and managed by the National Astronomical Observatories, Chinese Academy of Sciences. This work has made use of data from the European Space Agency (ESA) mission *Gaia* (<https://www.cosmos.esa.int/gaia>), processed by the *Gaia* Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the *Gaia* Multilateral Agreement. Funding for the Kepler mission is provided by the NASA Science Mission Directorate. STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5--26555.
[^1]: We used *Kepler* IDs in "tcomment\" columns in the latter catalog.
[^2]: Scatters of \(T_\mathrm{eff}\) from multi-epoch observations are typically smaller than the \(\sim 100\,\mathrm{K}\) uncertainty estimated for dwarfs. The bin size in Section [3](#sec:analysis){reference-type="ref" reference="sec:analysis"} is chosen to be larger than this latter value.
[^3]: The same appears to be the case for the analysis by, according to their Figure 3.
[^4]: The hottest bin does not follow this trend, but in Figure [\[fig:r_ro_all\]](#fig:r_ro_all){reference-type="ref" reference="fig:r_ro_all"} the data at \(\mathrm{Ro}\gtrsim 1\) are visually consistent with a steeper slope, which corresponds to larger \(\ro_\mathrm{break}\). We suspect that the current result might be biased due to a larger fraction of outliers at low and high \(\mathrm{Ro}\).
[^5]: <https://nexsci.caltech.edu/workshop/2012/keplergo/CalibrationSN.shtml>
[^6]: The \(K_p\) dependence also agrees with the estimates by using the Quarter 1 data, at least in the fainter end that is relevant to our discussion.
[^7]: <https://github.com/kemasuda/acheron/tree/main/kepler_prot_teff> | {'timestamp': '2022-06-07T02:28:30', 'yymm': '2206', 'arxiv_id': '2206.01595', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01595'} |
null | null |
# Introduction
Learning graph representations for relational data structures has been gaining increasing attention in the machine learning community . A graph is able to model real-world, complex systems by representing entities as nodes and interactions between them as edges. Since the information in graphs is often incomplete, e.g., missing node attributes or edges, relevant graph-related tasks for attaining new knowledge include node classification and link prediction. A variety of graph neural network (GNN) models have been developed , which learn node and edge embeddings in a low-dimensional vector space. Subsequently, these embeddings can be used to solve downstream tasks like node classification. Usually, the focus here lies on maximizing the accuracy--the proportion of nodes that are classified correctly. GNNs achieve good performance with respect to accuracy but are also black boxes and lack interpretability. Most machine learning models output confidence scores associated with the predictions, and the concept of calibration captures the idea that the score should reflect the ground-truth probability of the prediction's correctness. For example, if 100 instances have a score of 0.6 for a specific class \(k\), then 60 instances are expected to actually be of class \(k\). A real-world application is autonomous driving, where the model should not only be aware that the object in front of the car is more likely to be a plastic bag than a pedestrian but also know how much more likely it is. A score distribution of 0.99 for plastic bag and 0.01 for pedestrian or 0.51 for plastic bag and 0.49 for pedestrian could have a huge influence on the next action of the car. Generally, calibrated scores lead to a better interpretation of the results and increase the trustworthiness of machine learning models, which is especially important in safety-critical domains. The calibration of deep neural networks has been addressed in several works . The calibration of GNNs, however, has not been sufficiently explored yet, and existing calibration methods do not exploit the graph structure. Due to the different architectures of GNNs compared to neural networks, GNNs might exhibit different calibration characteristics. In this work, we are interested in the following research questions: **R1.** How are GNNs calibrated for the node classification task, and are existing calibration methods sufficient to calibrate GNNs? **R2.** How do model capacity (width and depth) and graph density influence the calibration? **R3.** Can a calibration error term be added to the loss function in a straightforward way to improve the calibration without hurting the accuracy? **R4.** Can the incorporation of topological information improve calibration? To better understand the calibration properties of GNNs, we conduct an empirical analysis of several GNN models in a node classification setting. Based on our experimental finding that the nodes in the graph express different levels of over-and underconfidence, we propose a topology-aware calibration method that takes the neighboring nodes into account. Our contributions are summarized as follows:
- We inspect the calibration of five representative GNN models on three benchmark citation datasets for node classification.
- We analyze the influence of model capacity, graph density, and a new loss function on the calibration of GNNs.
- We propose a calibration method that takes the graph topology into account and yields improved calibration compared to state-of-the-art post-processing calibration methods.
In Section [2](#sec:background){reference-type="ref" reference="sec:background"}, we define the necessary concepts and summarize related work. The existing GNNs and calibration methods used in this work are also described briefly. An experimental study on the calibration of GNNs is presented in Section [3](#sec:experiments){reference-type="ref" reference="sec:experiments"} (\(\rightarrow\) **R1**, **R2**, **R3**). In Section [4](#sec:ratio){reference-type="ref" reference="sec:ratio"} (\(\rightarrow\) **R1**, **R4**), we propose a topology-aware calibration method and show experimental results compared to state-of-the-art calibration baselines. The results are discussed in Section [5](#sec:discussion){reference-type="ref" reference="sec:discussion"}.
# Background {#sec:background}
## Definitions
### Node classification on graphs
An undirected graph is defined as \(G = (\mathcal{V} ,\mathcal{E})\), where \(\mathcal{V}\) is the set of nodes and \(\mathcal{E}\) the set of edges. An edge \(e = \{i,j\} \in \mathcal{E}\) connects the two nodes \(i\) and \(j\) in the graph. The information about the edges can be encoded in an adjacency matrix \(\mathbf{A} \in \left \{0, 1\right \}^{\left|\mathcal{V}\right| \times \left|\mathcal{V}\right|}\). With \(\mathbf{A}_{ij}\) being the entry in the \(i\)-th row and \(j\)-th column of \(\mathbf{A}\), we define \(\mathbf{A}_{ij} = 1\) if \(\{i,j\} \in \mathcal{E}\) and \(\mathbf{A}_{ij} = 0\) otherwise[^1]. Moreover, we define \(\mathcal{N}(i)\) as the set of neighbors of node \(i\). For attributed graphs, where each node \(i\) is associated with a \(d\)-dimensional feature vector \(\mathbf{X}_i \in \mathbb{R}^d\), we denote the feature matrix by \(\mathbf{X} \in \mathbb{R}^{|\mathcal{V}| \times d}\). The goal of the node classification task is to assign each node \(i \in \mathcal{V}\) a class label \(\hat{y}_i \in \mathcal{K} := \{1, 2, \dots, K\}\), where \(K\) stands for the total number of classes.
### Calibration
Let \(\mathbf{H}_i \in \mathbb{R}^h\) denote the node embedding and \(y_i \in \mathcal{K}\) the ground-truth label of sample (or node) \(i\in \mathcal{V}\). Let \(g: \mathbb{R}^h \rightarrow [0,1]^K\) be a function that takes \(\mathbf{H}_i\) as input and outputs a probability vector \(g(\mathbf{H}_i)\), where \(g(\mathbf{H}_i)_k\) represents the \(k\)-th element. The predicted class label for sample \(i\) is given by \(\hat{y}_i =\mathop{\arg\max}_{k \in \mathcal{K}}g(\mathbf{H}_i)_k\), where \(\hat{p}_i =\mathop{\max}_{k \in \mathcal{K}}g(\mathbf{H}_i)_k\) is called the corresponding confidence score for \(\hat{y}_i\). Perfect calibration is defined as \(\mathbb{P}(\hat{y}_i=y_i \mid \hat{p}_i=p)=p\) for all \(p \in [0,1]\) and any sample \(i\) . A reliability diagram plots accuracy against confidence to visualize the calibration of the model (see Fig. [\[fig:rds\]](#fig:rds){reference-type="ref" reference="fig:rds"}). More formally, the samples are grouped into \(M \in \mathbb{N}\) equally-spaced interval bins according to their confidences \(\hat{p}_i\). For each bin \(B_m\), \(m \in \{1,2, \dots, M\}\), the accuracy and average confidence are calculated according to \[\begin{aligned}
\mathrm{acc}(B_m) &= \frac{1}{|B_m|}\sum_{i \in B_m}\mathbbm{1}[\hat{y}_i=y_i] \quad \mathrm{and} \\ \quad \mathrm{conf}(B_m) &= \frac{1}{|B_m|}\sum_{i \in B_m}\hat{p}_i\;, \end{aligned}\] respectively, where \(|B_m|\) denotes the number of samples in bin \(B_m\) and \(\mathbbm{1}\) the indicator function. In case of perfect calibration, the equation \(\mathrm{acc}(B_m) = \mathrm{conf}(B_m)\) holds for all \(m\). Reliability diagrams also present a way to identify if the model is over-or underconfident. If the bars are above the diagonal line, it implies that the accuracy is higher than the average confidence, and the model is called underconfident. If the bars are below the diagonal, the model is overconfident. The expected calibration error (ECE) measures the miscalibration by averaging the gaps in the reliability diagram and is given by \[\sum_{m=1}^{M}\frac{|B_m|}{N}|\mathrm{acc}(B_m)-\mathrm{conf}(B_m)|\;, \label{eq:ece}\] where \(N\) is the total number of samples. The marginal ECE (MECE) approximates the marginal calibration error , which takes all classes into account. For each bin and every class \(k\), it compares the average confidence of samples for class \(k\) to the proportion of samples that has as ground-truth label class \(k\). The MECE is defined as \[\sum_{k=1}^K w_k \sum_{m=1}^{M} \frac{1}{N}\left|\sum_{i \in B_m}\mathbbm{1}[y_i=k]-\sum_{i \in B_m}g(\mathbf{H}_i)_k\right|\;, \label{eq:mece}\] where \(w_k\) is a class-dependent weight factor, which is set to \(1/K\) if all classes are equally important.
## Related work
Guo et al. showed that modern neural networks are miscalibrated and tend to be overconfident, i. e., the confidence scores are higher than the proportions of correct predictions. They proposed temperature scaling, a single-parameter variant of Platt scaling , to calibrate the results. Several other methods were introduced to improve the calibration of deep neural networks (e. g., mixup training and FALCON ). Methods that improve calibration by preventing overconfidence include label smoothing and focal loss . In GNNs, calibration issues have only been studied recently. A first evaluation of GNNs was done by Teixeira et al. , who performed experiments on multiple node classification datasets and concluded that GNNs are miscalibrated and existing calibration methods are not always able to improve the calibration to the desired extent.
## Methods
### Graph neural networks
Given an adjacency matrix \(\mathbf{A}\) and a feature matrix \(\mathbf{X}\), the idea of all GNNs is to learn node embeddings \(\mathbf{H} \in \mathbb{R}^{|\mathcal{V}| \times h}\). The embedding for node \(i\) is denoted by \(\mathbf{H}_i \in \mathbb{R}^h\), which can be fed to a task-specific decoder \(g\). For example, since we are concerned with node classification, we use a single-layer perceptron with softmax activation as decoder. For our experiments, we select the widely used models graph convolutional network (GCN) , graph attention network (GAT) , and simple graph convolution (SGC) . Further, we consider graph filter neural network (gfNN) , a straightforward extension of SGC, and approximate personalized propagation of neural predictions (APPNP) , a model with state-of-the-art performance. *GCN* applies a normalized adjacency matrix with self-loops \(\tilde{\mathbf{A}} = \hat{\mathbf{D}}^{-\frac{1}{2}} (\mathbf{A}+\mathbf{I}) \hat{\mathbf{D}}^{-\frac{1}{2}}\), where \(\mathbf{I}\) is the identity matrix and \(\hat{\mathbf{D}}\) the degree matrix of \(\mathbf{A}+\mathbf{I}\). Concretely, the hidden layer of a GCN is formed according to \(\mathbf{H}^{(l+1)} = \sigma(\tilde{\mathbf{A}}\mathbf{H}^{(l)}\mathbf{W}^{(l)}),\) where \({\mathbf{W}^{(l)}}\) is a trainable weight matrix, \(\sigma\) an activation function, and \(\mathbf{H}^{(0)} := \mathbf{X}\). GCN aggregates information from a node's neighbors by computing the normalized sum of adjacent node embeddings. *GAT* differs from GCN in the neighbor aggregation function by introducing an attention mechanism that scales the importance of neighbors when summing over their embeddings. For each node \(i\) and neighbor \(j \in \mathcal{N}(i)\), the attention \(\alpha_{ij}\) is computed as \[\alpha_{ij} = \frac{\exp(\mathrm{LeakyReLU}(\mathbf{a}^{T}[\mathbf{W}\mathbf{h}_{i};\mathbf{W}\mathbf{h}_{j}])}{\sum_{\tilde{j} \in \mathcal{N}(i)} \exp(\mathrm{LeakyReLU}(\mathbf{a}^{T}[\mathbf{W}\mathbf{h}_{i};\mathbf{W}\mathbf{h}_{\tilde{j}}]))} \label{2.9}\] where \(\mathbf{a}\) is a trainable vector, \(\mathbf{W}\) a trainable weight matrix, \(\mathbf{h}_i\) the hidden embedding of node \(i\), and \([\cdot\,;\cdot]\) the concatenation operation. *SGC* is a GCN without nonlinear activation functions between the layers, resulting from the authors' conjecture that the good performance of GCNs comes from the aggregation of local neighborhood information and not from the application of nonlinear feature maps. *gfNN* extends SGC with a nonlinear layer \(\sigma\) so that the node embeddings for layer \(l\) are obtained from \(\mathbf{H}^{(l)} = \sigma(\tilde{\mathbf{A}}^{l}\mathbf{X}\mathbf{W}),\) where \(\mathbf{W}\) is a trainable weight matrix. *APPNP* is based on the personalized PageRank (PPR) algorithm . The node embeddings in layer \(l+1\) are calculated via \(\mathbf{H}^{(l+1)} = (1-\alpha)\tilde{\mathbf{A}}\mathbf{H}^{(l)} + \alpha \mathbf{H}^{(0)},\) where \(\alpha \in (0,1]\) is a hyperparameter and \(\mathbf{H}^{(0)} := f(\mathbf{X})\), with \(f\) being a trainable neural network.
### Calibration methods
We consider the classical post-processing methods histogram binning, isotonic regression , and Bayesian binning into quantiles (BBQ) , which is a refinement of histogram binning. Further, we include temperature scaling as a multiclass calibration method and Meta-Cal , a recently introduced approach with state-of-the-art performance. *Histogram binning* divides the confidence scores \(\hat{p}_i\) into \(M\) bins and assigns a new score \(\hat{q}_m\) to each bin to represent the calibrated confidences. The scores \(\hat{q}_m\) are learned by minimizing \(\sum_{m=1}^{M} \sum_{i \in B_m}(\hat{q}_m-y_i)^2\). *Isotonic regression* learns a piecewise constant function \(f\) by minimizing \(\sum_{m=1}^M \sum_{i \in B_m}(f(\hat{p}_i)-y_i)^2\). It is a generalization of histogram binning where the bin boundaries and scores are jointly optimized. *BBQ* extends histogram binning and learns a distribution \(\mathbb{P}(\hat{q}_i \mid \hat{p}_i, \mathcal{D}_{\mathrm{val}})\) by marginalizing out all possible binnings, where \(\mathcal{D}_{\mathrm{val}}\) is the validation set. *Temperature scaling* is a single-parameter extension of Platt scaling for multiple classes. Given the output logit vector \(\mathbf{z}\) before the softmax activation, a rescaling \(\mathbf{z}/T\) depending on a temperature \(T > 0\) is applied. *Meta-Cal* combines temperature scaling as a base model with a bipartite ranking model to weaken the limitation of accuracy-preserving calibration methods. By investigating two practical constraints (miscoverage rate control and coverage accuracy control), the goal is to improve calibration depending on the bipartite ranking while controlling the accuracy.
# Experimental study {#sec:experiments}
## Setup
**Experiments** We first inspect the calibration of GNN models on benchmark citation datasets, where we take the best hyperparameter and training settings from the corresponding original papers. Then, we empirically analyze the influence of model capacity (width and depth) on calibration. It has been observed that stacking too many GCN layers drastically worsens the performance, which is partly attributed to a phenomenon called oversmoothing . Oversmoothing happens when repeated neighbor aggregation leads to similar node embeddings in the graph, and various methods have been proposed to tackle this problem . In the following, we investigate if increasing model depth also affects calibration. One of the core mechanisms of GNNs is the message aggregation from neighboring nodes. We examine how graph density, i. e., the ratio of the number of edges in the graph to the number of maximum possible edges, influences the calibration performance. Finally, we also test a new loss function [\[eq:loss\]](#eq:loss){reference-type="eqref" reference="eq:loss"} that combines the standard cross-entropy loss \(L_{\mathrm{ce}}\) with an ECE-inspired term \(L_{\mathrm{cal}}\) for optimizing the calibration. We define \(L_{\mathrm{cal}}\) as the cross entropy between the confidence of the sample and the accuracy of its corresponding bin, where the idea is that the confidence should stay close to the accuracy. Given the original cross-entropy loss \(L_{\mathrm{ce}}\), we define the new loss as
[\[tab:statistics\]]{#tab:statistics label="tab:statistics"}
\[\begin{aligned}
\label{eq:loss} L &= \alpha L_{\mathrm{ce}} + (1-\alpha) L_{\mathrm{cal}}\quad \mathrm{with}\\ L_{\mathrm{cal}} &=-\sum_{i=1}^{N}\mathrm{acc}(B_{m}(i))\cdot \log(\hat{p}_i)\;, \end{aligned}\] where \(\alpha \in (0,1)\) and \(B_{m}(i)\) denotes the bin that sample \(i\) belongs to. For the experiments on width, depth, graph density, and the new loss function, we focus on GCN and GAT, two of the basic and most widely used GNN models. **Datasets** Cora, Citeseer, and Pubmed[^2] are three commonly used benchmark datasets for node classification. They are citation networks, where nodes represent scientific publications and edges between pairs of nodes correspond to one publication citing the other. Each node comes with a \(d\)-dimensional feature vector that indicates the presence of words from a predefined vocabulary. The class label of a node is the topic of the corresponding publication. Similar to previous works , we operate under a semi-supervised setting, where only a small amount of labeled data is available during training. The statistics of the datasets are summarized in Table [1](#tab:statistics){reference-type="ref" reference="tab:statistics"}. **Implementation** The GNN models are implemented using the PyTorch-Geometric library[^3]. The bin number for calculating the ECE and MECE is set to 15. More information about the hyperparameters and experimental settings can be found in the supplementary material[^4].
## Results {#sec:gnn_calibration}
**Uncalibrated results** We run all GNNs on the three citation datasets and show the uncalibrated performance with respect to accuracy, ECE, and MECE in Table [2](#tab:uncal_performance){reference-type="ref" reference="tab:uncal_performance"}. The method APPNP is best on Cora and Pubmed in terms of accuracy (second-best on Citeseer), and it is also best calibrated on the datasets Citeseer and Pubmed. For Cora, gfNN has the lowest ECE and MECE. All models except for gfNN[^5] have stable calibration values with small standard deviations. The worst method with respect to the calibration performance is SGC, which is, apart from the softmax activation for normalization, the only linear model. Adding a nonlinear layer as in gfNN results in better calibration. Moreover, we find that GAT outperforms GCN in terms of ECE and MECE in two of three datasets.
**Influence of width** We compare the calibration of GCN and GAT for varying model width, i. e., the number of hidden dimensions per layer. While the accuracy basically stays constant, the ECE and MECE decrease with increasing number of hidden dimensions initially (see Fig. [\[fig:width\]](#fig:width){reference-type="ref" reference="fig:width"}). When a certain width is reached, the calibration values stagnate or slightly increase again. Generally, wider networks tend to be better calibrated.
**Influence of depth** We investigate the influence of model depth, i. e., the number of layers, on the calibration performance (see Fig. [\[fig:depth\]](#fig:depth){reference-type="ref" reference="fig:depth"}). Oversmoothing becomes particularly pronounced when the test accuracy decreases significantly with increasing number of layers. The ECE first improves when changing from two to three layers, then it increases again until five or six layers. Using an even larger model depth, the ECE eventually decreases again. **Influence of graph density** For this experiment, we remove different proportions of edges randomly from the dataset, ranging from \(0\%\) (original dataset) to \(100\%\) (no graph structure at all). The models GCN and GAT only differ in the aggregation mechanism, i. e., GAT introduces attention coefficients to weight the importance of neighbors. The results are shown in Fig. [\[fig:density\]](#fig:density){reference-type="ref" reference="fig:density"}. Similar to Table [2](#tab:uncal_performance){reference-type="ref" reference="tab:uncal_performance"}, the ECE of GAT is consistently lower than the ECE of GCN on Cora and Citeseer, while on Pubmed, GCN expresses partly better calibration. Generally, the graph density of Pubmed is the lowest. It might be that the attention weights in GAT are beneficial for calibration and especially useful when enough edges exist in the graph. **Influence of new loss function** Table [\[tab:cal_loss\]](#tab:cal_loss){reference-type="ref" reference="tab:cal_loss"} compares the results of the standard cross-entropy loss \(L_{\mathrm{ce}}\) and the new loss function \(L\) from [\[eq:loss\]](#eq:loss){reference-type="eqref" reference="eq:loss"}, which contains a calibration error term. The new loss \(L\) improves the model calibration in all cases while keeping the accuracy at the same level or even slightly increasing the accuracy. **Underconfidence vs. overconfidence** Taking the best hyperparameter and training settings from their corresponding publications, all GNNs exhibit underconfidence on all three datasets, i. e., the confidence scores are lower than the accuracy of the predictions (see Fig. [\[fig:rds\]](#fig:rds){reference-type="ref" reference="fig:rds"} and figures in the supplementary material). In some cases, however, we find that the model changes from underconfidence to overconfidence if it is trained without early stopping. The left plot in Fig. [\[fig:under_over_conf_epochs\]](#fig:under_over_conf_epochs){reference-type="ref" reference="fig:under_over_conf_epochs"} shows the test accuracy, scaled test negative log-likelihood (NLL), and scaled test ECE for a \(4\)-layer GCN on Cora during training, with a weight decay set to \(5\text{e-}4\). Around epoch \(150\), the NLL and accuracy become stable, while the ECE is still improving. At this point, GCN is underconfident, as shown in the left-most reliability diagram in Fig. [\[fig:under_over_conf_rd\]](#fig:under_over_conf_rd){reference-type="ref" reference="fig:under_over_conf_rd"}. In the epochs between \(200\) and \(300\), the ECE gains the best performance when the model changes from underconfidence to overconfidence (see the two diagrams in the middle of Fig. [\[fig:under_over_conf_rd\]](#fig:under_over_conf_rd){reference-type="ref" reference="fig:under_over_conf_rd"}). After epoch \(300\), GCN starts to overfit with repect to the ECE, while the NLL and accuracy remain rather unchanged. During this process, overconfidence aggravates, and the ECE increases to \(7.8\%\) in epoch 400, which is displayed in the right-most diagram in Fig. [\[fig:under_over_conf_rd\]](#fig:under_over_conf_rd){reference-type="ref" reference="fig:under_over_conf_rd"}. In summary, GCN first optimizes NLL and accuracy during training, then fits the confidence scores, and eventually starts to overfit regarding the ECE without influencing the NLL and accuracy. When we slightly increase the weight decay to \(7.5\text{e-}4\) (see right plot in Fig. [\[fig:under_over_conf_epochs\]](#fig:under_over_conf_epochs){reference-type="ref" reference="fig:under_over_conf_epochs"}), the ECE stabilizes after reaching the optimal value, and the values of NLL and accuracy also stay in a smaller range compared to the left plot.
# Ratio-binned scaling for calibrating GNNs {#sec:ratio}
## Same-class-neighbor ratio
From our experiments, we find that GNN models tend to be underconfident. Even though the overall model exhibits underconfidence, there might be differences depending on node-level properties, which have not been considered before. Especially for graph data, the topology could provide structural information that are useful for calibration. For node classification, the class labels and properties of a node's neighbors have a significant influence on the classification. We calculate for each node \(i\) the same-class-neighbor ratio, i. e., the proportion of neighbors that have the same class as node \(i\), and develop a new binning scheme that groups samples into bins based on the same-class-neighbor ratio for calibration. To evaluate the correlation between the same-class-neighbor ratio and the confidence of a model, we calculate the ratio for each node based on the ground-truth labels. In Fig. [\[fig:ratio\]](#fig:ratio){reference-type="ref" reference="fig:ratio"}, we group the nodes into \(5\) equally-spaced interval bins according to their ratios. Employing a trained GNN model, we compute the output of the classifier \(g\) for each node and draw the average confidence of the samples in each bin as a blue bar. The gap illustrates the difference between the average confidence and the accuracy in each bin. We observe that the average confidence increases with the same-class-neighbor ratio, where bins with higher ratios express underconfidence and bins with lower ratios overconfidence. Consequently, a binning scheme that groups samples depending on their same-class-neighbor ratios would take the graph structure into account and allow for an adaptive calibration depending on the confidence level of each bin.
## Ratio-binned scaling
We propose ratio-binned scaling (RBS), a topology-aware method, which first approximates the same-class-neighbor ratio for each sample, then groups the samples into \(M\) bins, and finally learns a temperature for each bin to rescale the confidence scores. In the semi-supervised setting, we only know the labels of a small number of nodes and therefore cannot use the true labels for binning. One natural option is to replace the nodes' ground-truth labels with their confidence scores for estimating the same-class-neighbor ratio. More precisely, we define the estimated ratio for node \(i\) as \[\hat{r}(i) = \frac{1}{\left | \mathcal{N}(i) \right |} \sum_{j \in \mathcal{N}(i)} g(\mathbf{H}_j)_{\hat{y}_i} \quad \in [0,1]\;,\] where \(\mathbf{H}_j\) is the node embedding of node \(j\), which is learned by a GNN model, and \(g\) is the classifier. \(g(\mathbf{H}_j)_{\hat{y}_i}\) denotes the confidence score of node \(j\) corresponding to the class \(\hat{y}_i\) that is predicted for the central node \(i\).
For the ratio-based binning scheme, let \(\{B_m \mid 1 \leq m \leq M\}\) be a set of bins that partitions the interval \([0,1]\) uniformly. After calculating the output for all nodes, each node \(i\) is assigned to a bin according to its estimated same-class-neighbor ratio \(\hat{r}(i)\), i. e., \(B_1 = \{i \in \mathcal{V} \mid \hat{r}(i) \in [0, \frac{1}{M}]\}\) and \(B_m = \{i \in \mathcal{V} \mid \hat{r}(i) \in (\frac{m-1}{M},\frac{m}{M}]\}\) for \(m \in \{2, \dots, M\}\). Let the output of \(g\) be in the form \(g(\mathbf{H}_i) = \sigma(\mathbf{Z}_i) \in \mathbb{R}^K\) for node \(i\), where \(\sigma\) is the softmax function and \(\mathbf{Z}_i\) the logits before normalization. For each bin \(B_m\), \(m \in \{1, \dots, M\}\), a temperature \(T_m > 0\) is learned on the validation dataset. The calibrated confidence for a test node \(j\) is then given by \[\mathbf{\hat{q}}_j = \sigma (\mathbf{Z}_{j}/T_{m}) \in [0,1]^K\quad \mathrm{if}\; j \in B_m.\] We apply temperature scaling for calibrating the nodes in each bin, but it would also be possible to apply other post-processing calibration methods for obtaining calibrated scores.
## Results {#results}
Table [\[tab:cal_performance\]](#tab:cal_performance){reference-type="ref" reference="tab:cal_performance"} summarizes the calibration performance of all considered post-processing calibration methods and our proposed method RBS in terms of ECE. All methods can improve the calibration of GNN models on Cora and Citeseer. In particular, the obtained ECE for a specific dataset and calibration method is rather similar for all GNNs regardless of the uncalibrated ECE. On Pubmed, most methods have difficulties improving calibration of APPNP, which already has low ECE. RBS gains the best performance in the majority of the cases and outperforms classical temperature scaling in \(11\) out of \(15\) experiments. Next to a good calibration performance, accuracy preservation is desirable for calibration methods. RBS and temperature scaling do not change the ranking of the classes and thus the accuracy stays unchanged. Meta-Cal trades good calibration for lower accuracy, while the other methods yield comparable or even slightly improved accuracy in some cases (see supplementary material).
## Effectiveness of real-ratio-binned scaling
Table [\[tab:cal_performance\]](#tab:cal_performance){reference-type="ref" reference="tab:cal_performance"} also shows the calibrated results of real-ratio-binned scaling (RRBS), where we assume that the ground-truth labels are available for all nodes. RRBS outperforms the best calibration method in 14 out of 15 experiments. Although the correct labels are not accessible in the semi-supervised setting, the results still indicate the effectiveness of the intuition of our proposed method.
# Discussion {#sec:discussion}
In general, the calibration performance depends on the specific GNN model and dataset, where all models perform best on Pubmed (see Table [\[tab:cal_performance\]](#tab:cal_performance){reference-type="ref" reference="tab:cal_performance"}). When using the hyperparameter and training settings from the original publications, all GNNs tend to be underconfident on all three datasets, in contrast to the finding that deep neural network models rather exhibit overconfidence . However, when plotting the reliability diagrams for a varying number of epochs, we observe that in some cases, underconfidence changes to overconfidence when the number of epochs increases. It seems that underconfidence or overconfidence is not necessarily a property of the model architecture but is also dependent on the training setting. Most GNNs suffer from oversmoothing, which becomes apparent when increasing the number of layers in the model . We observe that for large numbers of layers, the accuracy drops significantly, while the ECE improves. Oversmoothing results in similar node embeddings, which might be uninformative for the model. In this case, the model would most likely learn the distribution of classes in the training data as confidence scores. Therefore, all samples would be grouped into one bin, resulting in low ECE if the test distribution is close to the training distribution. However, such a model does not make use of the underlying graph structure and is probably not useful for application. The results for RBS and RRBS show the potential of a calibration method that takes the graph structure into account, where the binning scheme is constructed depending on node-level properties. RRBS almost always outperforms RBS, which suggests that RBS might be especially helpful for cases where the estimated ratios are close to the real ratios, i.e., for models with relatively high accuracy. The number of bins for RBS was chosen from \(\{2,3,4\}\), and it seems that even a small number of bins can lead to improved calibration compared to classical temperature scaling. It would further be interesting to apply RBS to other kinds of datasets, e.g., heterophilic graphs, where nodes from different classes are likely to be connected.
# Conclusion
We investigated the calibration of graph neural networks for node classification on three benchmark datasets. Graph neural networks seem to be miscalibrated, where the exact calibration depends on both the dataset and the model. Existing post-processing calibration methods are able to alleviate the miscalibration but do not consider the graph structure. Based on our experimental finding that the nodes in the graph express different levels of over-or underconfidence depending on their same-class-neighbor ratios, we proposed the topology-aware calibration metohd ratio-binned scaling. It takes the predictions of neighboring nodes into account and shows better performance compared to state-of-the-art baselines. For future work, it would be interesting to gain a more theoretical understanding of the calibration properties and conduct experiments on larger and a wider variety of datasets. | {'timestamp': '2022-06-06T02:14:34', 'yymm': '2206', 'arxiv_id': '2206.01570', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01570'} |
# Introduction
The physical properties of confined electronic structures have been widely investigated, mainly thanks to the possibility of tuning electronic states and its transport responses. Nanostructured systems emerge in different shapes and dimensions and further than being only dreamed theoretical idealization are being synthesized making use of multiple bottom-up and up-bottom growth processes. In particular, systems exhibiting dispersionless states in the electronic structure, named as flat bands have been pursued with the help of different strategies, such as controlling the angle of twisted bilayers, switching on electric and magnetic fields, and also engineering the growth of buckled graphene superlattices. Recently, the fractal dimension opened a new door in the study of the electronic properties in confined systems. The Sierpinski Triangle (ST) is a fractal mathematical structure that was used physically as a quantum corral to confine electrons. Despite the fact ST is represented in a bidimensional form, it has a fractal dimension, called Hausdorff dimension (D) given by \(D\approx1.585\). Graphene lattices and other structures, as illustrated in Fig. [\[sgtypes\]](#sgtypes){reference-type="ref" reference="sgtypes"}, have inspired theoretical studies on fractal Sierpinski triangle samples with different purposes such as exploring spin polarization features and the emergence of flat bands. In particular, extended eigenstates are verified in the spectrum of hexagonal ST \[Fig.1 (d)\], revealed as continuous bands as the lattice is threaded by magnetic fluxes. The role of fractality was already studied for diffusion, and random walks in photonic lattices, in the synthesis of fractal supramolecular nanostructures, all with interesting applications. One-dimensional molecular chains using Sierpinski triangles as building blocks have been reported using low-temperature Scanning Tunnelling Microscopy (STM). The systems were grown on Au(111) and the success of the chain formation depends on the molecular coverage and matching between molecular size and surface lattices. Experimental methods using self-assembly and templating are being used to synthesize ST flakes and nanoribbons with transition metals and metal-organic composites. Also, 2D crystals consisting of ST units are reported by exploiting benzene-like molecules and Fe atoms on Au(111) by combining molecular design and epitaxy control. Motivated by the rich experimental scenario based on different molecular constructions, we propose theoretical models to describe such synthesized lattices. Results of local electronic density of states of simplified molecular structures are compared with spatial charge distribution reported on distinct STM images. Here we explore electronic and transport properties of such molecular fractal chains, indicating the viability of tuning electronic states by construction, providing possible smart devices.
# Theoretical Model
A single-orbital tight-binding (TB) Hamiltonian is used to describe the ST systems studied, given by \[\begin{gathered}
H=\sum_{i}\varepsilon_{i}c_{i}^{\dagger}c_{i}+\sum_ {\left\langle ij\right\rangle} t_{ij}c_{i}^{\dagger}c_{j}+\sum_ {\left\langle\langle ij\right\rangle\rangle} t'_{ij}c_{i}^{\dagger}c_{j}+h.c.\label{Hamito} \end{gathered}\] with \(\varepsilon_i\) being the on-site energy for each atom located at site \(i\), \(c^{\dagger}_i\) (\(c_i\)) is the creation (annihilation) operator of an electron on site \(i\), and \(t_{i,j}\) and \(t'_{i,j}\) are the hopping energies for nearest and second-nearest neighboring atoms, respectively. Following the Green function formalism we obtain local density of states and we also investigate transport properties using the Landauer approach in which the system is decoupled into three parts: central conductor and right and left leads. We have considered semi-infinite ST chains as leads, matching perfectly with the central region. The central advanced (a) and retarded (r) Green functions are given as \[G^{a,r}_{c}(E)=\Big[\omega-{H_c}-\Sigma^{a,r}_L(E)-\Sigma^{a,r}_R(E)\Big]^{-1}\,\,,\] with \(\omega = E \pm i\eta\), \(\eta\) being an infinitesimal number. \({H}_{c}\) is the Hamiltonian of the central part, and \(\Sigma^{a,r}_{L,R} (E)\) correspond to left and right self-energies, given by the related surface Green functions, from which the coupling matrices are obtained: \[\Gamma^{L,R} (E)= i \Big(\Sigma_{L,R}^{r} (E)-\Sigma_{L,R}^{a}(E)\Big)\,\.\] Finally, to derive the electronic conductance in ST chains, G(E)=\(2e^2/h\mathcal{T}(E)\), we calculate the energy-dependent transmission given by \[{\mathcal{T}(E)}=\operatorname{Tr}\left[\boldsymbol\Gamma^{L} \boldsymbol{G}^{r}_{c} \boldsymbol\Gamma^{R} \boldsymbol{G}^{a}_{c} \right]\,\.\]
## Conventional Flake model
Inspired on the possibility of engineering real structures based on the geometries of the Sierpinski triangle we developed a recursive process to describe ST of different generations, based on the Green's function formalism. Our first emphasis is to investigate fractal features on the electronic properties of ST flakes. Here we show the results of the density of states (DOS) for conventional ST flakes in terms of the order generation \(l\).\
The electronic properties are obtained by using the single first neighbor tight binding Hamiltonian, described in Eq.(1) and putting \(t'=0\). The hopping energy matrices follow the corresponding connectivity of the \(l\)-order ST \[G(l)\], illustrated in Fig.[\[dcmethod\]](#dcmethod){reference-type="ref" reference="dcmethod"}. The DOS is given by\
\[\rho_{0}(E) =-\frac{1}{\pi} Im G_{00}(E).\label{eqDOS}\]\
The Dyson equations relating the real-space Green functions, among the 15-atoms of the second ST generation, shown in Fig.2, can be written as, \[\begin{gathered}
G_{00} = g_{0} + g_{0}t[G_{10} + G_{20} + G_{30} + G_{40}]\\ \nonumber G_{10} = g_{1}t[G_{00} +G_{20} +G_{60} +G_{50}]\\ \nonumber G_{20} = g_{2}t[G_{00} +G_{10} +G_{60} +G_{90}]\\ \nonumber G_{30} = g_{3}t[G_{00} +G_{40} +G_{70} +G_{110}]\\ \nonumber G_{40} = g_{4}t[G_{00} +G_{30} +G_{70} +G_{80}]\\ \nonumber G_{60} = g_{6}t[G_{10} +G_{20} +G_{50} +G_{90}]\\ \nonumber G_{70} = g_{7}t[G_{30} +G_{40} +G_{80} +G_{110}]\,\,, \nonumber \end{gathered}\] where the propagator \(g_i(E) = 1/(w-E_i)\) and \(t=t_{i,j}\) corresponds to the energy hopping between two first neighboring atoms i and j in the lattice, which were considered identical. After some algebraic manipulations involved in a decimation procedure, and considering \(g_i=g_0\), for all sites i, we obtain a renormalized Dyson equation for the \(G_{00}\) locator, \[G_{00} = \tilde{g_{0}} + \tilde{g_{0}}\tilde{t}[G_{50} +G_{80}+G_{90} +G_{110}]\,\,,\] corresponding to a reduced ST generation, with the dressed propagator and hopping energy given by, \[\tilde{g_{0}}=\frac{g_{0}}{1-\frac{4g_{0}^{2}t^{2}(1+g_{0}t)}{1-3g_{0}^{2}t^{2}-2g_{0}^{3}t^{3}}}\,\,,\] and \[\tilde{t}=\frac{g_{0}t^{2}(1+2g_{0}t)(1+g_{0}t)}{1-3g_{0}^{2}t^{2}-2g_{0}^{3}t^{3}}\,\.\] By realizing \(l\) iterative process it is possible to obtain the DOS of the \(l-th\) ST generation. The results for different numbers of iterative processes (1-4) are shown in Fig.[\[sgflakes\]](#sgflakes){reference-type="ref" reference="sgflakes"}. As being finite systems, the DOS are expected to exhibit a sequence of delta functions that increase as the number of atoms is increased, as depicted in Fig.[\[sgflakes\]](#sgflakes){reference-type="ref" reference="sgflakes"}. A self-similarity of the DOS, emerging from higher \(l\)-order STs is evident in the zoom (bottom panel) presenting the short energy range (-1.0t to 0.75t). Particular pinned localized states are found at the same energies, independently of the ST generation-order. More interesting to note is the fact that all states presented in a given iterative process also appear in the LDOS results of all subsequent iterative steps of higher order, in a cumulative process.
## Kagomé-Honeycomb ST Flake model
Motivated by fractal properties in more realistic systems we propose the Kagomé-Honeycomb ST flake (KHST). In this case we solve the tight-binding system up to second neighbors, as illustrated in Fig.[\[zz+armkagome\]](#zz+armkagome){reference-type="ref" reference="zz+armkagome"}-(a) for armchair-KHST (A-KHST) and in Fig.[\[zz+armkagome\]](#zz+armkagome){reference-type="ref" reference="zz+armkagome"}-(b) for zigzag-KHST (Z-KHST). The number of atoms for the armchair (\(N_{A}\)) and zigzag (\(N_{Z}\)) configurations are, respectively, \[\begin{gathered}
N_{A}^{l}=12*3^{l}+\sum_{i=1}^{l}3^{i}\nonumber\,\,, \\ N_{Z}^{l}=15*3^{l}+4-\sum_{i=0}^{l-1}3^{i}, \ {\rm for} \ l > 1 \ {\rm and} \ N_{Z}^{1}=49\,\. \end{gathered}\] All the calculations are made using on-site energies equal to \(\pm\epsilon= \pm0.25t_1\), with \(t_1\) being the nearest-neighbor hopping, for the blue (\(+\epsilon\)) and orange (\(-\epsilon\)) sites, following reports on two-dimensional covalent organic honeycomb frameworks, that are typical example of graphene-kagomé lattices. The second nearest neighbor hopping energy is chosen as \(t_2\)=0.08\(t_1\). The LDOS is calculated as a function of energy by \[LDOS(x,y,E)=\sum_n \abs{\Phi_n(x,y)}^2 \delta(E-E_n)\,\,,\] where \(x\) and \(y\) are the lattice position coordinates and \(\Phi(x,y)\) is the corresponding electronic wave function of the \(n^{th}\) state. To compute the delocalized electronic contribution around the sites, given by \(\abs{\Phi_n(x,y)}^2\) we transform the delta functions into Lorentzians with \(\Gamma = 0.08 t_1\). For both KHST edge geometries, the presence of a band gap is verified, for the cases of first and up to second-neighbors, as shown in Fig.[\[zz+armkagome\]](#zz+armkagome){reference-type="ref" reference="zz+armkagome"} (right panels), revealing a semiconducting feature to such Kagomé-like ST. Otherwise, the graphene ST flake results, reveal a semiconducting and metallic characteristic for armchair and zigzag GSTs, respectively in the energy-state maps. As noticeable from Figs.[\[zz+armkagome\]](#zz+armkagome){reference-type="ref" reference="zz+armkagome"}(a) and (b), the gap increases by \(\approx\) 0.16\(t_1\) when second nearest neighbours are taken into account in the TB model, for armchair and zigzag-edged KHSTs. An interesting feature found in the electronic properties of the KHST gaskets is the reminiscent flat band observed in 2-dimensional Kagomé lattice, highlighted here by the high degeneracy order of the E=-0.25\(t_1\) and-0.41\(t_1\) states, for first and second-neighbor models, respectively, at the gap threshold. The number of such localized degenerate states is expected to increase as the generation order of the gasket is increased, although the size of the gaps do not dependent on the generation order, as depicted in the insets of Fig.[\[zz+armkagome\]](#zz+armkagome){reference-type="ref" reference="zz+armkagome"}, for armchair and zigzag configurations. Compared with the gap dependence for GST flakes, for which the gap size saturates for \(l\) between 3 and 4, we must comment that the number of atoms in the KHST flakes are considerable superior at the same generation order, what may justify the constant gap size achieved already in the first generation. Due to the difference in the geometry and site numbers in the basic units forming the KHSTs in comparison to the conventional ST, another fractal dimension definition is used. Here we adopt the Minkowski--Bouligand dimension used for monofractals, given by \[D=\lim_{r \to 0} \frac{Log(N(r))}{Log(r^{-1})}\,\,,\] with \(N(r)\) and \(r\) being the number of squares covering the full system, and the square size, respectively. These parameters are used in the box-counting method to perform the dimension calculation. The dimension of each system is determined by the line slope coefficient of \(Log(N(r))\) vs \(Log(r^{-1})\) graph presented in Fig.[\[HAUSDORFF\]](#HAUSDORFF){reference-type="ref" reference="HAUSDORFF"}(a). Starting from the atomic spatial localization in the A-KHST flake (see Fig.[\[sgtypes\]](#sgtypes){reference-type="ref" reference="sgtypes"}(e)), the sizes (in pixels) of each square found to better adjust the slope are 10 up to 90 pixels. In both cases we take the LDOS image (360 x 360 pixels) and binarize it in a threshold of \(30\%\), to represent the most accurate electronic distribution. The details of such numerical process is described in Refs.. Perfect Kagomé nanoribbons and Kagomé 2D-lattices, exhibit symmetric band spectra, except for the presence of a flat band located exactly at \(\epsilon\) or-\(\epsilon\), depending on the hopping energy signal. For this reason, in this work we choose then the values of \(E=\pm \epsilon\) to calculate the maximum contribution in LDOS indicated by its brightest spots, which gives two different patterns, as can be seen in the binarized image shown in Fig.[\[HAUSDORFF\]](#HAUSDORFF){reference-type="ref" reference="HAUSDORFF"}-(b,c). In this system we have two energy values that the maximum contribution in LDOS is spread through the flake, given origin to the patterns (b) and (c) in Fig.5. Taking them both in the analysis, the box counting method yields a fractal dimension close to the conventional Sierpinski triangle gasket, \(\approx 1.585\), as can be verified by the slope values shown in Fig.[\[HAUSDORFF\]](#HAUSDORFF){reference-type="ref" reference="HAUSDORFF"}(a), for both situations (\(1^{st}\) and up to \(2^{nd}\) neighbor hopping). This occurs because the brightest spots in the LDOS maps are near from the geometrical points of a A-GST flake, that are also fractal. Otherwise, for energies where the bright spots (high LDOS values) do not reveal the real geometry of the flake, the fractal dimension varies between D=1.30-1.80, as verified by Kempkes *et al.* in hexagonal flakes, varying between low and high contributions of each site for the LDOS, respectively. The self-similarity observed in others ST flakes is also evidenced here for both configurations of the KHSTs (armchair and zigzag), independently on the order of the neighboring hopping taken into account. A further interesting feature to explore in zigzag and armchair KHST flakes, described by the first and second neighbors approximation, is the asymmetric distribution of the electronic probability distribution at \(E_1\)=-0.25\(t_1\) and \(E_2\)=0.25\(t_1\) states. Considering the first generation G(1), we found that while the \(E_1\) state is formed by four-fold degenerate eigenstates, the \(E_2\) state is non-degenerated. These states are spread differently among the orange and blue sites of the flake \[see Fig. [\[zz+armkagome\]](#zz+armkagome){reference-type="ref" reference="zz+armkagome"}\], giving raise to the LDOS exhibited in Fig.[\[HAUSDORFF\]](#HAUSDORFF){reference-type="ref" reference="HAUSDORFF"}(d) and (e), respectively. The same behavior is evidenced for higher orders of the KHSTs, considering both first and up to second-neighbor hopping models.
## ST Mirrored Chains
Following the experimental realization of 1-D molecular chains, with ST as building blocks grown on Au(111), we explore electronic properties of a similar quasi-1D chain as depicted in Fig [\[stchain\]](#stchain){reference-type="ref" reference="stchain"}(a). The system was idealized based on the coupling of two hexagonal STs spatially inverted (up and down), with mirror symmetry. Green and black lines connect nearest and second-nearest neighboring atoms, respectively, through the hopping terms \(t\) and \(t'\) in the tight binding Hamiltonian given in Eq.(1). Differently from the long-range ordered grown structures where Co atoms are present in intercalated benzene lattices, our simple model involves exclusively carbon atoms, denoted by blue dots. The electronic properties are calculated using on site energy \(\epsilon=0\) and following the relation t'=0.08t for the hopping parameters as recently for HSTs. The LDOS shown in Fig.[\[stchain\]](#stchain){reference-type="ref" reference="stchain"}-(b) is calculated via Eq.[\[eqDOS\]](#eqDOS){reference-type="ref" reference="eqDOS"}, where \(G_{00}\) is obtained using similar real-space decimation methods, properly constructed to infinite periodic systems. At energies close to the Fermi level, the electronic group velocity is near zero, giving origin to almost two dispersionless bands as can be seen in the corresponding electronic structure shown in Fig.[\[stchain\]](#stchain){reference-type="ref" reference="stchain"}-(c). These flat bands appear as highly peaked density of states, and are highlighted in the inset in Fig.[\[stchain\]](#stchain){reference-type="ref" reference="stchain"}-(b). The different curves correspond to the assigned sites marked with the same color in the unit cell displayed in part (a). As seen in Fig.[\[stchain\]](#stchain){reference-type="ref" reference="stchain"}-(b), at E=-0.16\(t\), the main contribution becomes from sites 1, 2 and 3, and from the symmetric upper sites 11, 10 and 9. A better visualization of the spatial electronic distribution through the system is displayed in Fig.[\[ldosst\]](#ldosst){reference-type="ref" reference="ldosst"} for distinct energies. While part (a) refers to a STM image adapted from ref., Figs.[\[ldosst\]](#ldosst){reference-type="ref" reference="ldosst"} (b), (c), and (d), are the LDOS theoretical results at \(E=-0.16t\) (flat band), \(0.96t\) and \(1.65t\). In the STM image a high charge distribution appears as bright spots at the relative atomic sites. Correspondingly, in our results the color maps refer to the LDOS intensity. The normalized LDOS results shown in Fig.[\[ldosst\]](#ldosst){reference-type="ref" reference="ldosst"} (b) reveal that at one of the flat bands \[E=-0.16t\] not all the atomic sites of the lattice are populated, in according with the previously discussion. This last state appears again in the Dephased ST Chain, and in special, it will not contribute to the transport properties of these systems, as we will see in the next session. Moreover, while at \(E=0.96t\), only sites 2 and 5, and their equivalents in the unit cell, contribute to the state \[see Fig.[\[stchain\]](#stchain){reference-type="ref" reference="stchain"}-(c)\], revealing a restricted charge distribution, for the state \(E=1.65t\), Fig.[\[stchain\]](#stchain){reference-type="ref" reference="stchain"}-(d), almost all the lattice is visited generating bright spots in the geometrical net. The later LDOS pattern strongly resembles the cited STM image \[Fig.[\[ldosst\]](#ldosst){reference-type="ref" reference="ldosst"}-(a)\] indicating that our simple model can be an important tool of modelling such organic molecular chain. Another experimental alternative of generating 1D mirrored chains made of Sierpiński triangles was proposed in Ref.. We call it as mirrored chain structure-type 2 (MSTC-2). In our model up and down triangles are now fully preserved resembling the STM images shown in Fig. [\[stchainalternative\]](#stchainalternative){reference-type="ref" reference="stchainalternative"}(a). The LDOS at a particular energy (\(E=2.4t\)) is depicted in part (b) where the region enclosing the nanoribbon unit cell is marked with red circles. The results for DOS and conductance are shown in Fig.[\[stchainalternative\]](#stchainalternative){reference-type="ref" reference="stchainalternative"} (c), highlighted with shadowed blue curves and compared with the results found for the previous molecular chain \[shaded gray curves\]. It is noticeable the emergence of two narrowed energy states near the Fermi Level for the MSTC-2, this is a characteristic property of double ST Chains that will be explored in next section. Also, the flat-like states in both chain configurations are preserved in the central gap in DOS, although they are suppressed in the electronic conductance response of each molecular designs due to the high localization features of the flat band.
## Dephased ST Chain
Following the experimental molecular structures presented in Ref., we address now another proposal for ST chain, as illustrated in Fig. [\[g1g2g3comparison\]](#g1g2g3comparison){reference-type="ref" reference="g1g2g3comparison"}-(a). Differently from the previous discussed molecular chains, the new structure are composed by dephased pairs of HSTs \[G(2)\], connected at particular lateral lattice sites, simulating the packing mode of STs produced by a combination of a Co atom and three BPyB molecules. The unit cells of the dephased molecular chain is marked with dashed red lines. It is interesting to note that our theoretical result for the LDOS at \(E=1.4t\), as shown in Fig.[\[g1g2g3comparison\]](#g1g2g3comparison){reference-type="ref" reference="g1g2g3comparison"}-(d), reproduces quite well the STM image presented in Fig.[\[g1g2g3comparison\]](#g1g2g3comparison){reference-type="ref" reference="g1g2g3comparison"}-(c) for the grown molecular nanostructured, reported in Ref.. The electronic properties of the G(2)-dephased ST is shown in Fig. [\[g1g2g3comparison\]](#g1g2g3comparison){reference-type="ref" reference="g1g2g3comparison"}-(b) via the density of states and conductance results. Due to the high electronic localization, the DOS peak at E=-0.16t does not contribute to the electronic conductance in G(2), what happens also for dephased chains of higher ST orders (not shown here). To investigate fractal properties on such nanostructured systems we concentrate now in the electronic band structures. We present in Fig. [\[g4g3g2\]](#g4g3g2){reference-type="ref" reference="g4g3g2"} (a-c) unit cells of three generations of the proposed molecular chain discussed in Fig.[\[g1g2g3comparison\]](#g1g2g3comparison){reference-type="ref" reference="g1g2g3comparison"}, G(2), G(3) and G(4), and the corresponding band structure of the 1D chains. The dashed regions marked in the band structures presented in Fig.[\[g4g3g2\]](#g4g3g2){reference-type="ref" reference="g4g3g2"}(a), related to the G(2) lattice, was enlarged in the subsequent panels at right, revealing a single flat band and other two bands presenting a small energy dispersion, at the lower energy range, highlighted by the red shaded area. These later are also noted in the DOS and conductance results (not shown). To analyze the role played by the generation order on the used HST blocks forming the chains, we show in Fig.[\[g4g3g2\]](#g4g3g2){reference-type="ref" reference="g4g3g2"} (b) and (c), for comparison, the energy bands of the three G(3) and G(4) dephased ST chains. The auto similarity of these systems can be evidenced looking firstly, at the same energy region of the three generation proposed, defined by the dashed energy range. Comparing the results, we notice in the last right panels that the common pattern of a single flat band (happening at 0.96t in the three cases), followed by two dispersive bands \[ending at 1.1t (a), 1.0t (b) and 0.97t (c), respectively\] is slightly preserved going from G(2) to G(4). As expected, for increasing ST generations, the repeated band structure pattern are found at smaller energy ranges. Moreover, the results for the energy bands reveal an increasing number of 1D-like van Hove singularities passing from G(2) to G(4) dephased chains. Similarly as reported before, the number of flat bands increases significantly with the ST order used to build the chains. Our findings suggest that properly tuning the tight binding parameters and playing with different connections between and inside the unit cells, better modeling of such nanostructured systems may be achieved. In particular, considering different on-site energies for Co atoms and BPyB molecules in our model (for instance taking \(E_C=-E_{Co}\)), simplified here as carbon atoms, only promotes a shift in the double flat states near the fermi level.
# Conclusion
We have explored fractal features in the electronic properties of ST flakes and molecular chains of different geometry details. The results found for the fractal dimension of the Kagomé-like ST flakes follow recent reports on graphene-like STs. The result is corroborated by the Kagomé ST LDOS at particular energy states that exhibit typical graphene ST spatial configurations. The self similarity of the energy states are found comparing different ST generation orders and also amplifying the energy ranges investigated, for both flakes and quasi 1D systems. In particular, the results for the local density of states of the theoretical molecular chains proposed here exhibit quite similar spatial charge distribution as experimental STM reports. The analysis of transport response of such quasi 1D molecular chains reveals localized states that do not contribute to the electronic transport. The study can be used as guide to propose a variety of architecture in the synthesis of real molecular chains. Further aspects of the theoretical framework adopted were explored such as changing the inter and intra unit cell connections of the proposed molecular chains. Some of the schemes induce degeneracy break of particular electronic states, favouring the disrupt of flat bands. Although a primary analysis on more realistic on-site energies, simulating metallic atoms in the ST chains that appear in the experiments, has not indicated great changes, more sophisticated theoretical framework must be used for a fully study of tuning the energy channel positions in the transport features of the chains. Probably, the inclusion of electron-electron interaction in the model will bring important light into the electronic occupation in conductive states, in special, in the flat states presented. | {'timestamp': '2022-06-06T02:15:43', 'yymm': '2206', 'arxiv_id': '2206.01610', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01610'} |
# Broader Impact
This work provides a technical advancement in the field of unsupervised representation learning, that already has great impact throughout many applications. In particular, in the medical context, including prior information can be of great value in order to convince clinicians to use AI solutions for computer-aided diagnosis. This work is also a theoretical advancement in representation learning for deep learning (DL) applications. We believe that it is crucial to provide theoretical guarantees of current DL systems in order to improve their reliability and robustness.
# More Empirical Evidence
In this section, we provide additional empirical evidence to confirm several claims and arguments developed in the paper.
## Decoupled Uniformity optimizes alignment
We empirically show here that Decoupled Uniformity optimizes alignment, even in the regime when the batch size \(n> d+1\), where \(d\) is the representation space dimension. We use CIFAR-10 and CIFAR-100 datasets and we optimize Decoupled Uniformity (without kernel) with all SimCLR augmentations with \(d=128\) and we vary the batch size \(n\). We report the alignment metric defined in as \(\cL_{align}=\bbE_{\cA(x|\bar{x})\cA(x'|\bar{x})p(\bar{x})} ||f(x)-f(x')||^2\).
## Multi-view Contrastive Learning with Decoupled Uniformity
When the intra-class connectivity hypothesis is full-filled, we showed that Decoupled Uniformity loss can tightly bound the classification risk for well-aligned encoders (see Theorem [\[th:boundness_intra_class_hyp\]](#th:boundness_intra_class_hyp){reference-type="ref" reference="th:boundness_intra_class_hyp"}). Under that hypothesis, we consider the standard empirical estimator of \(\mu_{\bar{x}}\approx \sum_{v=1}^V f(x^{(v)})\) for \(V\) views. Using all SimCLR augmentations, we empirically verify that increasing \(V\) allows for: 1) a better estimate of \(\mu_{\bar{x}}\) which implies a faster convergence and 2) better SOTA results on both small-scale (CIFAR10, CIFAR100, STL10) and large-scale (ImageNet100) vision datasets. We always use batch size \(n=256\) for all approaches with ResNet18 backbone for CIFAR10, CIFAR100 and STL10 and ResNet50 for ImageNet100. We report the results in Table [1](#tab:multiview_decoupled_unif){reference-type="ref" reference="tab:multiview_decoupled_unif"}.
## Influence of temperature and batch size for Decoupled Uniformity
InfoNCE is known to be sensitive to batch size and temperature to provide SOTA results. In our theoretical framework, we assumed that \(f(x)\in \bbS^{d-1}\) but we can easily extend it to \(f(x)\in \sqrt{t}\bbS^{d-1}\) where \(t>0\) is a hyper-parameter. It corresponds to write \(\cL_{unif}^d(f)=\bbE_{p(\bar{x})p(\bar{x}')}e^{-t||\mu_{\bar{x}}-\mu_{\bar{x}'}||^2}\). We show here that Decoupled Uniformity does not require very large batch size (as it is the case for SimCLR) and produce good representations for \(t\in [1, 5]\).
Datasets \(t=0.1\) \(t=0.5\) \(t=1\) \(t=2\) \(t=5\) \(t=10\)
---------- --------- --------- ------- ------- ------- --------
CIFAR10 73.91 83.01 84.72 85.82 83.05 74.82
CIFAR100 39.16 51.33 55.91 58.89 56.70 48.29
: Linear evaluation accuracy (%) after training for 400 epochs with batch size \(n=256\) and varying temperature in Decoupled Uniformity loss with SimCLR augmentations. \(t=2\) gives overall the best results, similarly to the uniformity loss in
Datasets Loss \(n=128\) \(n=512\) \(n=1024\) \(n=2048\)
---------- ---------------- --------- --------- ---------- ----------
CIFAR10 SimCLR 78.89 79.40 80.02 80.06
Decoupled Unif 82.67 82.12 82.74 82.33
CIFAR100 SimCLR 49.53 53.46 54.45 55.32
Decoupled Unif 54.61 54.12 55.56 55.20
: Linear evaluation accuracy (%) after training for 200 epochs with a batch size \(n\), ResNet18 backbone and latent dimension \(d=128\). Decoupled Uniformity is less sensitive to batch size than SimCLR thanks to its decoupling between positives and negatives, similarly to .
## Kernel choice on RandBits experiment
In our experiments on RandBits, we used RBF Kernel in Decoupled Uniformity but other kernels can be considered. Here, we have compared our approach with a cosine kernel on Randbits with \(k=10\) and \(k=20\) bits. There is no hyper-parameter to tune with cosine. From Table [2](#tab:kernel_randbits){reference-type="ref" reference="tab:kernel_randbits"}, we see that cosine gives comparable results for \(k=10\) bits with RBF but it is not appropriate for \(k=20\) bits.
# Geometrical Considerations about Decoupled Uniformity
In this section, we provide a geometrical understanding of Decoupled Uniformity loss from a metric learning point of view. In particular, we consider the Log-Sum-Exp (LSE) operator often used in CL as an approximation of the maximum. We consider the finite-samples case with \(n\) original samples \((\bar{x}_i)_{i\in[1..n]}\overset{iid}{\sim} p(\bar{x})\) and \(V\) views \((x_i^{(v)})_{v\in [1..V]}\overset{iid}{\sim}\cA(\cdot|\bar{x}_i)\) for each sample \(\bar{x}_i\). We make an abuse of notations and set \(\mu_i=\frac{1}{V}\sum_{v=1}^V f(x_i^{(v)})\). Then we have: \[\begin{aligned} \hat{\cL}_{unif}^d &= \log \frac{1}{n(n-1)}\sum_{i\neq j} \exp\left(-||\mu_i-\mu_j||^2\right)\\ &= \log \frac{1}{n(n-1)}\sum_{i\neq j} \exp\left(-s_i^+-s_j^+ + 2 s_{ij}^-\right) \label{eq:proofLdunif} \end{aligned}\] where \(s_i^+ = ||\mu_i||^2 = \frac{1}{V^2} \sum_{v, v'} s(x_i^{(v)}, x_i^{(v')})\), \(s_{ij}^-=\frac{1}{V^2}\sum_{v, v'} s(x_i^{(v)}, x_j^{(v')})\) and \(s(\cdot, \cdot)=\langle f(\cdot), f(\cdot)\rangle_2\) is viewed as a similarity measure. From a metric learning point-of-view, we shall see that minimizing Eq. [\[eq:proofLdunif\]](#eq:proofLdunif){reference-type="ref" reference="eq:proofLdunif"} is (almost) equivalent to looking for an encoder \(f\) such that the sum of similarities of all views from the same anchor (\(s_i^+\) and \(s_j^+)\) are higher than the sum of similarities between views from different instances (\(s_{ij}^-\)): \[s_i^+ + s_j^+ > 2s_{ij}^-+ \epsilon \quad \forall i\neq j\] where \(\epsilon\) is a margin that we suppose \"very big\" (see hereafter). Indeed, this inequality is equivalent to \(-\epsilon > 2s_{ij}^--s_i^+-s_j^+\) for all \(i\neq j\), which can be written as: \[\argmin_f \max(-\epsilon, \{2s_{ij}^--s_i^+-s_j^+\}_{i,j\in [1..n], j\neq i})\] This can be transformed into an optimization problem using the LSE (log-sum-exp) approximation of the \(\max\) operator: \[\argmin_f \log\left(\exp(-\epsilon) + \sum_{i\neq j} \exp{(-s_i^+-s_j^+ + 2s_{ij}^-)}\right)\] Thus, if we use an infinite margin (\(\lim_{\epsilon \rightarrow \infty}\)) we retrieve exactly our optimization problem with Decoupled Uniformity in Eq.[\[eq:proofLdunif\]](#eq:proofLdunif){reference-type="ref" reference="eq:proofLdunif"} (up to an additional constant depending on \(n\)).
# Additional general guarantees on downstream classification
## Optimal configuration of supervised loss
In order to derive guarantees on a downstream classification task \(\cD\) when optimizing our unsupervised decoupled uniformity loss, we define a supervised loss that measures the risk on a downstream supervised task. We prove in the next section that the minimizers of this loss have the same geometry as the ones minimizing cross-entropy and SupCon : a regular simplex on the hyper-sphere . More formally, we have:
This property notably implies that we can realize 100% accuracy at optima with linear evaluation (taking the linear classifier \(g(\bar{x})=W^*f^*(\bar{x})\) with \(W^*=(\mu_y^*)_{y\in \cY}\in \bbR^{C\times d}\)).
## General guarantees of Decoupled Uniformity
In its most general formulation, we tightly bound the previous supervised loss by Decoupled Uniformity loss \(\cL_{unif}^d\) depending on a variance term of the centroids \(\mu_{\bar{x}}\) conditionally to the labels:
Intuitively, it means that we will achieve good accuracy if all centroids \((\mu_{\bar{x}})_{\bar{x}\in \bar{\cX}}\) for samples \(\bar{x}\in \bar{\cX}\) in the same class are not too far. This theorem is very general since we do not require the intra-class connectivity assumption on \(\cA\); so any \(\cA \subset \cA^*\) can be used.
# Experimental Details
Code will be released upon acceptance of the manuscript. We provide a detailed pseudo-code of our algorithm as well as all experimental details to reproduce the experiments run in the manuscript.
## Pseudo-code
## Datasets
#### CIFAR
We use the original training/test split with 50000 and 10000 images respectively of size \(32\times 32\).
#### STL-10
In unsupervised pre-training, we use all labelled+unlabelled images (105000 images) for training and the remaining 8000 for test with size \(96\times 96\). During linear evaluation, we only use the 5000 training labelled images for learning the weights.
#### CUB200-2011
This dataset is composed of 200 fine-grained bird species with 5994 training images and 5794 test images rescaled to \(224\times 224\).
#### UTZappos
This dataset is composed of images of shoes from zappos.com. In order to be comparable with the literature on weakly supervised learning, we follow and split it into 35017 training images and 15008 test images resized at \(32\times 32\).
#### ImageNet100
It is a subset of ImageNet containing 100 random classes and introduced in . It contains 126689 training images and 5000 testing images rescaled to \(224\times 224\). It notably allows a reasonable computational time since we runt all our experiments on a single server node with 4 V100 GPU.
#### BHB
This dataset is composed of 10420 3D brain MRI images of size \(121\times 145\times 121\) with \(1.5mm^3\) spatial resolution. Only healthy subjects are included.
#### BIOBD
It is also a brain MRI dataset including 662 3D anatomical images and used for downstream classification. Each 3D volume has size \(121\times 145\times 121\). It contains 306 patients with bipolar disorder vs 356 healthy controls and we aim at discriminating patients vs controls. It is particularly suited to investigate biomarkers discovery inside the brain .
## Contrastive Models
#### Architecture.
For all small-scale vision datasets (CIFAR-10 , CIFAR-100 , STL-10 , CUB200-2011 and UT-Zappos ), we used official ResNet18 backbone where we replaced the first \(7\times 7\) convolutional kernel by a smaller \(3\times 3\) kernel and we removed the first max-pooling layer for CIFAR-10, CIFAR-100 and UTZappos. For ImageNet100, we used ResNet50 for stronger baselines as it is common in the literature. For medical images on brain MRI datasets (BHB and BIOBD, we used DenseNet121 as our default backbone encoder, following previous literature on these datasets . Following , we use the representation space after the last average pooling layer with 2048 dimensions to perform linear evaluation and use a 2-layers MLP projection head with batch normalization between each layer for a final latent space with 128 dimensions.
#### Batch size.
We always use a default batch size 256 for all experiments on vision datasets and 64 for brain MRI datasets (considering the computational cost with 3D images and since it had little impact on the performance ).
#### Optimization.
We use SGD optimizer on small-scale vision datasets (CIFAR-10, CIFAR-100, STL-10, CUB200-2011, UT-Zappos) with a base learning rate \(0.3\times \text{batch size}/256\) and a cosine scheduler. For ImageNet100, we use a LARS optimizer with learning rate \(0.02\times \sqrt{\text{batch size}}\) and cosine scheduler. In Kernel Decoupled Uniformity loss, we set \(\lambda=\frac{0.01}{\sqrt{\text{batch size}}}\) and \(t=2\). For SimCLR, we set the temperature to \(\tau=0.07\) for all datasets following. Unless mentioned otherwise, we use 2 views for Decoupled Uniformity (both with and without kernel) and the computational cost remains comparable with standard contrastive models.
#### Training epochs.
By default, we train the models for 200 epochs unless mentioned otherwise for all vision data-sets excepted CUB200-2011 and UTZappos where we train them for 1000 epochs, following . For medical datasets, we perform pre-training for 50 epochs, as in. For linear evaluation, we use a simple linear layer trained for 300 epochs with an initial learning rate 0.1 decayed by 0.1 on each plateau.
#### Augmentations.
We follow to define our full set of data augmentations for vision datasets including: *RandomResizedCrop* (uniform scale between 0.08 to 1), *RandomHorizontalFlip* and color distorsion (including color jittering and gray-scale). For medical datasets, we use cutout covering 25% of the image in each direction (\(1/4^3\) of the entire volume), following .
### Generative Models
#### Architecture.
For VAE, we use ResNet18 backbone with a completely symmetric decoder using nearest-neighbor interpolation for up-sampling. For DCGAN, we follow the architecture described in . We keep the original dimension for CIFAR-10 and CIFAR-100 datasets and we resize the images to \(64\times 64\) for STL-10. For BigBiGAN , we use the ResNet50 pre-trained encoder available at <https://tfhub.dev/deepmind/bigbigan-resnet50/1> with BN+CReLU features.
#### Training.
For VAE, we use PyTorch-lightning pre-trained model for STL-10 [^1] and we optimize VAE for CIFAR-10 and CIFAR-100 for 400 epochs using an initial learning rate \(10^{-4}\) and SGD optimizer with a cosine scheduler. We use the same pipeline on RandBits dataset. For DCGAN, we optimize it using Adam optimizer (following ) and base learning rate \(2\times 10^{-4}\).
# Omitted Proofs
## Estimation Error with Empirical Decoupled Uniformity
## Optimality of Decoupled Uniformity
## Optimality of Supervised Loss
## Generalization bounds for decoupled uniformity
## Generalization bound under intra-class connectivity assumption
## Conditional Mean Embedding Estimation
Let \(f\in \cF\) fixed.
## Generalization bound under extended intra-class connectivity hypothesis
Then, by definition of \(\hat{\mu}_{\bar{u}_j}\): \[\begin{aligned}
||\hat{\mu}_{\bar{u}_{j+1}}-\hat{\mu}_{\bar{u}_{j}}|| &= ||\sum_{k,l=1}^n \alpha_{k,l}K(\bar{x}_l, \bar{u}_{j+1}) f(x_k)-\sum_{k,l=1}^n \alpha_{k,l}K(\bar{x}_l, \bar{u}_{j}) f(x_k)||\\ &= ||A C|| \end{aligned}\] Where \(A = (\sum_{k=1}^n\alpha_{kj}f(x_k)^i)_{i,j}\in \bbR^{d\times n}\) (\(f(\cdot)^i\) is the i-th component of \(f(\cdot)\)) and \(C=(K(\bar{x}_l, \bar{u}_{j+1})-K(\bar{x}_l, \bar{u}_j))_{l}\in \bbR^{n\times 1}\). So, using the property of spectral \(\ell_2\) norm we have: \[\begin{aligned}
||\hat{\mu}_{\bar{u}_{j+1}}-\hat{\mu}_{\bar{u}_{j}}|| &= ||A C|| \le ||A||_2 ||C||_2\\ \end{aligned}\] Using the previous lemma and because \((\bar{u}_j, \bar{u}_{j+1})\in E_K\), we have: \(||C||_2^2=\sum_{i=1}^n(K(\bar{x}_i, \bar{u}_{j+1})-K(\bar{x}_i, \bar{u}_j))^2 \le \sum_{i=1}^n (\max(K(\bar{u}_{j+1}, \bar{u}_{j+1}), K(\bar{u}_{j}, \bar{u}_{j}))-K(\bar{u}_{j}, \bar{u}_{j+1}))^2 \le n\epsilon^2\). To conclude, we will prove that \(||A||_2 \le ||\mathbf{\alpha}||_2\) where \(\mathbf{\alpha} = (\alpha_{ij})_{i,j\in [1..n]^2}\). For any \(v\in \bbR^{n}\), we have: \[\begin{aligned}
||Av||^2 &= ||\sum_{k,j=1}^n \alpha_{k,j}v_j f(x_k)||^2 \stackrel{(1)}{\le} \left(\sum_{k,j=1}^n \alpha_{k,j}v_j\right)^2 = ||\alpha v ||^2 \stackrel{(2)}{\le} ||\alpha||_2^2 ||v||^2 \end{aligned}\] Where (1) holds with Cauchy-Schwarz inequality and because \(f(\cdot)\in \bbS^{d-1}\) and (2) holds by definition of spectral \(\ell_2\) norm. So we have \(\forall v\in \bbR^d, ||Av||\le ||\alpha||_2 ||v||\), showing that \(||A||_2 \le ||\alpha||_2\). So we can conclude that: \[\begin{aligned}
\sum_{j\in J}||\mu_{\bar{u}_{j+1}}-\mu_{\bar{u}_j}||& \le \sum_{j\in J} \left(\sqrt{n}||(K_n+\lambda n\mathbf{I}_n)^{-1}||_2 \epsilon + O(n^{-1/4}) \right) = |J|||(K_n+n\lambda \mathbf{I}_n)^{-1}||_2\sqrt{n} \epsilon + O(n^{-1/4}) \end{aligned}\] We set \(\beta_n(K_n)=\sqrt{n}||(K_n+\lambda n \mathbf{I}_n)^{-1}||_2\). In order to see that \(\beta_n(K_n) = (\frac{\lambda_{min}(K_n)}{\sqrt{n}} + \sqrt{n}\lambda)^{-1}\) with \(\lambda_{min}(K_n)>0\) the minimum eigenvalue of \(K_n\), we apply the spectral theorem on the symmetric definite-positive kernel matrix \(K_n\). Let \(0<\lambda_1 \le \lambda_2\le...\le \lambda_n\) the eigenvalues of \(K_n\). According to the spectral theorem, it exists \(U\) an unitary matrix such that \(K_n=UDU^T\) with \(D=\diag(\lambda_1,..., \lambda_n)\). So, by definition of spectral norm: \[\begin{aligned}
||(K_n+n\lambda \mathbf{I}_n)^{-1}||_2^2 &= \lambda_{max}\left(U(D+n\lambda \mathbf{I}_n)^{-1}U^T U(D+\lambda n\mathbf{I}_n)^{-1}U^T\right)\\ &= \lambda_{max}(U\tilde{D}U^T)\\ &= (\lambda_1 + n\lambda)^{-2} \end{aligned}\] where \(\tilde{D} = \diag(\frac{1}{(\lambda_1+n\lambda)^2},..., \frac{1}{(\lambda_n+n\lambda)^2})\). So we can conclude that \(\beta_n(K_n) = (\frac{\lambda_1}{\sqrt{n}} + \sqrt{n}\lambda)^{-1}=O(1)\) for \(\lambda=O(\frac{1}{\sqrt{n}})\).\
Finally, by pooling inequalities for edges over \(E\) and \(E_K\), we have: \[\begin{aligned}
||\mu_{\bar{x}}-\mu_{\bar{x}'}|| \le 2\epsilon' |I| + |J|\beta_n(K_n) \epsilon + O(n^{-1/4})\le D (2\epsilon' + \beta_n(K_n)\epsilon) + O(n^{-1/4}) \end{aligned}\] We can conclude by plugging this inequality in Theorem [\[th:downstream_gen\]](#th:downstream_gen){reference-type="ref" reference="th:downstream_gen"}. ◻
# Checklist {#checklist .unnumbered}
The checklist follows the references. Please read the checklist guidelines carefully for information on how to answer these questions. For each question, change the default to , , or . You are strongly encouraged to include a **justification to your answer**, either by referencing the appropriate section of your paper or providing a brief inline description. For example:
- Did you include the license to the code and datasets?
- Did you include the license to the code and datasets?
- Did you include the license to the code and datasets?
Please do not modify the questions and only use the provided macros for your answers. Note that the Checklist section does not count towards the page limit. In your paper, please delete this instructions block and only keep the Checklist section heading above along with the questions/answers below.
1. For all authors\...
1. Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
2. Did you describe the limitations of your work?
3. Did you discuss any potential negative societal impacts of your work?
4. Have you read the ethics review guidelines and ensured that your paper conforms to them?
2. If you are including theoretical results\...
1. Did you state the full set of assumptions of all theoretical results?
2. Did you include complete proofs of all theoretical results?
3. If you ran experiments\...
1. Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
2. Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
3. Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
4. Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets\...
1. If your work uses existing assets, did you cite the creators?
2. Did you mention the license of the assets?
3. Did you include any new assets either in the supplemental material or as a URL?
4. Did you discuss whether and how consent was obtained from people whose data you're using/curating?
5. Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
5. If you used crowdsourcing or conducted research with human subjects\...
1. Did you include the full text of instructions given to participants and screenshots, if applicable?
2. Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
3. Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
# Introduction
Contrastive Learning (CL) is a paradigm designed for representation learning which has been applied to unsupervised, weakly supervised and supervised problems. It gained popularity during the last years by achieving impressive results in the unsupervised setting on standard vision datasets (*e.g.* ImageNet) where it almost matched the performance of its supervised counterpart. The objective in CL is to increase the similarity in the representation space between *positive* samples (semantically close) while decreasing the similarity between semantically distinct *negative* samples. Despite its simple formulation, it requires the definition of a similarity function (that can be seen as an energy term ), positives and negatives. Similarity functions are defined on latent representations of an encoder \(f\in \cF\) (CNN or Transformer for vision datasets) and and is usually expressed as a Euclidean scalar product (*e.g.* InfoNCE). In supervised learning , positives are simply images belonging to the same class while negatives are images belonging to different classes. In unsupervised learning , since labels are unknown, positives are usually defined as transformed versions (*views*) of the same original image (a.k.a. the anchor) and negatives are the transformed versions of all other images. As a result, the augmentation distribution \(\cA\) used to sample both positives and negatives is crucial and it conditions the quality of the learnt representation. The most-used augmentations for visual representations involve aggressive crop and color distortion. Cropping induces representations with high occlusion invariance while color distortion may avoids the encoder \(f\) to take a shortcut while aligning positive sample representations and fall into the simplicity bias. Nevertheless, learning a representation that mainly relies on augmentations comes at a cost: both crop and color distortion induce strong biases in the final representation . Specifically, dominant objects inside images can prevent the model from learning features of smaller objects (which is not apparent in object-centric datasets such as ImageNet) and few, irrelevant and easy-to-learn features, that are shared among views, are sufficient to collapse the representation (a.k.a feature suppression). Finding the right augmentations in other visual domains, such as medical imaging, remains an open challenge since the anatomical differences between two classes (e.g., pathological and healthy) can be quite subtle and difficult to capture with usual transformations. If the augmentations are too weak or inadequate to remove irrelevant signal w.r.t. a discrimination task, then how can we define positive samples? In our work, we propose to integrate *prior information*, learnt from generative models or given as auxiliary attributes, into contrastive learning, to make it less dependent on data augmentation. Using the theoretical understanding of CL through the augmentation graph, we make the connection with kernel theory and introduce a novel loss with theoretical guarantees on downstream performance. Prior information is integrated into the proposed contrastive loss using a kernel. In the unsupervised setting, we leverage pre-trained generative models, such as GAN and VAE , to learn *a prior representation* of the data. We provide a solution to the feature suppression issue in CL and also demonstrate SOTA results with weaker augmentations on visual benchmarks. In the weakly supervised setting, we use instead auxiliary/prior information, such as image attributes (e.g. birds color or size) and we show better performance than previous conditional formulations based on these attributes . In summary, we make the following contributions: 1. We propose a new framework for contrastive learning allowing the integration of prior information, learnt from generative models or given as auxiliary attributes, into the positive sampling\
2. We derive theoretical bounds on the downstream classification risk that rely on weaker assumptions for data augmentations than previous works on CL\
3. We empirically show that our framework can benefit from the latest advances of generative models to learn a better representation while relying on less augmentations\
4. We show that we achieve SOTA results in the unsupervised and weakly supervised setting
# Related Works
In a weakly supervised setting, recent studies have shown that positive samples can be defined conditionally to an auxiliary attribute in order to improve the final representation, in particular for medical imaging. From an information bottleneck perspective, these approaches essentially compress the representation to be only predictive of the auxiliary attributes. This might harm the performance of the model when these attributes are too noisy to accurately approximate the true semantic labels for a given downstream task. Furthermore, the benefit of adding data augmentations when conditioning with respect to these attributes is not theoretically discussed. In an unsupervised setting, recent approaches used the encoder \(f\), learnt during optimization, to extend the positive sampling procedure to other views of different instances (*i.e.* distinct from the anchor) that are close to the anchor in the latent space. In order to avoid representation collapse, a support set , a momentum encoder or another small network can be used to select the positive samples. In clustering approaches , distinct instances with close semantics are attracted in the latent space using prototypes. These prototypes can be estimated through K-means or Sinkhorn-Knopp algorithm . All these methods rely on the past representation of a network to improve the current one. They require strong augmentations and they essentially assume that the closest points in the representation space belong to the same latent class in order to to better select the positives. This inductive bias is still poorly understood theoretically and may depend on the visual domain. Our work also relates to generative models for learning representations. VAE learns the data distribution by mapping each input to a Gaussian distribution that we can easily sample to reconstruct the original image. GAN instead sample directly from a Gaussian distribution to generate images that are classified by a discriminator in a min-max game. The discriminator representation can then be used as feature extractor. Other models (ALI , BiGAN and BigBiGAN) learn simultaneously a generator and an encoder that can be used directly for representation learning. All these models do not require particular augmentations to model the data distribution but they perform generally poorer than recent discriminative approaches for representation learning. A first connection between generative models and contrastive learning has emerged very recently . In , authors study the feasibility of learning effective visual representations using only generated samples, and not real ones, with a contrastive loss. Their empirical analysis is complementary to our work. Here, we leverage the representation capacity of the generative models, rather than their generative power, to learn prior representation of the data.
# Constrastive Learning with Decoupled Uniformity
## Problem setup
The general problem in contrastive learning is to learn a data representation using an encoder \(f\in \cF: \cX \rightarrow \bbS^{d-1}\) that is pre-trained with a set of \(n\) original samples \((\bar{x}_i)_{i\in [1..n]}\in \bar{\cX}\), sampled from the data distribution \(p(\bar{x})\)[^2] These samples are transformed to generate *positive samples* (*i.e.*, semantically similar to \(\bar{x}\)) in \(\cX\), space of augmented images, using a distribution of augmentations \(\cA(\cdot |\bar{x})\). Concretely, for each \(\bar{x}_i\), we can sample views of \(\bar{x}_i\) using \(x\sim \cA(\cdot |\bar{x}_i)\) (*e.g.*, by applying color jittering, flip or crop with a given probability). For consistency, we assume \(\cA(\bar{x})=p(\bar{x})\) so that the distributions \(\cA(\cdot |\bar{x})\) and \(p(\bar{x})\) induce a marginal distribution \(p(x)\) over \(\cX\). Given an anchor \(\bar{x}_i\), all views \(x\sim\cA(\cdot |\bar{x}_j)\) from different samples \(\bar{x}_{j\neq i}\) are considered as *negatives*. Once pre-trained, the encoder \(f\) is fixed and its representation \(f(\bar{\cX})\) is evaluated through linear evaluation on a classification task using a labeled dataset \(\cD=\{(\bar{x}_i, y_i)\}\in \bar{\cX}\times \cY\) where \(\cY=[1..K]\), with \(K\) the number of classes.\
**Linear evaluation.** To evaluate the representation of \(f\) on a classification task, we train a linear classifier \(g(\bar{x})=Wf(\bar{x})\) (\(f\) is fixed) that minimizes the multi-class classification error.
## Objective
The popular InfoNCE loss , often used in CL, imposes 1) alignment between positives and 2) uniformity between the views (\(x \overset{\text{i.i.d.}}{\sim} \cA(\cdot|\bar{x})\)) of all instances \(\bar{x}\) --two properties that correlate well with downstream performance. However, by imposing uniformity between *all* views, we essentially try to both attract (alignment) and repel (uniformity) positive samples and therefore we cannot achieve a perfect alignment *and* uniformity, as noted in . Moreover, InfoNCE has been originally designed for only two views (i.e., one couple of positive) and its extension to multiple views is not straightforward. Previous works have proposed a solution to either the first or second issue. Here, we propose a modified version of the uniformity loss, presented in , that solves both issues since it: i) decouples positives from negatives, similarly to and ii) is generalizable to multi-views as in . We introduce the Decoupled Uniformity loss as: \[\cL_{unif}^d(f) = \log \bbE_{p(\bar{x})p(\bar{x}')}e^{-||\mu_{\bar{x}}-\mu_{\bar{x}'}||^2} \label{def:decoupled_unif}\] where \(\mu_{\bar{x}}=\bbE_{\cA(x|\bar{x})}f(x)\) is called a *centroid* of the views of \(\bar{x}\). This loss essentially repels distinct centroids \(\mu_{\bar{x}}\) through an average pairwise Gaussian potential. Interestingly, it implicitly optimizes alignment between positives through the maximization of \(||\mu_{\bar{x}}||\)[^3], so we do not need to explicitly add an alignment term. It can be shown (see Appendix), that minimizing this loss brings to a representation space where the sum of similarities between views of the same sample is greater than the sum of similarities between views of different samples. From a physics point-of-view, we are trying to find the equilibrium state of \(|\bar{\cX}|\) particles linked with a pair Gaussian potential energy. We will study its main properties hereafter and we will see that *prior* information can be added during the estimation step of these centroids.
## Geometrical Analysis of Decoupled Uniformity
Theorem [\[th:optimal_decoupled_unif\]](#th:optimal_decoupled_unif){reference-type="ref" reference="th:optimal_decoupled_unif"} gives a complete geometrical characterization when the batch size \(n\) set during training is not too large compared to the representation space dimension \(d\). By removing the coupling between positives and negatives, we see that Decoupled Uniformity can realize both perfect alignment and uniformity, contrary to InfoNCE.\
**Remark.** The assumption \(n \le d+1\) is crucial to have the existence of a regular simplex on the hypersphere \(\bbS^{d-1}\). In practice, this condition is not always full-filled (e.g SimCLR with \(d=128\) and \(n=4096\)). Characterizing the optimal solution of \(\cL_{unif}^d\) for any \(n>d+1\) is still an open problem but theoretical guarantees can be obtained in the limit case \(n\rightarrow \infty\) (see below).
Empirically, we observe that minimizers \(f\) of \(\hat{\cL}_{unif}^d\) remain well-aligned when \(n>d+1\) on real-world vision datasets (see Appendix). Decoupled uniformity thus optimizes two properties that are nicely correlated with downstream classification performance --that is alignment and uniformity between centroids. However, as noted in , optimizing these two properties is necessary but not sufficient to guarantee a good classification accuracy. In fact, the accuracy can be arbitrary bad even for perfectly aligned and uniform encoders . Most recent theories about CL make the hypothesis that samples from the same semantic class have overlapping augmented views to provide guarantees on the downstream task when optimizing InfoNCE or Spectral Contrastive loss. This assumption, known as intra-class connectivity hypothesis, is very strong and only relies on the augmentation distribution \(\cA\). In particular, augmentations should not be \"too weak\", so that all intra-class samples are connected among them, and at the same time not \"too strong\", to prevent connections between inter-class samples and thus preserve the semantic information. Here, we prove that we can relax this hypothesis if we can provide a kernel (viewed as a similarity function between original samples \(\bar{x}\)) that is \"good enough\" to relate intra-class samples not connected by the augmentations (see Fig. [\[fig:extended_aug_graph\]](#fig:extended_aug_graph){reference-type="ref" reference="fig:extended_aug_graph"}). In practice, we show that generative models can define such kernel. We first recall the definition of the augmentation graph , and intra-class connectivity hypothesis before presenting our main theorems. For simplicity, we assume that the set of images \(\bar{\cX}\) is finite (similarly to ). Our bounds and theoretical guarantees will never depend on the cardinality \(|\bar{\cX}|\).
## Intra-class connectivity hypothesis
Previous analysis in CL make the hypothesis that it exists an optimal (accessible) augmentation module \(\cA^*\) that fulfills:
Under this hypothesis, Decoupled Uniformity loss can also tightly bound the downstream supervised risk for a bigger class of encoders than prior work . To show it, we define a measure of the risk on a downstream task \(\cD\). While previous analysis generally used the mean cross-entropy loss (as it has closer analytic form with InfoNCE), we use a supervised loss closer to decoupled uniformity with the same guarantees as the mean cross-entropy loss (see Appendix). Notably, the geometry of the representation space at optimum is the same as cross-entropy and SupCon and we can theoretically achieve perfect linear classification.
#### Remark.
This loss depends on centroids \(\mu_{\bar{x}}\) rather than \(f(\bar{x})\). Empirically, it has been shown that performing feature averaging gives better performance on the downstream task.
In practice, the diameter \(D\) can be controlled by a small constant in some cases (typically \(\le 4\)) but it remains specific to the dataset at hand. Furthermore, we observe (see Appendix) that \(f\) realizes alignment with small error \(\epsilon\) during optimization of \(\cL_{unif}^d(f)\) for augmentations close to the sweet spot \(\cA^*\) on CIFAR-10 and CIFAR-100. In the next section, we study the case when \(\cA^*\) is not accessible or very hard to find.
# Reconnect the disconnected: extending the augmentation graph with kernel
Having access to optimal augmentations is a strong assumption and, for many real-world applications (e.g medical imaging ), it may not be accessible. If we have only weak augmentations (*e.g.*, \(\supp\cA(\cdot |\bar{x}) \subsetneq \supp\cA^*(\cdot|\bar{x})\) for any \(\bar{x}\)), then some intra-class points might not be connected and we would need to reconnect them to ensure good downstream accuracy (see Theorem [\[th:downstream_gen\]](#th:downstream_gen){reference-type="ref" reference="th:downstream_gen"} in Appendix). Augmentations are intuitive and they have been hand-crafted for decades by using human perception (*e.g.*, a rotated chair remains a chair and a gray-scale dog is still a dog). However, we may know other *prior information* about objects that are difficult to transfer through invariance to augmentations (*e.g.*, chairs should have 4 legs). This prior information can be either given as image attributes (*e.g.*, age or sex of a person, color of a bird, etc.) or, in an unsupervised setting, directly learnt through a generative model (*e.g.*, GAN or VAE). Now, we ask: how can we integrate this information inside a contrastive framework to reconnect intra-class images that are actually disconnected in \(G_{\cA}\)? We rely on conditional mean embedding theory and use a kernel defined on the prior representation/information. This allows us to estimate a better configuration of the centroids in the representation space, with respect to the downstream task, and, ultimately, provide theoretical guarantees on the classification risk.
## \(\epsilon\)-Kernel Graph
**Example.** If we work with large natural images, assuming that we know a prior \(z(\bar{x})\) about our images (*e.g.*, given by a generative model), we can compute \(K_{\bar{\cX}}\) using \(z\) through \(K_{\bar{\cX}}(\bar{x}, \bar{x}')=\tilde{K}(z(\bar{x}), z(\bar{x}'))\) where \(\tilde{K}\) is a standard kernel (*e.g.*, Gaussian or Cosine). To link kernel theory with the previous augmentation graph, we need to define a *kernel graph* that connects images with high similarity in the kernel space.
The condition \(\max(K(\bar{x}, \bar{x}), K(\bar{x}', \bar{x}'))-K(\bar{x}, \bar{x}')\le \epsilon\) implies that \(d_K(\bar{x}, \bar{x}')\le 2\epsilon\) where \(d_K(\bar{x}, \bar{x}') = K(\bar{x}, \bar{x}) + K(\bar{x}', \bar{x}')-2K(\bar{x}, \bar{x'})\) is the kernel distance. For kernels with constant norm (*e.g.*, the standard Gaussian, Cosine or Laplacian kernel), it is in fact an equivalence. Intuitively, it means that we connect two original points in the kernel graph if they have small distance in the kernel space. We give now our main assumption to derive a better estimator of the centroid \(\mu_{\bar{x}}\) in the insufficient augmentation regime.
This assumption is notably weaker than Assumption [\[hyp:embed_intra_class_connect\]](#hyp:embed_intra_class_connect){reference-type="ref" reference="hyp:embed_intra_class_connect"} w.r.t augmentation distribution \(\cA\). Here, we do not need to find the optimal distribution \(\cA^*\) as long as we have a kernel \(K_{\bar{\cX}}\) such that disconnected points in the augmentation graph are connected in the \(\epsilon\)-kernel graph. If \(K\) is not well adapted to the data-set (i.e it gives very low values for intra-class points), then \(\epsilon\) needs to be large to re-connect these points and we will see that the classification error will be high. In practice, this means that we need to tune the hyper-parameter of the kernel (i.e., \(\sigma\) for a RBF kernel) so that all intra-class points are reconnected with a small \(\epsilon\).
## Conditional Mean Embedding
Decoupled Uniformity loss includes no kernel in its raw form. It only depends on centroids \(\mu_{\bar{x}}=\bbE_{\cA(x|\bar{x})} f(x)\). Here, we show that another consistent estimator of these centroids can be defined, using the previous kernel \(K_{\bar{\cX}}\). To show it, we **fix** an encoder \(f\in \cF\) and require the following technical assumption in order to apply conditional mean embedding theory .
**Intuition.** This theorem says that we can use representations of images close to an anchor \(\bar{x}\), according to our prior information, to accurately estimate \(\mu_{\bar{x}}\). Consequently, if the prior is \"good enough\" to connect intra-class images disconnected in the augmentation graph (i.e. fulfills Assumption [\[hyp:extended_intra_class_connectivity\]](#hyp:extended_intra_class_connectivity){reference-type="ref" reference="hyp:extended_intra_class_connectivity"}), then this estimator allows us to tightly control the classification risk. From this theorem, we naturally derive the empirical Kernel Decoupled Uniformity loss using the previous estimator.
#### Extension to multi-views.
If we have \(V\) views \((x_i^{(v)})_{v\in [1..V]}\) for each \(\bar{x}_i\), we can easily extend the previous estimator with \(\hat{\mu}_{\bar{x}_i}=\frac{1}{V}\sum_{v=1}^V \hat{\mu}_{\bar{x}_j}^{(v)}\) where \(\hat{\mu}_{\bar{x}_j}=\sum_{i=1}^n \alpha_{i,j}f(x_i^{(v)})\). The computational cost added is roughly \(O(n^3)\) (to compute the inverse matrix of size \(n\times n\)) but it remains negligible compared to the back-propagation time using classical stochastic gradient descent. Importantly, the gradients associated to \(\alpha_{i,j}\) are not computed.
## Contrastive Learning with Kernel
We show here that \(\hat{\cL}_{unif}^d(f)\) can tightly bound the supervised classification risk for well-aligned encoders \(f\in \cF\).
**Interpretation.** Theorem [\[th:final_bound_decoupled\]](#th:final_bound_decoupled){reference-type="ref" reference="th:final_bound_decoupled"} gives a tight bound on the classification loss \(\cL_{sup}(f)\) with few assumptions. In the special case \(\epsilon=0\) and \(\cA=\cA^*\) (i.e the augmentation graph is class-connected, a stronger assumption than [\[hyp:extended_intra_class_connectivity\]](#hyp:extended_intra_class_connectivity){reference-type="ref" reference="hyp:extended_intra_class_connectivity"}), we retrieve the standard bounds of Theorem [\[th:boundness_intra_class_hyp\]](#th:boundness_intra_class_hyp){reference-type="ref" reference="th:boundness_intra_class_hyp"}. As before, we don't require perfect alignment for \(f\in \cF\) and we don't have class collision term (even if the extended augmentation graph may contain edges between inter-class samples), contrarily to. Also, the estimation error doesn't depend on the number of views (which is low in practice))--as it was always the case in previous formulations --but rather on the batch size \(n\). Contrarily to CCLK , we don't condition our representation to weak attributes but rather we provide better estimation of the conditional mean embedding conditionally to the original image. Our loss remains in an unconditional contrastive framework driven by the augmentations \(\cA\) and the prior \(K_{\bar{\cX}}\) on input images.
# Experiments
Here, we study several problems where current contrastive frameworks fail at learning robust visual representations. In unsupervised learning, we show that we can leverage generative models to outperform current self-supervised models when the augmentations are insufficient to remove irrelevant signal from images. In weakly supervised, we demonstrate the superiority of our unconditional formulation when noisy auxiliary attributes are available. Implementation details in Appendix.
## Generative models as prior-Evading feature suppression
Previous investigations have shown that a few easy-to-learn irrelevant features not removed by augmentations can prevent the model from learning all semantic features inside images. We propose here a first solution to this issue.\
**RandBits dataset .** We build a RandBits dataset based on CIFAR-10. For each image, we add a random integer sampled in \([0, 2^k-1]\) where \(k\) is a controllable number of bits. To make it easy to learn, we take its binary representation and repeat it to define \(k\) channels that are added to the original RGB channels. Importantly, these channels will not be altered by augmentations, so they will be shared across views. We train a ResNet18 on this dataset with standard SimCLR augmentations and varying \(k\). For kernel decoupled uniformity, we use a \(\beta\)-VAE representation (ResNet18 backbone, \(\beta=1\)) to define \(K_{VAE}(\bar{x},\bar{x}')=K(\mu(\bar{x}), \mu(\bar{x}'))\) where \(\mu(\cdot)\) is the mean Gaussian distribution of \(\bar{x}\) in the VAE latent space and \(K\) is a standard RBF kernel.
## Filling the gap for medical imaging
Data augmentations on natural images have been handcrafted over decades to achieve current performance on ImageNet. However, they are not always adapted to medical datasets . We study bipolar disorder detection (BD), a challenging binary classification task, on the brain MRI dataset BIOBD . It contains 356 healthy controls (HC) and 306 patients with BD. We use BHB as a large pre-training dataset containing 10k 3D images of healthy subjects. While reconstruction-based model such as Model Genesis is SOTA for medical images, we show that contrastive approaches combined with VAE provide a new way to tackle hard classification problems.
[\[tab:medical_imaging\]]{#tab:medical_imaging label="tab:medical_imaging"}
## Weakly Supervised Learning
Now we assume to have access to image attributes that correlate well with true semantic labels (e.g birds color or size for birds classification). We use three datasets: CUB-200-2011 , ImageNet100 and UTZappos , following . CUB-200-2011 contains 11788 images of 200 bird species with 312 binary attributes available (encoding size, wing shape, color, etc.). UTZappos contains 50025 images of shoes from several brands sub-categorized into 21 groups that we use as downstream classification labels. It comes with 7 attributes. Finally, for ImageNet100 we follow and use the pre-trained CLIP model (trained on pairs (text, image)) to extract 512-d features considered as prior information. We compare our method with CCLK, a conditional contrastive model that defines positive samples only according to the conditioning attributes.
# Conclusion
In this work, we have showed that we can integrate prior information into CL to improve the final representation. In particular, we draw connections between kernel theory and CL to build our theoretical framework. We demonstrate tight bounds on downstream classification performance with weaker assumptions than previous works. Empirically, we show that generative models provide a good prior when augmentations are too weak or insufficient to remove easy-to-learn noisy features. We also show applications in medical imaging and in the weakly supervised setting. We hope that CL will benefit from the future progress in generative modelling with our theoretical framework.
[^1]: <https://github.com/PyTorchLightning/pytorch-lightning>
[^2]: With an abuse of notation, we define it as \(p(\bar{x})\) instead than \(p_{\bar{X}}\) to simplify the presentation, as it is common in the literature
[^3]: By Jensen's inequality \(||\mu_{\bar{x}}||\le\bbE_{\cA(x|\bar{x})}||f(x)||=1\) with equality iff \(f\) is constant on \(\supp\cA(\cdot|\bar{x})\). | {'timestamp': '2022-06-06T02:16:39', 'yymm': '2206', 'arxiv_id': '2206.01646', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01646'} |
null | null |
# Introduction
The sample efficiency of reinforcement learning (RL) algorithms crucially depends on the representation of the underlying system state they operate on. Sometimes, a low-dimensional (direct) representation of the state, such as the positions of the objects in the environment, is considered to make the resulting RL problem most efficient. However, such low-dimensional, direct state representations can have several disadvantages. On the one hand, a perception module, e.g., pose estimation, is necessary in the real world to obtain the representation from raw observations, which often is difficult to achieve in practice with sufficient robustness. On the other hand, if the goal is to learn policies that generalize over different object shapes, using a low-dimensional state representation is often impractical. Such scenarios, while challenging for RL, are common, e.g., in robotic manipulation tasks. Therefore, there is a large history of approaches that consider RL directly from raw, high-dimensional observations like images (e.g., ). Typically, an encoder takes the high-dimensional input and maps it to a low-dimensional latent representation of the state. The RL algorithm (e.g., the Q-function or the policy network) then operates on the latent vector as state input. This way, no separate perception module is necessary, the framework can extract information from the raw observations that are relevant for the task, and the RL agent, in principle, may generalize over challenging environments, such as if the object shapes are varied. While these are advantages in principle, jointly training encoders capable of processing high-dimensional inputs from the RL signal alone is challenging. To address this, one approach is to *pretrain* the encoder on a different task, e.g., image reconstruction, multi-view consistency, or a time-constrastive task. Alternatively, an auxiliary loss on the latent encoding can be added *during* the RL procedure. In both cases, the choice of the actual (auto-)encoder architecture and associated (auxiliary) loss function has a significant influence on the usefulness of the resulting latent space for the downstream RL task. Especially for image data, convolutional neural networks (CNNs) are commonly used for the encoder. However, 2D CNNs have a 2D (equivariance) bias, while for many RL tasks, the 3D structure of our world is essential. Architectures like Vision Transformers may process images with no such 2D bias, but they may require large scale data, which might be challenging in RL applications. Additionally, although multiple uncalibrated 2D image inputs can be used with generic image encoders, they do not benefit from 3D inductive bias, which may help for example in resolving ambiguities in 2D images such as occlusions and object permanence. Recently, Neural Radiance Fields (NeRFs) have shown great success in learning to represent scenes with a neural network that enables to render the scene from novel viewpoints, and have sparked broad interest in computer vision. NeRFs exhibit a strong 3D inductive bias, leading to better scene reconstruction capabilities than methods composed of generic image encoders (e.g., ). In the present work, we investigate whether incorporating these 3D inductive biases of NeRFs into learning a state representation can benefit RL. Specifically, we propose to train an encoder that maps multiple RGB image views of the scene to a latent representation through an auto-encoder structure, where a (compositional) NeRF decoder provides the self-supervision signal using an image reconstruction loss for each view. In the experiments, we show for multiple environments that supervision from NeRF leads to a latent representation that makes the downstream RL procedure more sample efficient compared to supervision via a 2D CNN decoder, a contrastive loss on the latent space, or even hand-engineered, perfect low-level state information given as keypoints. Commonly, RL is trained on environments where the objects have the same shape. Our environments include hanging mugs on hooks, pushing objects on a table, and a door opening scenario. In all of these, the objects' shapes are not fixed, and we require the agent to generalize over all shapes from a distribution. To summarize our main contributions: (i) we propose to train state representations for RL with NeRF supervision, and (ii) we empirically demonstrate that an encoder trained with a latent-conditioned NeRF decoder, especially with an object-compositional NeRF decoder, leads to increased RL performance relative to standard 2D CNN auto-encoders, contrastive learning, or expert keypoints.
# Related Work
**Neural Scene/Object Representations in Computer Vision, and Applications.** To our knowledge, the present work is the first to explore if neural scene representations like NeRFs can benefit RL. Outside of RL, however, there has been a very active research field in the area of neural scene representations, both in the representations themselves and their applications; see for recent reviews. Within the family of NeRFs and related methods, major thrusts of research have included: improving modeling formulations, modeling larger scenes, addressing (re-)lighting, and an especially active area of research has been in improving speed, both of training and of inference-time rendering. In our case, we are not constrained by inference-time computation issues, since we do not need to render images, and only have to run our latent-space encoder (with a runtime of approx. 7 ms on an RTX3090). Additionally of particular relevance, various methods have developed latent-conditioned or compositional/object-oriented approaches for NeRFs, although they, nor other NeRF-style methods to our knowledge, have been applied to RL. Neural scene representations have found application across many fields (i.e., augmented reality and medical imaging ) and both NeRFs and other neural scene approaches have started to be used for various problems in robotics, including pose estimation, trajectory planning, visual foresight, grasping, and rearrangement tasks. **Learning State Representations for Reinforcement Learning.** One of the key enabling factors for the success of deep RL is its ability to find effective representations of the environment from high-dimensional observation data . Extensive research has gone into investigating different ways to learn better state representations using various auxiliary objective functions. Contrastive learning is a common objective and has shown success in unsupervised representation learning in computer vision applications . Researchers built upon this success and have shown such learning objectives can lead to better performance and sample efficiency in deep RL , where the contrasting signals could come from time alignment , camera viewpoints , and different sensory modalities , with applications in real-world robotic tasks . Extensive efforts have investigated the role of representation learning in RL , provided a detailed analysis of the importance of different visual representation pretraining methods , and shown how we can improve training stability in the face of multiple auxiliary losses . There is also a range of additional explorations on pretraining methods with novel objective functions (e.g., bisimulation metrics and temporal cycle-consistency loss ) and less-explored data sources (e.g., in-the-wild images and action-free videos ). Please check the survey for more related work in this direction . Our method is different in that we explicitly utilize a decoder that includes strong 3D inductive biases provided by NeRFs, which we empirically show improves RL for tasks that depend on the geometry of the objects.
# Background {#sec:background}
## Reinforcement Learning {#sec:RL}
This work considers decision problems that can be described as discrete-time Markov Decision Processes (MDPs) \(M = \langle \mathcal{S}, \mathcal{A}, T, \gamma, R, P_0 \rangle\). \(\mathcal{S}\) and \(\mathcal{A}\) are the sets of all states and actions, respectively. The transition probability (density) from \(s\) to \(s'\) using action \(a\) is \(T(s'\mid s,a)\). The agent receives a real-valued reward \(R(s,a,s')\) after each step. The discount factor \(\gamma \in [0,1)\) trades off immediate and future rewards. \(P_0: \mathcal{S} \rightarrow \mathbb{R}_0^+\) is the distribution of the start state. RL algorithms try to find the optimal policy \(\pi^*: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}_0^+\), where \(\pi^* = \argmax_{\pi} \sum_{t=0}^\infty \gamma^t \mathbb{E}_{s_{t+1} \sim T(\cdot\mid s_t,a_t),\, a_{t} \sim \pi(\cdot\mid s_t), s_0 \sim P_0}\left[ R(s_t,a_t,s_{t+1}) \right].\) Importantly, in this work, we consider RL problems where the state \(s\) encodes both the position and the shape of the objects in the scene. We require the RL agent to generalize over all of these shapes at test time. We can therefore think of the state as a tuple \(s=(s_p, s_s)\), where \(s_p\) encodes positional information, and \(s_s\) encodes the shapes involved. We focus the experiments on sparse reward settings, meaning \(R(s,a,s') = R_0 > 0\) for \(s'\in\mathcal{S}_g\) and \(R(s,a,s') = 0\) for \(s\in\mathcal{S}\backslash\mathcal{S}_g\), where the volume of \(\mathcal{S}_g \subset \mathcal{S}\) is much smaller than the volume of \(\mathcal{S}\). The state space \(\mathcal{S}\) usually is low-dimensional or a minimal description of the degrees of freedom of the system. In this work, we consider that the RL algorithm has only access to a (high-dimensional) observation \(y\in\mathcal{Y}\) of the scene (e.g., RGB images). In particular, this means that the policy has observations as input \(a \sim \pi(\cdot \mid y)\). Since we assume that the underlying state \(s=(s_p, s_s)\) is fully observable from \(y\), we can treat \(y\) like a state for an MDP. **Reinforcement Learning with Learned Latent Scene Representations.** The general idea of RL with learned latent scene representations is to learn an *encoder* \(\Omega\) that maps an observation \(y\in\mathcal{Y}\) to a \(k\)-dimensional *latent vector* \(z = \Omega(y)\in\mathcal{Z}\subset \mathbb{R}^k \label{eq:latentRL}\) of the scene. The actual RL components, e.g., the Q-function or policy, then operate on \(z\) as its state description. For a policy \(\pi\), this means that the action \(a \sim \pi(\cdot\mid z) = \pi(\cdot\mid \Omega(y))\) is conditional on the latent vector \(z\) instead of the observation \(y\) directly. The dimension \(k\) of the latent vector is typically (much) smaller than that of the observation space \(\mathcal{Y}\), but larger than that of the state space \(\mathcal{S}\).
## Neural Radiance Fields (NeRFs) {#sec:backgroundNerfs}
The general idea of NeRF, originally proposed by, is to learn a function \(f = (\sigma, c)\) that predicts the emitted RGB color value \(c(x)\in\mathbb{R}^{3}\) and volume density \(\sigma(x)\in\mathbb{R}_{\ge0}\) at any 3D world coordinate \(x\in\mathbb{R}^3\). Based on \(f\), an image from an arbitrary view and camera parameters can be rendered by computing the color \(C(r)\in\mathbb{R}^3\) of each pixel along its corresponding camera ray \(r(\alpha) = r(0) + \alpha d\) through the volumetric rendering relation \[\begin{aligned}
C(r) = \int_{\alpha_n}^{\alpha_f}T_f(r, \alpha)\sigma(r(\alpha))c(r(\alpha))\diff \alpha ~~~~~~~\text{with}~~~~~~~T_f(r, \alpha) = \exp\left(-\int_{\alpha_n}^\alpha \sigma(r(u))\diff u \right). \label{eq:nerfC} \end{aligned}\] Here, \(r(0)\in\mathbb{R}^3\) is the camera origin, \(d\in\mathbb{R}^3\) the pixel dependent direction of the ray and \(\alpha_n, \alpha_f \in \mathbb{R}\) the near and far bounds within which objects are expected, respectively. The camera rays are determined from the camera matrix \(K\) (intrinsics and extrinsics) describing the desired view.
# Learning State Representations for RL with NeRF Supervision {#sec:RLWithNeRF}
This section describes our proposed framework, in which we use a latent state space for RL that is learned from NeRF supervision. For learning the latent space, we use an encoder-decoder where the decoder is a latent-conditioned NeRF, which may either be a global or a compositional NeRF decoder. To our knowledge, no prior work has used such NeRF-derived supevision for RL. In Sec. [4.1](#subsec:nerf-latent-for-rl){reference-type="ref" reference="subsec:nerf-latent-for-rl"} we describe this proposition, Sec. [4.2](#subsec:comp-nerf-overivew){reference-type="ref" reference="subsec:comp-nerf-overivew"} provides an overview of the encoder-decoder training, Sec. [4.3](#sec:nerfDecoder){reference-type="ref" reference="sec:nerfDecoder"} and Sec. [4.4](#sec:objectEncoder){reference-type="ref" reference="sec:objectEncoder"} introduce options for the NeRF decoder and encoder, respectively.
## Using Latent-Conditioned NeRF for RL {#subsec:nerf-latent-for-rl}
We propose the state representation \(z\) on which an RL algorithm operates to be a latent vector produced by an encoder that maps images from multiple views to a latent \(z\), which is trained with a (compositional) latent-conditioned NeRF decoder. As will be verified in experiments, we hypothesize that this framework is beneficial for the downstream RL task, as it produces latent vectors that represent the actual 3D geometry of the objects in the scene, can handle multiple objects well, as well as fuse multiple views in a consistent way to deal with occlusions by providing shape completion, all of which is relevant to solve tasks where the geometry is important. There are two steps to our framework, as shown in Fig. [\[fig:RLWithLearnedStateRepresentations\]](#fig:RLWithLearnedStateRepresentations){reference-type="ref" reference="fig:RLWithLearnedStateRepresentations"}. First, we train the encoder + decoder from a dataset collected by random interactions with the environment, i.e., we do not yet need a trained policy. Second, we take encoder trained in the first step, which we leave frozen, and use the latent space to train an RL policy. Note that we investigate two variants of the auto-encoder framework, a global one, where the whole scene is represented by one single latent vector, and a compositional one, where objects are represented by their own latent vector. For the latter, objects are identified by masks in the views.
## Overview: Auto-Encoder with Latent-Conditioned NeRF Decoder {#subsec:comp-nerf-overivew}
Assume that an observation \(y = \left(I^{1:V}, K^{1:V}, M^{1:V}\right)\) of the scene consists of RGB images \(I^i\in\mathbb{R}^{3\times h \times w}\), \(i=1,\ldots, V\) taken from \(V\) many camera views, their respective camera projection matrices \(K^i\in\mathbb{R}^{3\times 4}\) (including both intrinsics and extrinsics), and per-view image masks \(M^{1:V}\). For a *global* NeRF decoder, these are global non-background masks \(M_{\text{tot}}^i\in\{0,1\}^{h\times w}\), and for a *compositional* NeRF decoder as in, these are sets of binary masks \(M_j^i\in\left\{0,1\right\}^{h\times w}\) that identify the objects \(j=1,\ldots, m\) in the scene in view \(i\). The global case is equivalent to \(m=1\), \(M^i_{j=1}=M^i_{\text{tot}}\). The encoder \(\Omega\) maps these posed image observations from the multiple views into a set of latent vectors \(z_{1:m}\), where each \(z_j\) represents each object in the scene separately in the compositional case, or the single \(z_1\) all objects in the scene. This is achieved by querying \(\Omega\) on the masks \(M_j^{1:V}\), i.e., \[\begin{aligned}
z_{j} = \Omega\left(I^{1:V}, K^{1:V}, M^{1:V}_{j}\right) \in\mathbb{R}^k \label{eq:omega} \end{aligned}\] for object \(j\). The supervision signal to train the encoder is the image reconstruction loss \[\begin{aligned}
\mathcal{L}^i = \left\|I^i\circ M^i_\text{tot}-D\left(\Omega\left(I^{1:V}, K^{1:V}, M^{1:V}_{1:m}\right), K^i\right)\right\|_2^2 \end{aligned}\] on the input view \(i\) where the decoder \(D\) renders an image \(I = D(z_{1:m}, K)\) for arbitrary views specified by the camera matrix \(K\) from the set of latent vectors \(z_{1:m}\). Both the encoder and decoder are trained end-to-end at the same time. The target images for the decoder are the same in both the global and compositional case: the global-masked image \(I^i\circ M^i_\text{tot}\) (\(\circ\) is the element-wise product). In the compositional case this can be computed with \(M^i_\text{tot} = \bigvee_{j=1}^m M_j^i\). By fusing the information from multiple views of the objects into the latent vector from which the decoder has to be able to render the scene from multiple views, this auto-encoder framework can learn latent vectors that represent the 3D configurations (shape and pose) of the objects in the scene.
## Latent-Conditioned NeRF Decoder Details {#sec:nerfDecoder}
**Global.** The original NeRF formulation learns a fully connected network \(f\) that represents one single scene (Sec. [3.2](#sec:backgroundNerfs){reference-type="ref" reference="sec:backgroundNerfs"}). In order to create a decoder from NeRFs within an auto-encoder to learn a latent space, we condition the NeRF \(f(\cdot, z)\) on the latent vector \(z\in\mathbb{R}^k\). While approaches such as use the latent code to represent factors such as lighting or category-level generalization, in our case the latent code is intended to represent the scene variation, i.e., shape *and* configuration of objects, such that a downstream RL agent may use this as a state representation. **Compositional.** In the compositional case, the encoder produces a set of latent vectors \(z_{1:m}\) describing each object \(j=1,\ldots, m\) individually, this leads to \(m\) many NeRFs \((\sigma_j(x), c_j(x)) = f_j(x) = f(x, z_j)\), \(j=1,\ldots, m\) with their associated volume density \(\sigma_j\) and color value \(c_j\). Note that while one could use different networks \(f_j\) with their own network weights for each object, we have a single network \(f\) for all objects. This means that both the object's pose as well as its shape and type are represented through the latent code \(z_j\). In order to force those conditioned NeRFs to learn the 3D configuration of each object separately, we compose then into a global NeRF model with the composition formulas (proposed e.g., by ): \(\sigma(x) = \sum_{j=1}^m \sigma_j(x)\), \(c(x) = \frac{1}{\sigma(x)}\sum_{j=1}^m \sigma_j(x)c_j(x).\) As this composition happens in 3D space, the latent vectors will be learned such that they correctly represent the actual shape and pose of the objects in the scene with respect to the other objects, which we hypothesize may be useful for the downstream RL agent.
## Encoder Details {#sec:objectEncoder}
The encoder \(\Omega\) operates by fusing multiple views together to estimate the latent vector for the RL task. Since the scientific question of this work is to investigate whether a decoder built from NeRFs to train the encoder end-to-end is beneficial for RL, we consider two different encoder architectures. The first one is a 2D CNN that averages feature encodings from the different views, where each encoding is additionally conditioned on the camera matrix of that view. The second one is based on a learned 3D neural vector field that incorporates 3D biases by fusing the different camera views in 3D space through 3D convolutions and camera projection. This way, we are able to distinguish between the importance of 3D priors incorporated into the encoder versus the decoder. **Per-image CNN Encoder ("Image encoder").** For the global version, we utilize the network architecture from as an encoder choice. In order to work with multiple objects in the compositional case, we modify the architecture from by taking the object masks into account as follows. For each object \(j\), the 2D CNN encoder computes \[\begin{aligned}
z_j = \Omega_\text{CNN}\left(I^{1:V}, K^{1:V}, M^{1:V}_{j}\right) = h_\text{MLP}\!\left(\frac{1}{V}\sum_{i=1}^Vg_\text{MLP}\!\left(E_\text{CNN}\!\left(I^i\circ M^i_j\right), K^i\right) \right).\label{eq:CNNEncoder} \end{aligned}\] \(E_\text{CNN}\) is a ResNet-18 CNN feature extractor that determines a feature from the masked input image \(I^i\circ M^i_j\) of object \(j\) for each view \(i\), which is then concatenated with the (flattened) camera matrix. The output of the network \(g_\text{MLP}\) is hence the encoding of each view, including the camera information, which is averaged and then processed with \(h_\text{MLP}\), to produce the final latent vector. Note that in the global case, we set \(m=1\), \(M^i_{j=1}=M_\text{tot}^i\) such that \(\Omega_\text{CNN}\) produces a single latent vector. **Neural Field 3D CNN Encoder ("Field encoder").** Several authors have considered to incorporate 3D biases into learning an encoder by computing pixel-aligned features from queried 3D locations of the scene to fuse the information from the different camera views directly in 3D space. We utilize the encoder architecture from, where the idea is to learn a neural vector field \(\phi\left[I^{1:V}, M^{1:V}_j\right]:\mathbb{R}^3\rightarrow\mathbb{R}^E\) over 3D space, conditioned on the input views and masks. The features of \(\phi\) are computed from projecting the query point into the camera coordinate system from the respective view. To turn \(\phi\) into a latent vector, it is queried on a workspace set \(\mathcal{X}_h\in \mathbb{R}^{d_\mathcal{X} \times h_\mathcal{X} \times w_\mathcal{X}}\) (a 3D grid) and then processed by a 3D convolutional network, i.e., \(z_j = E_\text{3D CNN}\left(\phi\left[I^{1:V}, M^{1:V}_j\right](\mathcal{X}_h)\right)\). This method differs from by computing a latent vector from the pixel-aligned features.
# Baselines / Alternative State Representations {#sec:baselines}
In this section, we briefly describe alternative ways of training an encoder for RL, which we will investigate in the experiments as baselines and ablations. **Conv. Autoencoder.**[\[sec:2DCNNDecoder\]]{#sec:2DCNNDecoder label="sec:2DCNNDecoder"} This baseline uses a standard CNN decoder based on deconvolutions instead of NeRF to reconstruct the image from the latent representation, similar to. Therefore, with this baseline we investigate the influence of the NeRF decoder relative to CNN decoders. We follow the architecture of for the deconvolution part for the global case. In the compositional case, we modify the architecture to be able to deal with a set of individual latent vectors instead of a single, global one. The image \(I = D_\text{deconv}(g_\text{MLP}(\frac{1}{m}\sum_{j=1}^m z_j), K)\) is rendered from \(z_{1:m}\) by first averaging the latent vectors and then processing the averaged vector with a fully connected network \(g_\text{MLP}\), leading to an aggregated feature. This aggregated feature is concatenated with the (flattened) camera matrix \(K\) describing the desired view and then rendered into the image with \(D_\text{deconv}\). In the experiments, we utilize this decoder as the supervision signal to train the latent space produced by the 2D CNN encoder from Sec. [4.4](#sec:objectEncoder){reference-type="ref" reference="sec:objectEncoder"}. In the compositional version, the 2D CNN encoder [\[eq:CNNEncoder\]](#eq:CNNEncoder){reference-type="eqref" reference="eq:CNNEncoder"} use the same object masks as the compositional NeRF-RL variant. **Contrastive Learning.** As an alternative to learning an encoder via a reconstruction loss, the idea of contrastive learning is to define a loss function directly on the latent space that tries to pull latent vectors describing the same configurations together (called positive samples) while ones representing different system states apart (called negative samples). A popular approach to achieve this is with the InfoNCE loss. Let \(y_i\) and \(\tilde{y}_i\) be two *different* observations of the *same* state. Here, \(\tilde{\cdot}\) denotes a perturbed/augmented version of the observation. For a mini-batch of observations \(\left\{(y_i, \tilde{y}_i)\right\}_{i=1}^n\), after encoding those into their respective latent vectors \(z_i = \Omega(y_i)\), \(\tilde{z}_i = \Omega(\tilde{y}_i)\) with the encoder \(\Omega\), the loss for that batch would use (\(z_i\), \(\tilde{z}_i\)) as a positive pair, and (\(z_i\),\(\tilde{z}_{\ne i}\)) as a negative pair, or some similar variation. A crucial question in contrastive learning is how the observation \(y\) is perturbed/augmented into \(\tilde{y}\) to generate positive and negative training pairs, described in the following. **CURL.** In CURL, the input image is randomly cropped to generate \(y\) and \(\tilde{y}\). We closely follow the hyperparameters and design of. CURL operates on a single input view and we carefully choose a view for this baseline from which the state of the environment can be inferred as best as possible. **Multi-View CURL.** This baseline investigates if the neural field 3D encoder (Sec. [4.4](#sec:objectEncoder){reference-type="ref" reference="sec:objectEncoder"}) can be trained with a contrastive loss. As this encoder operates on multiple input views we *double* the amount of available camera views. Half of the views are the same as in the other experiments, the other half are captured from sightly perturbed camera angles. We use the same loss as CURL, but with different contrastive pairs--rather than from augmentation, the contrastive style is taken from TCN: the positive pairs come from different views but at the same moment in time, while negative pairs come from different times. Therefore, this baseline can be seen as a multi-view adaptation of CURL. **Direct State / Keypoint Representations.**[\[sec:keypoint\]]{#sec:keypoint label="sec:keypoint"} Finally, we also consider a direct, low-dimensional representation of the state. Since we are interested in generalizing over different object shapes, we consider multiple 3D keypoints that are attached at relevant locations of the objects by expert knowledge and observed with a perfect keypoint detector. See Fig. [\[fig:mug:keypoints\]](#fig:mug:keypoints){reference-type="ref" reference="fig:mug:keypoints"} for a visualization of those keypoints. The keypoints both provide information about object shape and its pose. Furthermore, as seen in Fig. [\[fig:mug:keypoints\]](#fig:mug:keypoints){reference-type="ref" reference="fig:mug:keypoints"}, they have been chosen to reflect those locations in the environment relevant to solve the task. Additionally, we report results where the state is represented by the poses of the objects--as this cannot represent object shape, in this case we use a constant object shape for training and test.
# Experiments {#sec:exp}
We evaluate our proposed method on different environments where the geometry of the objects in the scene is important to solve the task successfully. Please also refer to the video <https://dannydriess.github.io/nerf-rl>. Commonly, RL is trained and evaluated on a single environment, where only the poses are changed, but the involved object shapes are kept constant. Since latent-conditioned NeRFs have been shown to be capable of generalizing over geometry, we consider experiments where we require the RL agent to generalize over object shapes within some distribution. Answering the scientific question of this work requires environments with multi-view observations --- and for the compositional versions object masks as well. These are *not provided in standard RL benchmarks*, which is the reason for choosing the environments investigated in this work. We use PPO as the RL algorithm and four camera views in all experiments. Refer to the appendix for more details about our environments, parameter choices, and training times.
## Environments {#sec:exp:environments}
**Mug on Hook.** In this environment, adopted from and visualized in Fig. [\[fig:mug:keypoints\]](#fig:mug:keypoints){reference-type="ref" reference="fig:mug:keypoints"}, the task is to hang a mug on a hook. Both the mug and the hook shape are randomized. The actions are small 3D translations applied to the mug. This environment is challenging as we require the RL agent to generalize over mug and hook shapes and the tolerance between the handle opening and the hook is relatively small. Further, the agent receives a sparse reward only if the mug has been hung stably. This reward is calculated by virtually simulating a mug drop after each action. If the mug would not fall onto the ground from the current state, a reward of one is assigned, otherwise zero. **Planar Pushing.** The task in this environment, shown in Fig. [\[fig:box:images\]](#fig:box:images){reference-type="ref" reference="fig:box:images"}, is to push yellow box-shaped objects into the left region of the table and blue objects into the right region with the red pusher that can move in the plane, i.e., the action is two dimensional. This is the same environment as in with the same four different camera views. Each run contains a single object on the table (plus the pusher). If the box has been pushed inside its respective region, a sparse reward of one is received, otherwise zero. The boxes in the environment have different sizes, two colors and are randomly initialized. In this environment, we cannot use keypoints for the multi-shape setting, as the reward depends on the object color; we evaluate the keypoints baseline only in the single shape case (Appendix). **Door Opening.** Fig. [\[fig:door:images\]](#fig:door:images){reference-type="ref" reference="fig:door:images"} shows the door environment, where the task is to open a sliding door with the red end-effector that can be translated in 3 DoFs as the action. To solve this task, the agent has to push on the door handle. As the handle position and size is randomized, the agent has to learn to interact with the handle geometry accordingly. Interestingly, as can be seen in the video, the agent often chooses to push on the handle only at the beginning, as, afterwards, it is sufficient to push the door itself at its side. The agent receives a sparse reward if the door has been opened sufficiently, otherwise, zero reward is assigned.
## Results {#sec:exp:results}
Figs [\[fig:mug:learning\]](#fig:mug:learning){reference-type="ref" reference="fig:mug:learning"}, [\[fig:box:learning\]](#fig:box:learning){reference-type="ref" reference="fig:box:learning"}, [\[fig:door:learning\]](#fig:door:learning){reference-type="ref" reference="fig:door:learning"} show success rates (averaged over \(6\) independent experiment repetitions and over \(30\) test rollouts per repetition per timestep) as a function of training steps. Also shown are the \(68\%\) confidence intervals. These success rates have been evaluated using randomized object shapes and initial conditions, and therefore reflect the agent's ability to generalize over these. In all these experiments, a latent space trained with compositional NeRF supervision as the decoder consistently outperformed all other learned representations, both in terms of sample efficiency and asymptotic performance. Furthermore, our proposed framework with compositional NeRF even outperforms the expert keypoint representation. For the door environment, the 3D neural field encoder plus NeRF decoder (NeRF-RL comp. + field) reaches nearly perfect success rates. For the other two environments, the compositional 2D CNN encoder plus NeRF decoder (NeRF-RL comp. + image) was slightly better than with the neural field encoder but not significantly. This shows that the *decoder* built from compositional NeRF is relevant for the performance, not so much the choice of the encoder. Training the 3D neural field encoder with a contrastive loss as supervision signal for different camera views as positive/negative training pairs is not able to achieve significant learning progress in these scenarios (Multi-CURL). However, the other contrastive baseline, CURL, which has a different encoder and uses image cropping as data augmentation instead of additional camera views, is able to achieve decent performance and sample efficiency on the door environment, but not for the pushing environment. In the mug environment, CURL initially is able to make learning progress comparable to our framework, but never reaches a success rate above 59% and then becomes unstable. Similarly, the global CNN autoencoder baseline shows decent learning progress initially on the mug and pushing scenario (not for the door), but then becomes unstable (mug) or never surpasses 50% success rate (pushing). Such variations in performance or instable learning across the different environments have not been observed with our method, which is stable in all cases. The compositional variant (NeRF-RL comp.) of our framework achieves the highest performance. Since the conv. comp. autoencoder baseline has worse performance than its global variant, compositionality alone is not the sole reason for the better performance of our state representation. Indeed, the global NeRF-RL + image variant in the pushing env. is also better than all other baselines.
# Discussion {#sec:discussion}
#### Why NeRF provides better supervision.
The NeRF training objective [\[eq:nerfC\]](#eq:nerfC){reference-type="eqref" reference="eq:nerfC"} strongly forces each \(f(\cdot, z_j)\) to represent each object in its actual 3D configuration and relative to other objects in the scene (compositional case), including their shape. This implies that the latent vectors \(z_j\) have to contain this information, i.e., they are trained to determine the object type, shape and pose in the scene. In the global case, \(z_1\) has to represent the geometry of the whole secne. As the tasks we consider require policies to take the geometry of the objects into account, we hypothesize that a latent vector that is capable of parameterizing a NeRF to reconstruct the scene in the 3D space has to contain enough of the relevant 3D information of the objects also for the policy to be successful.
#### Masks.
In order for the auto-encoder framework to be compositional, it requires object masks. We believe that instance segmentation has reached a level of maturity that this is a fair assumption to make. As we also utilize the individual masks for the compositional conv. autoencoder and the multi-view CURL baseline, which do not show good performance, it indicates that the masks are not the main reason that our state representation achieves higher performance. This is further supported by the fact that the global NeRF-RL variant which does not rely on individual object masks on the pushing scenario achieved a performance higher than all baselines, i.e., masks will increase the performance of NeRF-RL as they enable the compositional version, but they do not seem essential.
#### Offline/Online.
In this work, we focused on pretraining the latent representation offline from a dataset collected by random actions. During RL, the encoder is fixed and only the policy networks are learned. This has the advantage that the same representation can be used for different RL tasks and the dataset to train the representation not necessarily has to come from the same distribution. However, if a policy is needed to explore reasonable regions of the state space, collecting a dataset offline to learn a latent space that covers the state space sufficiently might be more challenging for an offline approach. This was not an issue for our experiments where data collection with random actions was sufficient. Indeed, we show generalization over different starting states of the same environment and with respect to different shapes (within distribution). Future work could investigate NeRF supervision in an online setup. Note that the reconstruction loss via NeRF is computationally more demanding than via a 2D CNN deconv. decoder or a contrastive term, making NeRF supervision as an auxiliary loss at each RL training step costly. One potential solution for this is to apply the auxiliary loss not at every RL training step, but with a lower frequency. Regarding computational efficiency, this is where contrastive learning has an advantage over our proposed NeRF-based decoder, as the encoding with CURL can be trained within half a day, whereas the NeRF auto-encoder took up to 2 days to train for our environments. However, when using the encoder for RL, there is no difference in inference time.
#### Multi-View.
The auto-encoder framework we propose can fuse the information of multiple camera views into a latent vector describing an object in the scene. This way, occlusions can be addressed and the agent can gain a better 3D understanding of the scene from the different camera angles. Having access to multiple camera views and their camera matrices is an additional assumption we make, although we believe the capability to utilize this information is an advantage of our method.
# Conclusion {#sec:conclusion}
In this work, we have proposed the idea to utilize Neural Radiance Fields (NeRFs) to train latent spaces for RL. Our environments focus on tasks where the geometry of the objects in the scene is relevant for successfully solving the tasks. Training RL agents with the pretrained encoder that maps multiple views of the scene to a latent space consistently outperformed other ways of learning a state representation and even keypoints chosen by expert knowledge. Our results show that the 3D prior present in compositional NeRF as the decoder is more important than priors in the encoder. **Broader Impacts.** Our main contribution is a method to learn representations that improve the efficiency of vision-based RL, which could impact automation. As such, our work inherits general ethical risks of AI, like the question of how to address the potential of increased automation in society. | {'timestamp': '2022-06-06T02:16:23', 'yymm': '2206', 'arxiv_id': '2206.01634', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01634'} |
null | null |
null | null |
# Introduction
Ground source heat pump (GSHP) is a highly efficient renewable energy technology which allows sufficient reduction in energy consumption for building heating and cooling. An important part of each GSHP system is a ground heat exchanger (GHE) installed in a vertical borehole or horizontal trench. Thermal energy is extracted from the ground or absorbed by the ground due to the effective heat exchange between the fluid, which circulates in the GHE, and the adjacent soil. The effective design of GHEs requires a good understanding of the thermal processes in the system GHE-surrounding soil. Three main approaches to the investigation of the system GHE-surrounding soil can be distinguished: analytical, numerical and experimental. Analytical approach allows obtaining analytical dependencies under a number of strong simplifications about the thermal processes. More complicated 2D or 3D geometry models require numerical simulation, but use the results of analytical approach and experimental approaches for their testing. A lot of information on the modeling of thermal processes for various types of GHE can be found in reviews. In, a list of commonly used software packages for GSHP design can be found as well. While most models focus on normal GHE operation with above-zero inlet temperature, the effect of phase change on GHE operation is of great importance in cold regions, where the circulating fluid inlet temperature must be below zero to extract the required amount of heat. Freezing-thawing processes around the GHE change the thermal characteristics of the soil thus affecting the thermal exchange between the pipe fluid and the surrounding soil. In particular, a number of experiments show that freezing increases the heat transfer rate and the coefficient of performance. As a result, utilization of latent heat from the groundwater freezing allows, for example, the reduction of vertical GHE length for the GSHP with the same characteristics. Another consequence is the possible damage of the pipe due to the ice lens formation in the grout and the impact of freeze-thaw cycles on GHE hydraulic conductivity. In, the influence of freeze-thaw cycles on aging effect of U-shaped GHEs were studied using the large-scale test rig determining the hydraulic conductivity of GHE systems. In, a pilot-scale experiment was built to test a 1-m section of a typical GHE under freezing-thawing conditions. Freeze-thaw cycles seem to be a leading cause for a potential violation of the system's integrity. In, the thermal processes in the saturated porous medium around the GHE were simulated using FEFLOW 2D axisymmetric model with C++ plug-in for phase changes between solid and liquid phases. Possible ways to avoid GHE pipes deformation and reduce flow resistance under freezing conditions were studied via experimental tests in and using numerical simulation in. Note that the damage of the pipes is mainly studied by experimental tests. On the contrary, the role of numerical simulation in studying the efficiency of thermal transfer processes in the system GHE-surrounding soil is very large. Up to now a number of math models for this aim was developed, which take into account freezing-thawing processes in the soil around the vertical and horizontal GHEs. The major part of all math models is the heat transfer equation for the surrounding soil. The freezing-thawing processes are described in the model using a latent heat term in the heat transfer equation and the heat transfer model with effective heat capacity to describe the three phases--solid, liquid and mushy. This method allows avoiding one of the difficulties in numerical simulation for phase change: how to treat the movable interface between the phases. The other important part of math model is temperature distribution inside the GHE pipe. Some investigations do not consider this temperature distribution, operating with the total heat gain/loss in GHE. The others (see, for example ) use a dual continuum approach. Under such approach, the soil is described in 3D geometry, while the two branches of U-shaped GHE are described in 1D geometry. To simplify the description of the temperature distribution inside the U-shaped pipe, the pipe is replaced by a mono-tube with equivalent diameter. Without a doubt, detailed 3D or even 2D calculation of thermal processes in the GHE and surrounding soil can provide accurate results. But such calculations are in most cases based on commercial software packages, such as FEFLOW, COMSOL, ABAQUS, OpenGeoSys, CFD ANSYS FLUENT and so on. Commercial software packages are also very useful at modeling the arrays of GHEs. An obvious drawback of these packages is the impossibility of numerical modeling of phase transfer processes in porous soil. To overcome this drawback, some authors use C++ or FORTRAN plug-ins to commercial software package for modeling phase transition between solid and liquid phases in porous surrounding soil. Another disadvantage of detailed 3D simulation is the large time of calculation and large computational resources necessary for the long-term simulation of the system GHE-surrounding soil with phase change processes. To reduce the computational resources, simplified 1D and 2D model for vertical GHEs were developed. For example in, an improved thermal resistance and capacity model (RC) was proposed for the heat transfer modeling between vertical single U-tube GHEs and the frozen soil. The model was verified through experiment tests and numerical simulation on basis of 3D CFD ANSYS model. In, a 2D model for temperature field modeling around the GHE was developed. The model is based on heat transfer equation in cylindrical geometry with a source term and Dirichlet boundary conditions to represent the GHE. In, 1D radial numerical heat transfer model is developed to evaluate heat transfer from the borehole wall to the ground with taking into account three phases: ice-soil, water-soil and transition phase. For the boundary conditions at the borehole wall and the outer boundary of computational area the temperature, measured experimentally, was used. Another approach to reduce the computational time was demonstrated in for the modeling of a GHE array applied to an existing historical building in Venice (Northern Italy). In this paper, two different scale models were developed. The results of the coarser large scale model were used as initial and boundary conditions for the fully discretized small scale model, developed for the detailed description of freezing-thawing processes in the close vicinity of the GHE. The processes of freezing-thawing are closely related to the processes of groundwater seepage and flow. The influence of these two factors--freezing-thawing together with ground water seepage, are taken into account in in modeling the temperature field in the surrounding soil for the array of vertical U-shaped GHEs. The temperature field equation for the porous surrounding soil with convection and ice/water phase and Brinkman equation for water transport in porous media are the basis of the 3D model, which is solved by the COMSOL software. The model does not describe the temperature inside the pipes. Instead it uses the iteration processes to determine the inlet and output temperature on basis of the soil model. The similar model was developed in for artificial freezing. This paper proposes a combined mathematical model containing a 1D model of heat transfer in a fluid circulating inside a vertical coaxial GHE and a 2D model of heat conduction in a cylindrically symmetric surrounding soil. The model takes into account temperature changes in the inner and annular pipes, soil freezing processes, geothermal temperature gradient and horizontal soil stratification. Numerical implementation of the model, based on the finite difference method and the MatLab **ode15s** numerical solver, makes it possible to investigate the effect of soil freezing on the outlet temperature of the coaxial heat exchanger. Coaxial GHE have a number of advantages in comparison with U-shaped one, especially in the case of deep boreholes (1000 to 3000 meters depth). Among them is a simple installation procedure, moderate temperature difference between the secondary fluid and the surrounding ground, the possibility to use water as a secondary fluid even in colder countries and to ignore its local thermal resistance in comparison with U-shaped GHE . The comparison of different types of GHE can be found in, where the equivalent ground thermal conductivity was evaluated by Thermal Response Test. A review on various approaches for investigation of coaxial GHEs can be found, for example, in. Geothermal gradient was taken into account in an analytical model and semi-analytical models proposed for shallow (100 to 200 m depth,, ) and deep (1000 to 3000 meters depth, ) coaxial BHE. The vertical geological structure of the adjacent ground was considered in at numerical modeling of a deep coaxial GHE. However, to our knowledge, freezing-thawing processes have not yet been considered for coaxial GHE. It should be noted that, in deep GHEs, only pure water (without anti-freeze) with positive inlet temperature is allowed to use to prevent pollution on deep groundwater resources. Therefore, in our paper, we focus on a shallow vertical coaxial GHE with negative inlet temperature of circulating fluid.
# Formulation of the problem
## Scheme of coaxial ground heat exchanger
We imagine a heat exchanger in the form of two coaxial pipes located in a vertical borehole and fixed in it with the help of a grout. The cross section of the heat exchanger is shown schematically, not to scale, in the Figure [\[she\]](#she){reference-type="ref" reference="she"}. The cross section of the inner pipe is shown in the center of the figure, \(r_{11}\) and \(r_{12}\) are the inner and the outer radius of this pipe, respectively. Coaxial with the inner pipe is the outer pipe, its inner and outer radius are designated \(r_{21}\) and \(r_{22}\), respectively. The radius of the borehole is denoted by \(r_b\), the space between the outer pipe and the soil outside the borehole is filled with grout. In the heat exchanger scheme, we take into account the thickness of the pipe walls and the thickness of the grout layer to estimate the magnitude of heat flows between the pipes and between the fluid and the soil. In the process of heat exchange, the heat-carrying fluid moves along the inner tube and in the annular space between the tubes in opposite directions. We will consider a heat exchanger in which the fluid moves from top to bottom along the annular area between the pipes and rises up the inner pipe. Such a mode is usually used to extract heat from the ground. The temperature of the fluid in the inner pipe will be denoted by \(T_1\), in the space between the pipes--by \(T_2\), and through \(T_s\) we will denote the temperature of the soil. In the general case, the temperatures \(T_1\), \(T_2\), and \(T_s\) are unknown functions of time and spatial coordinates.
## Mathematical statement of the problem
To calculate the temperature of the fluid and the soil surrounding the borehole, we will solve the initial-boundary value problem in a circular cylinder of radius \(r_d\) and height \(H\) with a heat exchanger located along the vertical axis of the cylinder. The problem will be solved in a cylindrical coordinate system with the origin located in the center of the upper end of the inner pipe and the axis \(z\) directed downward and coinciding with the axis of the inner pipe. We assume that the temperature of the fluid \(T_1\) and \(T_2\) can depend only on the time \(t\) and the coordinate \(z\), \(T_1(t,z)\) and \(T_2(t,z)\). We do not take into account the thermal conductivity of the fluid, assuming that heat in the exchanger is distributed only due to the flow of the fluid and heat transfer through the walls of pipes and borehole. Regarding the soil, we assume that its physical properties may depend on the coordinates \(r\) and \(z\), but do not depend on the azimuth angle, \(T_s=T_s(t,r,z)\). The side wall of the cylinder at \(r = r_d\) is considered to be thermally insulated, and a constant temperature is maintained on the bases of the cylinder. Moreover, we will take into account the geothermal gradient, which is essential for long heat exchangers. Thus, we will not take into account seasonal changes in temperature on the surface of the earth, as well as changes in the properties of the surface itself. It can be assumed that the upper part of the heat exchanger is located at a depth of about 10 m under the earth surface, where the soil temperature is practically constant throughout the year and is equal to the average annual air temperature on the surface. Based on the assumptions made, the problem of calculating the temperature of the fluid and the soil can be written in the following form. \[\begin{aligned}
&&c_f\rho_fa_1\frac{\partial T_1(t,z)}{\partial t} = c_fv_f\frac{\partial T_1(t,z)}{\partial z}-q_1(T_1,T_2), \\ &&c_f\rho_fa_2\frac{\partial T_2(t,z)}{\partial t} = -c_fv_f\frac{\partial T_2(t,z)}{\partial z} + q_1(T_1,T_2)-q_2(T_2,T_{sb}), \\ &&c_s\rho_s\frac{\partial T_s(t,r,z)}{\partial t} = \frac{1}{r}\frac{\partial}{\partial r}\left[r\lambda_s\frac{\partial T_s(t,r,z)}{\partial r}\right] + \frac{\partial}{\partial z}\left[\lambda_s\frac{\partial T_s(t,r,z)}{\partial z}\right]. \end{aligned}\] Here are the constant parameters of the fluid: \(c_f\), \[J/(kg\(\cdot\)K)\], is the heat capacity; \(\rho_f\), \[kg/m\(^3\)\], is the density; and \(v_f\), \[kg/s\], is the mass velocity. The cross section area of the inner pipe and the annular region between the pipes are denotes as \(a_1=\pi r_{11}^2\) and \(a_2=\pi r_{21}^2-\pi r_{12}^2\). In contrast to the fluid parameters, the soil parameters \(c_s\) and \(\rho_s\), as well as the thermal conductivity of the soil, \(\lambda_s\) \[W/(m\(\cdot\)K)\], in the general case, can depend on spatial coordinates and on temperature \(T_s\). However, in order not to complicate the formulation of the problem and the numerical solution, we will consider the smooth dependence of the soil parameters on the coordinates. The terms \(q_1\) and \(q_2\), \[W/m\], describe heat transfer between fluids flowing down and upwards, and between fluid and soil, respectively. As usual, we calculate the values of these terms from the boundary value problem for the stationary heat equation in the rings \(r_{11}<r<r_{12}\) and \(r_{21}<r<r_b\) with Robin boundary conditions at the pipe walls and the Dirichlet condition at the borehole wall. From the solution of these problems, \[\begin{aligned}
&&q_1(T_1,T_2) = \frac{2\pi\lambda_1(T_1-T_2)}{\ln\left(\frac{r_{12}}{r_{11}}\right) + \frac{\lambda_1}{\alpha_{11}r_{11}} + \frac{\lambda_1}{\alpha_{12}r_{12}}}, \\ &&q_2(T_2,T_{sb}) = \frac{2\pi\lambda_2(T_2-T_{sb})}{\ln\left(\frac{r_{22}}{r_{21}}\right) + \frac{\lambda_2}{\lambda_g}\ln\left(\frac{r_b}{r_{22}}\right) + \frac{\lambda_2}{\alpha_{21}r_{21}}}. \end{aligned}\] Here \(\lambda_1,\, \lambda_2,\, \lambda_g\) are the thermal conductivity coefficients of the walls of the inner and outer pipes and the grout, respectively; \(\alpha_{11},\,\alpha_{12},\, \alpha_{21}\) are the coefficients of convective heat transfer between the fluid and pipe walls: \(T_{sb}\) is the temperature of the soil adjacent to the borehole wall. Note that the equations, similar to equations (1)-(5), were written in for a deep coaxial GHE, without taking into account the grout and the freezing-thawing processes. The equations (1), (2) are defined in segment \(0<z<H\) with the boundary conditions \[T_1(t,H)=T_2(t,H);\; T_2(t,0) = T_{in},\] where \(T_{in}\) is the fluid temperature at the heat exchanger inlet. In addition, we will need the outlet temperature \(T_{out}(t)=T_1(t,0)\), as well as the heat extraction rate \[q_{ex}(t) = c_fv_f(T_{out}(t)-T_{in}).\] The equation (3) is defined in the rectangle \(\{r_b<r<r_d,\, 0<z<H\}\). In accordance with the assumptions made above, we set the boundary conditions for Eq. (2) as follows \[\begin{aligned}
&& \lambda_sr_b\frac{\partial T_s(t,r,z)}{\partial r}\bigg|_{r=r_b} = q_2,\; \frac{\partial T_s(t,r,z)}{\partial r}\bigg|_{r=r_d} = 0, \\ && T_s(t,r,z)\big|_{z = 0} = T_{top},\; T_s(t,r,z)\big|_{z = H} = T_{bot}, \end{aligned}\] where \(T_{top}\) is the specified soil temperature at the top of the heat exchanger (assumed to be equal to the average annual temperature of the ambient air), and \(T_{bot}\) is calculated by the equation \[T_{bot} = T_{top}+G_tH,\] where \(G_t\) is the geothermal gradient As the initial conditions for equations (1)-(3) we put \[\label{tf0} T_1(0,z)=T_2(0,z)=T_{top}+G_tz,\] and also \[\label{ts0} T_s(0,r,z)=T_{top}+G_tz,\]
## Numerical scheme
We will solve problem (1)-(11) numerically using the finite difference method. For this purpose, we introduce a uniform grid along the \(z\) coordinate, \(z_i=ih\), \(i = 0,1,\dots,N_z\), \(h = H/N_z\); and an exponential grid along the \(r\) coordinate, \[r_j = r_b\exp\left[\frac{j-0.5}{N_r}\ln\left(\frac{r_d}{r_b}\right)\right],\, j = 1,2,\dots,N_r.\] An exponential grid has been used in several papers cited above and is needed to record the rapid change in soil temperature near borehole. Obviously, it is inspired by the fundamental solution of the Laplace equation in polar coordinates, \(T(r)\sim \ln r\). Passing to grid functions, \(T^1_i(t) = T_1(t,z_i)\), \(T^2_i(t) = T_2(t,z_i)\), \(T^s_{i,j}(t) = T_s(t,r_j,z_i)\); and replacing spatial derivatives in equations (1)-(3) with finite differences, we obtain a system of \(N_z(N_r+2)\) ordinary differential equations. \[\begin{aligned}
\label{et1} &&c_f\rho_fa_1\frac{dT^1_i}{dt} = c_fv_f\frac{T^1_{i+1}-T^1_i}{h}-q_1(T^1_i,T^2_i), \\ &&i = 0,1,\dots,N_z-1,\, T^1_{N_z} = T^2_{N_z}; \nonumber \\ \label{et2} &&c_f\rho_fa_2\frac{dT^2_i}{dt} =-c_fv_f\frac{T^2_i-T^2_{i-1}}{h}+q_1(T^1_i,T^2_i)-q_2(T^2_i,T^s_{i,1}), \\ &&i =1,2,\dots,N_z,\, T^2_0 = T_{in}; \nonumber \\ \label{ets} &&(c_s\rho_s)_{i,j}S_j\frac{dT^s_{i,j}}{dt} = R_{i,j}+Z_{i,j},\, i = 1,2...,N_z, \, j = 1,2...,N_r; \end{aligned}\] where \(S_j = \pi(r^{*2}_j-r^{*2}_{j-1})\). \(r^*_0 = r_b\), \(r^*_j = 0.5(r_j+r_{j+1})\). \[\begin{aligned}
&&R_{i,1} = q_2(T^2_i,T^s_{i,1})-2\pi r^*_1\lambda_{i,1}^r\frac{T^s_{i,1}-T^s_{i,2}}{r_2-r_1}; \nonumber\\ &&R_{i,j} = 2\pi r^*_{j-1}\lambda_{i,j-1}^r\frac{T^s_{i,j-1}-T^s_{i,j}}{r_j-r_{j-1}}-2\pi r^*_j\lambda_{i,j}^r\frac{T^s_{i,j}-T^s_{i,j+1}}{r_{j+1}-r_j},\\ &&j = 2,3...,N_r-1; \nonumber \\ &&R_{i,N_r} = 2\pi r^*_{N_r-1}\lambda_{i,N_r-1}^r\frac{T^s_{i,N_r-1}-T^s_{i,N_r}}{r_{N_r}-r_{N_r-1}}, \nonumber\\ &&i = 1,2...,N_z, \nonumber \end{aligned}\] where \[\lambda_{i,j}^r = \frac{2\lambda_{i,j}\lambda_{i,j+1}}{\lambda_{i,j}+\lambda_{i,j+1}}.\] \[\begin{aligned}
&&Z_{i,j} = \frac{S_j}{h^2}\left[\lambda_{i-1,j}^z(T^s_{i-1,j}-T^s_{i,j})-\lambda_{i,j}^z(T^s_{i,j}-T^s_{i+1,j})\right], \\ &&i = 1,2...,N_z,\, j = 1,2...,N_r; \nonumber \end{aligned}\] where \[\lambda_{i,j}^z = \frac{2\lambda_{i,j}\lambda_{i+1,j}}{\lambda_{i,j}+\lambda_{i+1,j}}.\; T^s_{0,j} = T_{top},\, T^s_{N_z+1,j} = T_{bot}.\] We have written finite-difference equations for the general case when soil parameters depend on spatial coordinates. Obviously, for a homogeneous soil, the equations are somewhat simplified.
# Model parameters
For simulations that demonstrate the capabilities of the model, it is necessary to set the values of the parameters of the heat exchanger, heat-carrying fluid and ground. The parameters of shallow ground heat exchangers can take values in fairly wide ranges. In this work, we do not consider any special heat exchanger. Therefore, for numerical simulation, we have chosen the typical values of the parameters that are found in the literature. These values are presented in Table 1.
Description Value
--------------- ------------------------------------ ----------------------
\(H\) Pipes length 100 m
\(r_{11}\) Inner radius of inner pipe 0.04 m
\(r_{12}\) Outer radius of inner pipe 0.043 m
\(r_{21}\) Inner radius of outer pipe 0.074 m
\(r_{22}\) Outer radius of outer pipe 0.08 m
\(r_b\) Radius borehole 0.1 m
\(\lambda_1\) Thermal conductivity of inner pipe 0.4 W/(m\(\cdot\)K)
\(\lambda_2\) Thermal conductivity of outer pipe 40 W/(m\(\cdot\)K)
\(\lambda_g\) Thermal conductivity of grout 1.5 W/(m\(\cdot\)K)
\(\alpha_{11}\) Convective coefficient at \(r_{11}\) 1000 W/(m\(^2\cdot\)K)
\(\alpha_{12}\) Convective coefficient at \(r_{12}\) 500 W/(m\(^2\cdot\)K)
\(\alpha_{21}\) Convective coefficient at \(r_{21}\) 500 W/(m\(^2\cdot\)K)
\(c_f\) Fluid heat capacity 4500 J/(kg\(\cdot\)K)
\(\rho_f\) Fluid density 1000 kg/m\(^3\)
\(v_f\) Fluid mass velocity 0.3 kg/s
: Heat exchanger parameters
Difference equations (12)-(16) make it possible to calculate the temperature distribution and heat fluxes in inhomogeneous soil. To demonstrate this possibility in this work, we have chosen an imaginary soil composed of several horizontal layers of minerals with different physical properties. The composition of the soil and the location of minerals in depth are presented in Table 2. The properties of minerals are described by the parameters, \(\rho_s^0\), \(c_s^0\), \(\lambda_s^0\)--density, heat capacity and thermal conductivity of dry matter, and volumetric moisture content in the substance of the layer, \(m\). The bottom part of the table shows the thermophysical parameters for water and ice. In the next section, the parameters of water are denoted by the subscript \(w\) (\(\rho_w\), \(c_w\), \(\lambda_w\)), and of ice--by the subscript \(ice\) (\(\rho_{ice}\), \(c_{ice}\), \(\lambda_{ice}\)).
----------- ----------- ------------ --------------- --------------- -------------
Soil type \(z_t-z_b\) \(\rho_s^0\) \(c_s^0\) \(\lambda_s^0\) \(m\)
m kg/m\(^3\) J/(m\(\cdot\)K) W/(m\(\cdot\)K) m\(^3\)/m\(^3\)
clay 0-5 1700 920 1.1 0.2
limestone 5-60 2500 840 1.0 0.15
mudstone 60-90 2600 800 1.8 0.1
granite 90-100 2700 790 1.1 0.05
Average 0-100 2510 827 1.25 0.13
water 997 4200 0.57
ice 919 2108 2.25
----------- ----------- ------------ --------------- --------------- -------------
: Strati-graphic column of borehole
# Simulation results
Equations ([\[et1\]](#et1){reference-type="ref" reference="et1"})-([\[ets\]](#ets){reference-type="ref" reference="ets"}) can be written in matrix form as \[\label{mat} \frac{d{\bm T}}{dt} ={\bf A} {\bm T} + {\bm R},\] where \(\bm T\) is a column vector containing \(N_e=(N_r+2)N_z\) components. \[T_i=T^1_i,\; T_{N_z+i}=T^2_i,\; i = 1,\dots,N_z;\] and \[\{T_{(j+1)N_z+i}=T^s_{i,j},\; i = 1,\dots,N_z\},\; j=1,\dots,N_r.\] The matrix \(\bf A\) depends on the fluid and soil parameters, while the vector \(\bm R\) takes into account the boundary conditions of the problem.The elements of the matrix \(\bf A\) and the column vector \(\bm R\) are easily determined from equations ([\[et1\]](#et1){reference-type="ref" reference="et1"})-([\[ets\]](#ets){reference-type="ref" reference="ets"}). The matrix \(\bf A\) contains \(N_e^2\) elements, but it is essentially 5-diagonal. Namely, only \(11N_z + (N_r-1)(5N_z-2)-5\) of them are nonzero. With the values of \(N_z\) and \(N_r\) taken in our calculations, the ratio of the number of non-zero elements to the total number of elements is of the order of \(10^{-4}\). Thus, the matrix \(\bf A\) is very strongly sparse. In this case, when solving system ([\[mat\]](#mat){reference-type="ref" reference="mat"}), it is very efficient to use the technique of sparse matrices, which makes it possible to radically reduce the size of RAM and the computation time. In this work, the solution of system ([\[mat\]](#mat){reference-type="ref" reference="mat"}) was carried out in MatLab using the **ode15s** function, which effectively uses the sparseness of the matrix and overcomes the stiffness of the problem. During the operation of the heat exchanger, the temperature of the fluid and the surrounding soil can drop below the freezing temperature of the ground moisture, \(T_f\) (for definiteness, we assume \(T_f=0^\circ\)C). Antifreeze can be used as a heat carried fluid, but ground moisture will freeze at low temperatures, and then, when the temperature rises, melt. Thus, in ground moisture, a water-ice phase transition will occur. A change in the state of aggregation of moisture leads to a change in the physical properties of the soil and is accompanied by the release or absorption of the latent heat of the phase transition, \(L\) = 334 kJ/kg. Therefore, adequate modeling of heat exchangers operating at the phase transition temperature must take this transition into account. However, due to computational difficulties, it is often ignored. In this work, for comparison, we present the results of calculations obtained both with and without taking into account the freezing-thawing of ground moisture. In calculations that do not take into account the freezing of moisture, we assume that moisture behaves like antifreeze. In the following, a model with this assumption will be referred to as *model A*. In model A, at any temperature, the ground parameters in any stratum are calculated by the formula \[\label{psp} p_s = (1-m)p^0_s + mp_w,\] where \(p\) denotes \(\rho\), \(c\) or \(\lambda\). To model the freezing-thawing of ground moisture, we use the method of apparent heat capacity. In this method, it is assumed that the water-ice transition occurs in a fairly narrow temperature range, \(T_f-\Delta_f<T<T_f+\Delta_f\). When the temperature \(T\) lies in this interval, then it is assumed that water and ice are simultaneously present in the soil. Moreover, the proportion of water, \(w\), in the mixture is calculated by the formula \[\label{wmix} w = \frac{1}{2}\left(\frac{T}{\Delta_f}+1\right).\] And the ground parameters are determined by the equations \[\begin{aligned}
\label{fsp} && \rho_s = (1-m)\rho^0_s +m[w\rho_w+(1-w)\rho_{ice}], \nonumber \\ && c_s = (1-m)c^0_s +m[wc_w+(1-w)c_{ice}+L/(2\Delta_f)], \\ && \lambda_s = (1-m)\lambda^0_s + m[w\lambda_w+(1-w)\lambda_{ice}]. \nonumber \end{aligned}\] At \(T > T_f+\Delta_f\) the soil parameters are calculated by Eq. [\[psp\]](#psp){reference-type="ref" reference="psp"}. While at \(T < T_f-\Delta_f\), when all the moisture in the soil is frozen, the soil parameters are calculated by the equation \[\label{msp} p_s = (1-m)p^0_s + mp_{ice},\] Allowance for moisture freezing makes system ([\[mat\]](#mat){reference-type="ref" reference="mat"}) non-linear and significantly increases the computation time. When solving ([\[mat\]](#mat){reference-type="ref" reference="mat"}) numerically, the matrix elements are recalculated only in the vicinity of the phase transition band. As \(\Delta_f\) decreases, the number of elements to be updated decreases too, but the stiffness of the problem increases and the computation time increases. Therefore, the value of \(\Delta_f\) is chosen so as to balance these opposite tendencies. In the calculations presented below, it is assumed that \(\Delta_f = 0.1^\circ\)C. In the following, the model that takes into account the freezing of soil moisture will be referred to as the *model F*.
## Stationary mode
Under unchanged boundary conditions, the solution \({\bm T}(t,r,z)\) of system ([\[mat\]](#mat){reference-type="ref" reference="mat"}) at \(t\to\infty\) asymptotically approaches the stationary solution \({\bm T}_{st}(r,z)\), which satisfies the system of algebraic equations \[\label{mats} {\bf A}{\bm T}_{st} + {\bm R} = 0.\] Figure 2 shows graphs of the stationary temperature of the fluid and the soil adjacent to the borehole wall, calculated both with allowance for moisture freezing (solid lines) and without freezing (dotted lines). Freezing of ground moisture in our model leads to an increase in the thermal conductivity of the soil, and hence to an increase in the heat flux from the external soil to the heat exchanger. Therefore, in Fig. 2 temperature graphs for the model with freezing are shifted to the right relative to the graphs for the model without freezing. In particular, under given conditions, the temperature of the fluid at the outlet of the heat exchanger \(T_{out} =-8.2^\circ\)C in the model without freezing and \(-8.0\) with freezing. Under considered conditions, a difference of 0.2 degrees in temperature means a difference of 972 kJ/h in heat extraction rate or 10 percent of the heat extraction rate in model A. On the graphs of the temperature of the adjacent soil, a protrusion is distinguished at a depth of 60 to 90 m. This protrusion corresponds to a layer in which the thermal conductivity coefficient is significantly higher than in neighboring layers. Therefore, the heat flux from the outer soil in this layer is greater than in the neighboring ones, and, accordingly, the soil temperature is higher. In our model and in many other models of heat exchangers, it is assumed that the heat flow between the coolant and the adjacent soil is proportional to the temperature difference between the soil and the coolant. Figure 3 shows the graphs of the temperature difference between the incoming liquid and the adjacent soil for model A and for model F. The color map of the temperature field in the vicinity of the borehole is shown in Fig. 4, on the left--for model A and on the right--for model F. The numbers near the temperature level lines show the temperature value on the corresponding line. It can be seen that the soil temperature in the model with freezing (F) is lower than the soil temperature in the model without freezing. This is understandable, since it follows from Figures 2 and 3 that the amount of heat extracted from the soil per unit time in model F is greater than in model A. However, it also follows from Fig. 2 that the temperature of the soil adjacent to the borehole in model F is higher than in model A. This apparent contradiction is clarified by Fig. 5, which shows plots of soil temperature depending on the radius for \(z\) = 50 m (the middle of the exchanger). The temperature graphs intersect at a distance of approximately 10 cm from the borehole wall. To the left of the intersection point, the temperature in model F is higher than in model A, and to the right, it is vice versa. In stationary mode, it is convenient to evaluate the sensitivity of various characteristics of the heat exchanger to changes in its parameters. In particular, let us estimate the sensitivity of the stationary heat extraction rate, \(q_{st}\). In a linear approximation, the change in \(q_{st}\) with a change in the parameter \(p\) is calculated by the formula \[\Delta q_{st} \approx \frac{\partial q_{st}}{\partial p}\Delta p,\] where \(p\) is one of the GHE parameters listed in Table 1. We estimate the partial derivatives numerically using the central finite difference by calculating \(q_{st}\) at \(p = (1\pm 0.01)p_0\), where \(p_0\) is the value of parameter \(p\) specified in the table. The results of calculation are presented in Table 3.
\(T_{in}=+2^\circ\)C \(T_{in}=-10^\circ\)C, A \(T_{in}=-10^\circ\)C, F
--------------- -------------------------------------------------- -------------------------------------------------- --------------------------------------------------
\(r_{11}\) \(-\)`<!-- -->`{=html}1.732e+04 \(-\)`<!-- -->`{=html}3.758e+04 \(-\)`<!-- -->`{=html}4.582e+04
\(r_{12}\) 1.568e+04 3.401e+04 4.148e+04
\(r_{21}\) 7.572e+01 1.672e+02 2.825e+02
\(r_{22}\) 1.868e+03 4.125e+03 5.444e+03
\(r_b\) 2.985e+02 7.039e+02 1.517e+01
\(\lambda_1\) \(-\)`<!-- -->`{=html}1.231e+02 \(-\)`<!-- -->`{=html}3.563e+01 \(-\)`<!-- -->`{=html}3.293e+02
\(\lambda_2\) 1.092e\(-\)`<!-- -->`{=html}02 2.412e\(-\)`<!-- -->`{=html}02 2.833e\(-\)`<!-- -->`{=html}02
\(\lambda_g\) 2.223e+01 4.909e+01 6.518e+01
\(\alpha_{11}\) \(-\)`<!-- -->`{=html}6.808e\(-\)`<!-- -->`{=html}03 \(-\)`<!-- -->`{=html}1.477e\(-\)`<!-- -->`{=html}02 \(-\)`<!-- -->`{=html}1.786e\(-\)`<!-- -->`{=html}02
\(\alpha_{12}\) \(-\)`<!-- -->`{=html}2.533e\(-\)`<!-- -->`{=html}02 \(-\)`<!-- -->`{=html}5.495e\(-\)`<!-- -->`{=html}02 \(-\)`<!-- -->`{=html}6.507e\(-\)`<!-- -->`{=html}02
\(\alpha_{21}\) 1.121e\(-\)`<!-- -->`{=html}02 2.475e\(-\)`<!-- -->`{=html}02 4.180e\(-\)`<!-- -->`{=html}02
\(c_f\) 3.964e\(-\)`<!-- -->`{=html}02 8.725e\(-\)`<!-- -->`{=html}02 1.094e\(-\)`<!-- -->`{=html}01
\(\rho_f\) 1.137e\(-\)`<!-- -->`{=html}12 \(-\)`<!-- -->`{=html}1.439e\(-\)`<!-- -->`{=html}11 \(-\)`<!-- -->`{=html}1.105e\(-\)`<!-- -->`{=html}11
\(v_f\) 5.946e+02 1.309e+03 1.642e+03
\(T_{in}\) \(-\)`<!-- -->`{=html}1.124e+02 \(-\)`<!-- -->`{=html}1.242e+02 \(-\)`<!-- -->`{=html}1.382e+02
: Parameters sensitivity
Table 3 shows estimates of partial derivatives for both positive, \(T_{in}=2^\circ\)C, and negative, \(T_{in}=-10^\circ\)C, inlet temperatures. Moreover, for a negative temperature, the derivatives were calculated in both models A and F. In all calculations for the heat extraction rate \(q_{st}\), \(r_{11}\) and \(r_{12}\) turned out to be the most sensitive parameters. The derivatives with respect to them have different signs. The derivative with respect to \(r_{11}\) is negative. An increase in \(r_{11}\) leads to a decrease in the thermal resistance of the inner pipe wall and to a decrease in the linear velocity of the fluid in the inner pipe. Therefore, the fluid cools down more strongly when moving through the inner pipe, which leads to a decrease in \(q_{st}\). The derivative with respect to \(r_{12}\) is positive. An increase in \(r_{12}\) leads to an increase in thermal resistance between the fluids in the outer and inner pipes, so the outgoing fluid cools less and \(q_{st}\) increases. Similar conclusions about the effect of inlet temperature and inner and outer pipes diameters on the thermal performance of a coaxial GHE were reported in for a deep GHE. Note that under given conditions, all derivatives of \(q_{st}\) with respect to the heat exchanger parameters have the same signs and almost all of them are of the same order of magnitude. The exceptions are the derivatives with respect to \(r_b\) and \(\lambda_1\). The derivative with respect to \(r_b\) in model A is 46 times greater than in model F, and the derivative with respect to \(\lambda_1\) in model A is approximately 10 times smaller in absolute value than in model F. The stationary solution is easy to calculate and provides useful information about the heat extraction process. However, with a significant difference between the average soil temperature (11.5\(^\circ\)C in our example) and the inlet temperature \(T_{in}\), the stationary solution turns out to be practically unattainable due to the large relaxation time. Figure 6 gives an idea of the relaxation time for the assumed conditions. The relaxation time increases rapidly as the difference between the natural ground temperature and the inlet fluid temperature increases. Moreover, the temperature of the fluid in the heat exchanger approaches the equilibrium temperature much faster than the temperature of the soil. We estimate the time of approaching the soil temperature to equilibrium by the time \(t_*\) at which the inequality \(\max_{r,z}|T(t,r,z)-T_{st}|<\delta\) is satisfied for the first time, where \(T\) is the solution of ([\[mat\]](#mat){reference-type="ref" reference="mat"}) and \(\delta=1^\circ\)C. For the fluid, we estimate the time \(t_\circ\) at which the inequality \(T_{out}(t)-T_{out}^{st}<\delta\) begins to hold for \(\delta=0.1^\circ\)C. The graphs of \(t_*\) and \(t_\circ\) in dependence on \(T_{in}\) are shown in Figure 6.
## Transient mode
Let us now return to system ([\[mat\]](#mat){reference-type="ref" reference="mat"}) and present the time-dependent solution of the system immediately after the start of working. Here and below, all calculations are performed with the initial conditions ([\[tf0\]](#tf0){reference-type="ref" reference="tf0"}), ([\[ts0\]](#ts0){reference-type="ref" reference="ts0"}). Figure 7 shows graphs of fluid temperature, \(T_{out}(t)\), at the outlet of the heat exchanger for the first five days of operation at \(T_{in}=-10^\circ\)C for both models A and F. During the first day of operation, the temperature \(T_{out}(t)\) drops rapidly. By the end of the fifth day, the temperature drop noticeably slows down, although it is still very far from the equilibrium temperature, \(-8.2^\circ\)C for model A and \(-8.0^\circ\)C for model F (see Fig. 2). The convergence of the solution of both models to the stationary solution is very slow. From Figure 6, we can conclude that the relaxation time for \(T_{in} =-10^\circ\)C must be very large and thermal equilibrium in this example is practically unattainable. By the end of the fifth day the outlet temperature in model A equals to \(-5.85^\circ\)C and in model F it equals \(-5.10^\circ\)C. The heat extraction rate \(q_{ex}\) is proportional to the difference \(T_{out}-T_{in}\). Thus, under given conditions, by the end of the fifth day of work, \(q_{ex}\) calculated using model F is 18% more than \(q_{ex}\) calculated using model A. The evolution of soil temperature near the borehole in the beginning five days of operation is shown in Figures 8 and 9. These figures show temperature versus time graphs at nine spatial grid nodes with coordinates \(z\) = 0.25 m and \(r\) = 0.106, 0.120, 0.136, 0.154, 0.175, 0.198, 0.224, 0.254, 0.288 m, from bottom to top (borehole radius \(r_d\) = 0.1 m). These nodes are located in horizontal plane just near the top of the exchanger (\(z\) = 0). Figure 8 shows plots of soil temperature in model A, obtained without taking into account moisture freezing. Here the temperature changes smoothly, decreasing rapidly on the first day and slowing down noticeably at the end of the fifth day. Note that the temperature curves are almost equidistant. This fact is a consequence of the exponential grid in the radial direction. A different view have the soil temperature curves at the same grid nodes in Figure 9, calculated using the model F. When the temperature in the node and in the corresponding cell of the spatial grid decreases to \(T_f-\Delta_f\), moisture begins to freeze in the cell. During freezing, latent heat \(L\) is released, which maintains the temperature in the interval (\(T_f-\Delta_f,\, T_f+\Delta_f)\) in the grid cell and slows down the temperature change in neighboring cells in the radial direction. This phenomenon is known as the zero-curtain effect. The farther the node is from the borehole, the larger the volume of the cell and the more latent heat is released. Therefore, with the distance from the borehole, the freezing time of moisture in the cell increases. Temperature stabilization in the cell for the time of moisture freezing slows down the temperature change not only in the cells with a negative temperature located closer to the borehole wall, but also in several adjacent cells with a positive temperature, which lie farther from the borehole in the same horizontal plane. The release of the latent heat of freezing increases the heat flux to the heat exchanger and increases the temperature of the fluid in the outer pipe, which is confirmed by the graphs in Figure 7. Obviously, the stepped form of the curves is a consequence of the discretization of the problem. With refinement of the spatial grid, the number of steps will increase, their height will decrease, and the curves will be smoothed out. Figure 10 shows color maps of soil temperature near the borehole--in the area of the most rapid temperature change with distance from the borehole. The general slope of the level lines to the right is due to the geothermal gradient. Soil temperature at \(z\) = 100 (13\(^\circ\)C) is three degrees higher than at \(z\) = 0. The breaks in the temperature level lines on the left map (model A) correspond to the boundary between the soil layers with different physical properties. These kinks are also visible on the right map (model F), but here one can also see much more kinks associated with temperature stabilization when moisture freezes. Level lines corresponding to the same temperature in model A are located farther from the borehole than in model F. Therefore, at the end of the fifth day, the amount of heat in the soil calculated by model A is less than in model F. There is no contradiction here with Fig. 4, because in Fig. 4 shows the stationary (final) temperature distribution in the soil, and Fig. 10 shows the unsteady transitional distribution.
## Controlled mode
And, finally, we give an example of a controlled mode of operation of the heat exchanger. Suppose that during the heating season (October 1-April 31) every day the heat exchanger must eject from the soil the amount of heat \(Q_d\), proportional to the difference between the average daily ambient temperature per day \(d\), \(T_a^d\), and room temperature \(T_r=20^\circ\)C, \[Q_d=k_q(T_a^d-T_r),\] where \(k_q\) \[J/K\] is a constant coefficient. For definiteness, let's take the Moscow region. The graph of the average daily air temperature in Moscow is shown in Fig. 11. Let the temperature of the fluid at the inlet to the heat exchanger \(T_{in}\) during the \(d\)-th day be constant and equal to \(T_{in}^d\), then the heat exchanger takes away on the \(d\)-th day the amount of heat \(Q_{ex}^d\) equal to \[Q_{ex}^d(T_{in}^d)=c_fv_f\int_d{(T_{out}(t)-T_{in}^d)\,dt},\] where \(T_{out}(t)\) is obtained from the solution of Eq. ([\[mat\]](#mat){reference-type="ref" reference="mat"}) with \(T_{in}=T_{in}^d\) and the temperature field at the end of the previous day as the initial condition. The integral is taken over \(d\)-th day. To obtain the required amount of heat on each day of the heating season, it is necessary to solve the following chain of equations for \(T_{in}^d\), \[\label{eqq} Q_{ex}^d(T_{in}^d)=Q_d,\; d = 1,2...\] Figure 12 shows the results of calculations of the heating period, performed according to the F and A models for two values of the coefficient \(k_q\), 10\(^4\) and \(2\times 10^4\). The top panel shows plots of extracted heat \(Q_{ex}(t)\). On the scale of the Figure, the graphs of \(Q_{ex}(t)\) and the required heat \(Q_d(t)\) coincide. The lower panel shows graphs of the inlet temperature of the fluid \(T_{in}\) corresponding to the required amount of heat \(Q_d\), calculated by model A and model F. From Fig. 12 it follows that for both considered values of \(k_q\), negative fluid temperatures are necessary to satisfy the required amount of heat extracted from the soil. Note that if at \(k_q=10^4\) the difference in the inlet temperatures of the fluid in the heat exchanger, calculated according to models A and F, is small, then at \(k_q=2\times 10^4\) the maximum difference reaches 37 percent: \(-23.9^\circ\)C in model A and \(-17.4^\circ\)C in model F. The numerical solution of equation ([\[eqq\]](#eqq){reference-type="ref" reference="eqq"}) for the \(d\)-th day was obtained by a method similar to the method of chords for solving nonlinear equations. First, the amount of heat \(Q_1\) extracted at the inlet temperature \(T_1=T_{in}^{d-1}\) is calculated. Then the amount of heat \(Q_2\) extracted from the soil at the inlet temperature \(T_2 = T_1 +\delta\), if \(Q_1>Q_d\), or at \(T_2=T_1-\delta\), if \(Q_1< Q_d\) is calculated. Then \[T_3=T_1+\frac{Q_d-Q_1}{Q_2-Q_1}(T_2-T_1)\] was assigned and the amount of heat \(Q_3\) extracted at the inlet temperature \(T_3\) was calculated. The described iterations stopped when the condition \[|Q_3-Q_d|/Q_d < 0.001.\] Note that in all our calculations, this condition was already satisfied at the first iteration. The temperature field in the soil in the vicinity of the borehole at the end of the heating period is shown in Fig. 13. The temperature is calculated according to the F model, i.e. taking into account the freezing/thawing of moisture, on the left--for \(k_q=10^4\), on the right--for \(k_q=2\times 10^4\). At \(k_q=10^4\), the ice has completely melted by the end of the heating period, and the ground temperature is everywhere positive. At \(k_q=2\times 10^4\), a significant part of the moisture in the soil remains frozen. In the upper part of the borehole, the maximum ice boundary approximately runs at a distance of 0.78 m from the borehole axis. Thus, at the end of the heating season, in both considered cases, the soil temperature near the borehole is significantly lower than the undisturbed soil temperature. Therefore, the question arises of restoring the soil temperature by the beginning of the next heating season. In this paper, we do not consider this issue, but note that the natural temperature recovery due to the influence of boundary conditions cannot be correctly described by our model. Our model does not take into account the convective heat transfer in fluids in the vertical direction. When fluid moves in a working heat exchanger, this transfer can be ignored. However, in a stagnant fluid in the summer, it can have a significant effect on the restoration of the soil temperature near the borehole, in particular, on the melting of adjacent ice.
# Conclusion
The paper proposes a new mathematical model of heat transfer in a shallow vertical coaxial ground heat exchanger and in the surrounding ground. The model takes into account soil heterogeneity and geothermal gradient. At a negative temperature of the heat-carrying fluid, the process of freezing-thawing of ground moisture is taken into account. The water-ice phase transition is modeled using the apparent heat capacity method. The numerical implementation of the model was performed using the finite difference method on a non-homogeneous spatial grid under the assumption of axial symmetry of the problem. The computer implementation of the model is made in the Matlab environment on a personal computer, which makes the model accessible to a wide range of specialists. In the considered examples, three modes of operation of the heat exchanger are modeled: stationary, transitional and controlled. In the calculations, the main attention was paid to demonstrating the differences in the results obtained with and without taking into account the water-ice phase transition in ground moisture. From the above examples, it follows that the greatest difference is noted in the controlled extraction of heat from the soil during the heating season. The proposed model will be useful in the design of heat pump installations for heating and cooling buildings if it is necessary to work with negative temperatures of the coolant, taking into account the structural features and composition of the soil. To simulate the release of heat into the soil, it is enough to change the direction of fluid movement in the model. The model can also be used to optimize the parameters of the heat exchanger, taking into account the specific conditions of its operation. | {'timestamp': '2022-06-06T02:12:56', 'yymm': '2206', 'arxiv_id': '2206.01537', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01537'} |
# Introduction
People in the Information Age read reviews from online review websites when making decisions to buy a product or use a service. The proliferation of such reviews has driven research on opinion mining, where the ultimate goal is to glean information from multiple reviews so that users can make decisions more effectively. Opinion mining has assumed several facets in its history: among others, there are sentiment analysis, that reduces a single review into a sentiment label, opinion extraction, that produces a list of aspect-sentiment pairs representing opinions mentioned in the reviews, and most notably *opinion summarization*, which creates a textual summary of opinions that are found in multiple reviews about a certain product or service. Opinion summarization is arguably the most effective solution for opinion mining, especially when assisting the user in making decisions. Specifically, textual opinion summaries provide users with information that is both more concise and more comprehensible compared to other alternatives. Thus, opinion mining research on the IR community has geared its focus towards opinion summarization in recent years (see Table [1](#tab:sols){reference-type="ref" reference="tab:sols"}). The task of summarizing opinions in multiple reviews can be divided into two subtasks: opinion retrieval and summary generation. Opinion retrieval selects opinions from the reviews that are salient and thus need to be included in the summary. Summary generation produces a textual summary given the retrieved opinions that is concise yet informative and comprehensible for users to read and make decisions effectively. The summary can be generated from scratch with possibly novel tokens (i.e., *abstractive* summarization; ) or spans of text directly extracted from the input (i.e., *extractive* summarization; ). Traditionally, these subtasks correspond to a pipeline of natural language generation models where opinion retrieval and summary generation are treated as content selection and surface realization tasks, respectively. Thanks to advancements in neural networks, most of the recent methods use an end-to-end approach where both opinion retrieval and summary generation are done by a single model optimized to produce well-formed and informative summaries. There are two broad types of challenges in opinion summarization: *annotated data scarcity* and *usability*. As reviews-summary pairs are expensive to create, this has resulted in annotated dataset scarcity. However, the exceptional performance of neural networks for text summarization is mostly driven by large-scale supervised training , which makes opinion summarization challenging. The second challenge--usability--stems from a number of practical requirements for industrial applications. First, for real-world products and service we often need to summarize many thousands of reviews. This is largely infeasible due to the high computational and memory costs of modelling that many reviews with neural architectures . Second, state-of-the-art text summarizers are prone to hallucinations . In other words, a summarizer might mistakenly generate a summary with information not covered by input reviews, thus misinforming the user. Third, generic summaries often cannot address specific user needs. This, in turn, calls for ways to learn summarizers producing personalized summaries.
This opens exciting avenues to develop methods for solving these major challenges in opinion summarization. In this light, the aim of the tutorial is to inform interested researchers and practitioners, especially in opinion mining and text summarization, about recent and ongoing efforts to improve the state of the art and make opinion summarization systems useful in real-world scenarios. And the tutorial will make the audience well-equipped for addressing these challenges in terms of methods, ideas, and related work.
# Tutorial Content and Outline
The tutorial will be 3 hours long and consist of the following five parts, which we describe in detail below.
## Part I: Introduction \[30 min\]
Opinion summarization focuses on summarizing opinionated text, such as customer reviews, and has been actively studied by researchers from the natural language processing and data mining community for decades. There are two major types of opinion summaries: non-textual summaries, such as aggregated ratings , aspect-sentiment tables , and opinion clusters , and textual summaries, which often consist of a short text. Compared to non-textual summaries, which may confuse users due to their complex formats, textual summaries are considered much more user-friendly . Thus, in recent years, the considerable research interest in opinion summarization has shifted towards textual opinion summaries. In this tutorial, we will also focus on recent solutions for generating textual opinion summaries. Like single document summary, textual opinion summary can also be either extractive or abstractive. However, unlike single document summarization, opinion summarization can rarely rely on gold-standard summaries at training time due to the lack of large-scale training examples in the form of review-summary pairs. Meanwhile, the prohibitively many and redundant input reviews also pose new challenges for the task. In this part of the tutorial, we will first describe the opinion summarization task, its history, and the major challenges that come with the task. We will then provide a brief overview of existing opinion summarization solutions.
[\[tab:sols\]]{#tab:sols label="tab:sols"}
## Part II: Solutions To Data Scarcity \[90 min\]
In this part of the tutorial, we will present multiple existing opinion summarization models, as also summarized in Table [1](#tab:sols){reference-type="ref" reference="tab:sols"}. These models attempt to solve the annotated data scarcity problem and are classified into four parts: pre-neural models, autoencoder-based models, models that use synthetic data, and models that leverage low-resource annotated data.
### Autoencoders \[30/90 min\] {#subsubsec:autoencoder}
Due to the lack of training examples, one major approach is to use autoencoders for unsupervised opinion summarization. The autoencoder model consists of an encoder that transforms the input into latent representations and a decoder that attempts to *reconstruct* the original input using a reconstruction objective. It has a wide range of applications in both CV and NLP communities . Autoencoders can also help models obtain better text representations, which allows easier text clustering, aggregation, and selection. Thus, it benefits both extractive and abstractive solutions. In this tutorial, we will first introduce the basics of autoencoders and then describe how to use autoencoders for both extractive and abstractive opinion summarization.
### Synthetic Dataset Creation \[30/90 min\] {#subsubsec:synthetic}
The supervised training of high-capacity models on large datasets containing hundreds of thousands of document-summary pairs is critical to the recent success of deep learning techniques for abstractive summarization . The absence of human-written summaries in a large-scale calls for creative ways to synthesize datasets for supervised training of abstractive summarization models. Customer reviews, available in large quantities, can be used to create synthetic datasets for training. Such datasets are created by sampling one review as a pseudo-summary, and then selecting or generating a subset of reviews as input to be paired with the pseudo-summary. Subsequently, the summarizer is trained in a supervised manner to predict the pseudo-summary given the input reviews. This *self-supervised* approach, as has been shown in a number of works \[[@zhang2020pegasus], *inter alia*\], is effective for training summarizers to generate abstractive opinion summamaaries. In this tutorial, we will introduce various techniques to create synthetic datasets, contrast them, and present results achieved by different works.
### Low-Resource Learning \[30/90 min\] {#subsubsec:fewshot}
Modern deep learning methods rely on large amounts of annotated data for training. Unlike synthetic datasets, automatically created from customer reviews, annotated datasets require expensive human effort. Consequently, only datasets with a handful of human-written summaries are available, which lead to a number of few-shot models. These models alleviate annotated data-scarcity using specialized mechanisms, such as parameter subset fine-tuning and summary candidate ranking. An alternative to human-written are editor-written summaries that are scraped from the web and linked to customer reviews. This setup is challenging because each summary can have hundreds of associated reviews. In this tutorial, we will present both methods that are few-shot learners and that scale to hundreds of input reviews.
## Part III: Improving Usability \[30 min\] {#subsubsec:advanced}
In order to make opinion summarizers more useful in industrial settings, a number of features need to be improved. In this part of the tutorial, we will discuss the following three major features and recent solutions the community has proposed:
- **Scalability:** The ability to handle a massive number of input reviews. To handle large scale input, the ability to retrieve salient information, e.g., reviews or opinions, becomes a important yet challenging feature for opinion summarization solutions.
- **Input Faithfulness:** The ability of a summarizer to generate summaries covered in content by input reviews. In other words, the summarizer should not confuse entities or introduce novel content into summaries.
- **Controllability:** The ability to produce constrained summaries, such as a hotel summary that only includes room cleanliness or a product summary that only covers the negative opinions.
## Part IV: Evaluation and Resources \[20 min\]
As is common in other areas of natural language processing, in opinion summarization, researchers often rely on automatic metrics. These metrics, such as ROUGE, are based on word overlaps with the reference summary. However, word overlap metrics are limited and can weakly correlate with human judgment. To address these shortcomings, human evaluation is often used, where human annotators assess various aspects of generated summaries. In this tutorial, we will present different kinds of human evaluation experiments, how they are designed, and how they are performed.
## Part V: Future Work \[10 min\]
To conclude the tutorial, we will present several notable open questions for opinion summarization, such as the need for additional annotated resources, common issues with the generated summary (e.g., repetition, hallucination, coherency, and factuality), and the ability to handle various type of input data (e.g., images and knowledge bases). Based on these open questions, we will also present future work on opinion summarization.
# Objectives
In this tutorial, we will cover a wide range of techniques from pre-neural approaches to the most recent advances for opinion summarization. In addition, we will also introduce the commonly used resources and evaluation metrics. Our goal for this tutorial is to increase the interest of the IR community towards the opinion summarization problem and help researchers to start working on relevant problems.
# Relevance to the IR community
Sentiment analysis has been a major research area in the IR community. Since the tutorial will cover cutting-edge research in the field, it would attract a wide variety of IR researchers and practitioners. We would also like to emphasize that the interest in opinion mining and summarization techniques in the IR community has been rapidly and significantly increased in recent years. To the best of our knowledge, we are the first ones to offer a tutorial covering the series of recent opinion summarization approaches.[^1]
# Broader Impact
Methods presented in the tutorial also have applications beyond customer reviews. The amount of opinions on various topics expressed online is vast. These opinions address various sports, politics, and public events. In turn, this calls for ways to summarize this information for the benefit of the user. As we will discuss, the methods presented in the tutorial can be applied to other opinion domains, such as social media and blogs.
# Relevance to ACL
The models we will discuss in this tutorial were originally developed for opinion summarization, but the techniques can be used for many other NLP tasks such as aspect-based sentiment analysis, opinion extraction, unsupervised text generation, multi-document, and long-form summarization. The tutorial also covers open-ended problems and future directions that are also relevant in other tasks such as personalization, multimodality, and multilinguality, among others.
# Tutorial Type
This is a **cutting-edge** tutorial that primarily covers recently published papers. More than 80% of the papers were published within the last two years. The techniques that will be discussed in the tutorial, especially abstractive opinion summarization methods, are currently active and emerging research areas.
# Breadth
At least \(70\%\) of papers covered by the tutorial are not authored by the tutorial presentors.
# Prerequisites
The tutorial is self-contained, which will cover essential background knowledge for each technique. We assume that the audience has basic knowledge of Machine Learning and NLP. More advanced knowledge in Neural Networks such as encoder-decoder models, the attention mechanism, and pre-trained language models are preferred.
# Reading List
Reading the following papers will be helpful for attendees but are not necessary:
- Opinion Mining:
- Text summarization:
- Opinion Summarization:
# Target Audience
We believe that we are the first to offer a tutorial that covers the series of recent opinion summarization models. Thus, we expect the tutorial to attract a wide variety of IR researchers and practitioners, especially in the areas of opinion mining and text summarization. We estimate that at least 100 people will attend the tutorial. To the best of our knowledge, there is no tutorial that covers a similar set of recent works for opinion summarization.
# Open Access
We will publicly share all the materials used during the tutorial, including presentation slides and software.
# Instructors
*Reinald Kim Amplayo* is a Research Scientist at Google. He received his PhD from the University of Edinburgh, where his thesis focused on controllable and personalizable opinion summarization. He is a recepient of a best student paper runner-up at ACML 2018. *Arthur Bražinskas* is a Research Scientist at Google working on natural language generation for Google Assistant. His PhD on low-and high-resource opinion summarization is supervised by Ivan Titov and Mirella Lapata at the University of Edinburgh. *Yoshi Suhara* is an Applied Research Scientist at Grammarly. Previously, he was a Senior Research Scientist at Megagon Labs, an Adjunct Instructor at New College of Florida, a Visiting Scientist at the MIT Media Lab, and a Research Scientist at NTT Laboratories. He received his PhD from Keio University in 2014. His expertise lies in NLP, especially Opinion Mining and Information Extraction. *Xiaolan Wang* is a Senior Research Scientist at Megagon Labs. She received her PhD from University of Massachusetts Amherst in 2019. Her research interests include data integration, data cleaning, and natural language processing. She co-instructed the tutorial, *Data Augmentation for ML-driven Data Preparation and Integration*, at VLDB 2021. *Bing Liu* is a Distinguished Professor of Computer Science at the University of Illinois at Chicago (UIC). He has published extensively in top conferences and journals. He also authored four books about lifelong learning, sentiment analysis and Web mining. Three of his papers received Test-of-Time awards: two from SIGKDD and one from WSDM. He has served as the Chair of ACM SIGKDD from 2013-2017, as program chair of many leading data mining conferences, including KDD, ICDM, CIKM, WSDM, SDM, and PAKDD, and as associate editor of leading journals such as TKDE, TWEB, DMKD and TKDD. He is a recipient of ACM SIGKDD Innovation Award, and he is a Fellow of the ACM, AAAI, and IEEE.
[^1]: More than 80% of the papers were published within the last three years. | {'timestamp': '2022-06-06T02:13:13', 'yymm': '2206', 'arxiv_id': '2206.01543', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01543'} |
null | null |
# Introduction
Probing the large scale structure of the Universe is a crucial step towards precision cosmology as we try to constrain the nature of dark energy, non-Gaussian fluctuations in the Universe's primordial density field, and test general relativity. Typically, this is done using galaxy surveys with spectroscopic or photometric redshifts in the optical or near-infrared. At radio wavelengths, we use the redshifted neutral hydrogen ([H[i]{.smallcaps}]{.roman}) hyperfine transition line, with a rest-frame wavelength of 21cm, to measure redshift. Given the ubiquitous nature of [H[i]{.smallcaps}]{.roman} in the Universe, we can use it to trace the distribution of dark matter at low and high redshifts. The faintness of the [H[i]{.smallcaps}]{.roman} emission line makes it challenging to resolve individual galaxies at higher redshifts over large volumes. However, for cosmology we are interested in the bulk fluctuations on large (Mpc) scales, so we can use the [H[i]{.smallcaps}]{.roman} intensity mapping technique. This technique relaxes the requirement of galaxy detection by integrating all 21cm emission within relatively large spatial voxels. This delivers high survey speeds over large volumes, providing a novel solution to the current challenges of observational cosmology. One of the main challenges in detecting the [H[i]{.smallcaps}]{.roman} intensity mapping signal is the presence of foregrounds that are orders of magnitude brighter. Removing these requires precise instrumental calibration. Cross-correlating with galaxy surveys helps to mitigate residual systematics from foregrounds, Radio Frequency Interference (RFI) and thermal noise. Moreover, it can improve constraints on cosmological parameters and provide insight into the [H[i]{.smallcaps}]{.roman} astrophysics of the correlated galaxies. Currently, the [H[i]{.smallcaps}]{.roman} intensity mapping signal has only been detected in cross-correlation with galaxy surveys. Both MeerKAT and the future SKA Observatory (SKAO) have been put forward as state-of-the-art intensity mapping instruments capable of complementing and extending cosmological measurements at other wavelengths. Using the single-dish data from each element of the array, we can access the large cosmological scales inaccessible by the interferometer due to its lack of very short baselines. In we presented the first calibrated maps of a pilot MeerKAT survey, which utilised this single-dish method on the multi-dish interferometer array. In this article, we use this data to measure the cross-correlation power spectrum between the [H[i]{.smallcaps}]{.roman} signal and overlapping WiggleZ data. With only \(10.5\,\text{hrs}\) of data for each of the 64 dishes over an effective survey area of \({\sim}\,200\,\text{deg}^2\), this detection shows the power of this approach and paves the way towards probing large cosmological scales with much larger surveys on MeerKAT and SKAO. The paper is structured as follows; in we introduce the data products used in this study. introduces our approach to foreground cleaning in the MeerKAT intensity maps. The formalism adopted for the power spectrum estimation and modelling is discussed in . We present our main results in and finally conclude in .
# Observational data {#sec:ObsData}
The MeerKAT [H[i]{.smallcaps}]{.roman} intensity mapping data was collected using the L-band with \(10.5\) hours of observation per dish over six nights between February and July 2019. The L-band has 4096 frequency channels at \(856-1712\,\text{MHz}\), but in this work, we only use 199 channels at \(973.2-1014.6\,\text{MHz}\) (\(0.400\,{<}\,z\,{<}\,0.459\)). Of these channels, a further 32 are removed due to their dominant contributions to the eigenmodes of the principal component analysis (see later discussion in ) for those particular channels. This aggressive strategy can be seen as a final RFI-flagging stage[^1] and was necessary in order for us to obtain a cross-correlation detection. The survey area spans around \(200\,{\rm deg}^2\) and targeted the WiggleZ 11hr field, covering \(153^\circ\,{<}\,\text{R.A.}\,{<}\,172^\circ\) and \(-1^\circ\,{<}\,\text{Dec.}\,{<}\,8^\circ\). In this analysis, only the auto-correlation of MeerKAT interferometer visibility data, i.e. the single dish mode observation, is used. The time-ordered data for each dish, \(d\), are projected into the map space via the map-making process, \[\begin{aligned}
\hat{m} = \left(A^\text{T} N^{-1} A\right)^{-1} A^\text{T} N^{-1} d \end{aligned}\] in which \(A\) is the pointing matrix mapping the time-ordered data to the map coordinates and \(N\) is the noise covariance matrix between time stamps. As described in, the time-ordered data (TOD) is calibrated against a noise diode which is fired every \(20\) seconds. This removes long-term noise correlation, so-called \(1/\text{f}\) noise, due to receiver chain gain variations on time scales longer than 20 s. On shorter time scales, \(1/\text{f}\) noise is negligible compared to thermal noise fluctuations. We also subtract the average signal every 220 s in the TOD which suppresses long-term sloping gain changes. This should reduce the overall variance of the signal but can potentially have the adverse effect of removing [H[i]{.smallcaps}]{.roman} signal over large angular scales. In this work we assume any signal loss from this process is sub-dominant relative to the signal loss from foreground cleaning and thus do not attempt any reconstruction. The noise covariance matrix \(N\) is assumed to be diagonal with constant variance during the observation, but we allow the variance to differ between dishes. The noise covariance is also projected to the map space via, \[\begin{aligned}
\hat{n} = \left(A^\text{T} N^{-1} A\right)^{-1}, \end{aligned}\] where \(\hat{n}\) is the pixel noise variance. The inverse of the pixel noise variance, \(w_\textrm{H\textsc{i}}\,{=}\,1/\hat{n}\), is used as the inverse-variance weight in the analysis[^2]. We use the flat-sky approximation and grid the map into square pixels with a width of \(0.25^\circ\). To create the final maps, we average over all individual dish maps (full details of the single-dish calibration are provided in ). The observed intensity maps are shown in . The top panel shows the frequency-averaged intensity map before foreground cleaning. The bottom panel shows the residual map after foreground cleaning, a process we outline in the following section.
# Foreground Cleaning {#sec:FGcleaning}
Here we discuss the foreground cleaning performed on the MeerKAT intensity maps. We provide detailed descriptions on each stage in the following sub-sections but begin with a summary of the foreground cleaning method we adopt. The foreground cleaning is performed using a blind Principal Component Analysis (PCA) method, which relies on the foregrounds being the dominant signal and correlated in frequency. Thus, by removing the first \(N_\text{fg}\) principal component modes in frequency from each pixel, the majority of their contribution is suppressed (see for further discussion). Before cleaning, the map is resmoothed using a Gaussian window function with kernel size \(1.2\) times the largest beam size within the frequency range (see for details). Foreground cleaning is imperfect, and the cleaned maps contain residual foreground structure which can be seen by comparing both panels in . However, the amplitude of the map has decreased by several orders of magnitude, thus the foreground residuals should dominate less over the [H[i]{.smallcaps}]{.roman} fluctuations. Furthermore, some [H[i]{.smallcaps}]{.roman} signal will be removed, typically on larger scales where modes are most degenerate with the spectrally smooth foregrounds. We aim to reconstruct this lost signal with a foreground transfer function, which we discuss in .
## Reconvolution of maps {#sec:Resmooth}
It is understood that a frequency-dependent beam size can cause the foregrounds to leak into a greater number of spectral modes, requiring more aggressive cleaning. A way to potentially mitigate this issue is to convolve all maps to a common resolution before performing the foreground clean, as done in previous experiments. However, recent tests on simulations suggest that a simple Gaussian resmoothing of the data to a common resolution does not improve blind foreground removal techniques, even if the true beam is a perfect Gaussian. For real data though, it is beneficial to resmooth to homogenise some of the systematic contributions from e.g. residual RFI or polarisation leakage. For this reason we perform a weighted resmoothing on the MeerKAT [H[i]{.smallcaps}]{.roman} intensity maps prior to foreground cleaning. An intensity map \(\delta T^\prime\) which has a frequency dependent beam (denoted by the \({^\prime}\) index) with a full-width-half-maximum \(\theta_\text{FWHM}(\nu)\) in degrees, and an angular separation between pixels given by \(\Delta\theta\), is convolved with the following kernel: \[\label{eq:ResmoothKernel} K(\Delta \theta,\nu)=\exp \left[-\frac{\Delta \theta^{2}}{2[\gamma \sigma_{\max }^{2}-\sigma^{2}(\nu)]}\right]\,,\] where \(\sigma(\nu)=\theta_{\mathrm{FWHM}}(\nu) /(2 \sqrt{2 \ln 2})\), \(\sigma_\text{max}\) is the maximum \(\sigma(\nu)\) value and \(\gamma\) is a scaling factor which governs how much the final effective resolution is decreased by. In previous Green Bank Telescope (GBT) studies, a choice of \(\gamma\,{=}\,1.4\) was used. In this work, due to the already large MeerKAT beam, we use a smaller value of \(\gamma\,{=}\,1.2\). We experimented with this choice, finding \(\gamma\,{=}\,1\) produced a noisier power spectrum with worse model agreemen. The higher choice of \(\gamma\,{=}\,1.4\) delivered a similarly good model agreement compared to \(\gamma\,{=}\,1.2\), but had a slightly lower cross-correlation detection significance due to the increased damping at high-\(k\). Not performing any reconvolution still delivered a clear detection but resulted in a particularly noisy power spectrum at small-\(k, indicating the presence of residual foreground and systematics which are mitigated by this resmoothing procedure. The kernel in \autoref{eq:ResmoothKernel} is normalised such that the sum over all pixels is equal to unity, then the weighted convolution used to resmooth the maps is given by \begin{equation} \delta T(\theta,\nu) = \frac{\left[\delta T^\prime(\theta,\nu)\,w_\textrm{H\textsc{i}}^\prime(\theta,\nu)\right] * K(\Delta\theta,\nu)}{w_\textrm{H\textsc{i}}^\prime(\theta,\nu) * K(\Delta\theta,\nu)}\,, \end{equation} where\)w\_[H[i]{.smallcaps}]{.roman}\^()\(is the inverse variance weight. The\)\*\(denotes a convolution performed separately in each frequency channel e.g.\)w\_[H[i]{.smallcaps}]{.roman}\^() \* K() = \_i w\_[H[i]{.smallcaps}]{.roman}\^(\_i)K(-\_i)\(. To ensure the weight field still represents the inverse variance of the new resmoothed field, the weight\)w\_[H[i]{.smallcaps}]{.roman}\^\(is convolved according to \begin{equation} w_\textrm{H\textsc{i}}(\theta,\nu) = \frac{\left[w_\textrm{H\textsc{i}}^\prime(\theta,\nu) * K(\Delta\theta,\nu)\right]^2}{w_\textrm{H\textsc{i}}^\prime(\theta,\nu) * K^2(\Delta\theta,\nu)}\. \end{equation} \subsection{PCA foreground cleaning}\label{sec:PCA} In this work, a PCA-based blind foreground subtraction method is used. The observed intensity maps can be represented by a matrix\)\_\(with dimensions\)N\_ N\_\(where\)N\_\(is the number of frequency channels along the line-of sight and\)N\_\(is the number of pixels. The assumption behind blind-foreground cleaning is that the data can be represented by the linear system\)\_ = + \(, where\[represents the mixing matrix and\]are the\)N\_\(separable source maps identified by projecting the mixing matrix along the data\) = \^\_\(. In PCA, the mixing matrix is extracted from the eigen-decomposition of the frequency covariance matrix of the mean-centred data, defined by\) = (\_)\^(\_)/(N\_ - 1)\(, where\[are the inverse variance weights recast into\)N\_ N\_\(matrices. The eigen-decomposition is then given as\)=\(, where\]is the diagonal matrix of eigenvalues ordered by descending magnitude, and\[are the eigenvectors, the first\)N\_\(of which supplies the set of functions used to construct the mixing matrix. We assume the subtraction of\]in the linear system will remove dominant foregrounds, leaving behind in the residuals\[any \textrm{H\textsc{i}}\ signal not removed in the subtraction. \subsection{Transfer function construction}\label{sec:TF} We compensate the signal loss due to the foreground cleaning with a foreground transfer function. Following previous literature, the transfer function can be constructed by injecting mock intensity mapping data\)\_[H[i]{.smallcaps}]{.roman}\(into the true observed data\)\_\(, which includes foregrounds and observational systematics. By running a PCA clean on this combination, we can measure (and compensate for) the signal loss in the cleaned mock data; \begin{equation}\label{eq:MockforTF} \mathbf{M}_\text{c} = [\mathbf{M}_\textrm{H\textsc{i}} + \mathbf{X}_\text{obs}]_\text{PCA}-[\mathbf{X}_\text{obs}]_\text{PCA}\. \end{equation} The\)\[ \]\_\(notation represents performing the PCA clean (outlined in \secref{sec:PCA}) on the quantities inside the brackets, treating them as a single combination. For example, the mixing matrix is not determined separately for both mock and data in\)\[\_[H[i]{.smallcaps}]{.roman} + \_\]\_\(, but determined for the combination of\)\_[H[i]{.smallcaps}]{.roman} + \_\(. We also subtract the PCA clean of the data,\)\[\_\]\_\(, since this only adds uncorrelated variance, thus subtracting it makes convergence to a smooth transfer function more efficient, requiring fewer mock iterations. After calculating\)\_\(, we measure the cross-power spectrum with a corresponding mock galaxy map\)\_\(, then divide this by a foreground-free equivalent to estimate the signal loss at each mode; \begin{equation}\label{eq:TF} \mathcal{T}(k) = \left\langle \frac{\mathcal{P}(\mathbf{M}_\text{c}\, ,\, \mathbf{M}_\text{g})}{\mathcal{P}(\mathbf{M}_\textrm{H\textsc{i}}\, ,\, \mathbf{M}_\text{g})} \right\rangle \,. \end{equation}\)()\(denotes an operator which measures the cross-power spectrum, then spherically averages modes into the same\)k\(-bins as the data. The angled brackets represent an ensemble average over a large number of mocks (we use 100 in this work). The \textrm{H\textsc{i}}\ mocks are generated with the lognormal method, sampled from a model \textrm{H\textsc{i}}\ power spectrum (the same as that used in the power spectrum fitting, see \secref{sec:Model}) with a Gaussian smoothing applied perpendicular to the line-of-sight to approximately emulate the MeerKAT beam. Similarly, for the galaxy mocks, we generate a lognormal density field with the same random seed as the \textrm{H\textsc{i}}, then Poisson sample galaxies onto the field with the same number count as the WiggleZ catalogue. These steps ensure the mock fields emulate the amplitudes of real maps as realistically as possible. There is evidence suggesting the transfer function is not overly sensitive to the choice of fiducial cosmology, but further investigation into how much it can be relied on for precision cosmology is required. Any power spectrum measurement we make on the data is divided by\)(k)\(to correct for the signal loss (unless clearly stated in demonstrative figures). For the \textrm{H\textsc{i}}\ auto-power spectrum used in the error estimation (discussed later in \secref{sec:Pkest}), we also multiply through by\)`<!-- -->`{=html}1/(k)\(. Previous studies have opted to use\)`<!-- -->`{=html}1/(k)\^2\(as a correction instead, motivated by the assumption that in auto-correlation, signal loss occurs in both maps, so there should be twice the reconstruction of power needed. However, from simulation tests, we found this over-corrected the signal loss. Furthermore, our analytical error estimation on the cross-power spectrum (see \secref{sec:Pkest}), which uses the auto-\textrm{H\textsc{i}}\ power spectrum, is found to be in good agreement with other approaches of error estimation using the WiggleZ randoms and jackknife tests.\)\^2\_\(analysis also suggests our errors are not over-estimated in any case. This changes if we opt for the\)`<!-- -->`{=html}1/(k)\^2\(correction where it becomes clear that the errors have been over-estimated, suggesting that the signal loss in the auto-\textrm{H\textsc{i}}\ power spectrum has been over-corrected. We defer further investigation into signal loss in the \textrm{H\textsc{i}}\ auto-correlation to future work. \section{Power spectrum estimation \& modelling}\label{sec:Pkest} The power spectrum estimation process we adopt is based on the optimal weighting method outlined in (see for applications to \textrm{H\textsc{i}}\ intensity mapping cross-correlations). We define the two Fourier transformed fields of the \textrm{H\textsc{i}}\ temperature fluctuation maps\)T\_[H[i]{.smallcaps}]{.roman}\(and the galaxy count field\)n\_\(as \begin{equation} \tilde{F}_\textrm{H\textsc{i}}(\mathbf{k})=\sum_{\mathbf{x}} \delta T_\textrm{H\textsc{i}}(\mathbf{x}) w_\textrm{H\textsc{i}}(\mathbf{x}) \exp (i \mathbf{k}{\cdot}\mathbf{x}) \end{equation} \begin{equation} \tilde{F}_\text{g}(\mathbf{k})=\sum_{\mathbf{x}} n_\text{g}(\mathbf{x}) w_\text{g}(\mathbf{x}) \exp (i \mathbf{k}{\cdot}\mathbf{x})-N_\text{g}\tilde{W}_\text{g}(\mathbf{k})\,, \end{equation} where\)N\_ = n\_\(is the total number of galaxies in the optical map and\)\_\(is the weighted Fourier transform of the selection function (\)W\_\() which is normalised such that\)\_W\_() = 1\(; \begin{equation} \tilde{W}_\text{g}(\mathbf{k})=\sum_{\mathbf{x}} W_\text{g}(\mathbf{x}) w_\text{g}(\mathbf{x}) \exp (i \mathbf{k}{\cdot}\mathbf{x})\. \end{equation} The selection function\)W\_\(accounts for incompleteness in the WiggleZ survey and is constructed by stacking the random catalogues generated in, reproducing the varying target and redshift completeness. The map weights in the above equations are the inverse variance map for the \textrm{H\textsc{i}}\ field\footnote{Optimal weights for \textrm{H\textsc{i}}\ intensity mapping are derived in and we leave it to future work on data with higher \)S/N\( to experiment with these.},\)w\_[H[i]{.smallcaps}]{.roman}() = 1/()\((defined in \secref{sec:ObsData}) and for the galaxies we use the optimal weighting as per \begin{equation}\label{eq:FKPweight} w_\text{g}(\mathbf{x}) = 1\bigg/\left(1+\frac{W_\text{g}(\mathbf{x})N_\text{g}P_0}{V_\text{cell}}\right)\,, \end{equation} where\)P_0\(is a chosen value of the power spectrum at the scale where an optimal measurement is sought. We use\)P_0 = 5000 h\^-3\^3\(, approximately corresponding to\)k = 0.1 h \^-1\(.\)V\_\(is the volume of a grid cell in\)h\^-3\^3\(. The maps are enclosed onto a cuboid in comoving space, using a fiducial Planck18 cosmology with sides of length\)(l\_,l\_,l\_) \~ (812,264,176) /h\(, divided into\)(n\_,n\_,n\_) = (151,49,250)\(equal sized cells. Assuming a flat-sky approximation, and using the map pixel's angular coordinates and redshifts, we grid the data using a nearest grid point assignment\footnote{The scales that we probe and the \)S/N\( of this data means potential impact from aliasing effects are sub-dominant and we do not need to correct for them.}. To taper the edges along the line-of-sight, we apply a Blackman window along the z-axis of the galaxy and \textrm{H\textsc{i}}\ intensity maps, as well as their weights and selection functions. The cross-correlation between the \textrm{H\textsc{i}}\ intensity maps and the galaxies is then estimated with \begin{equation} \small \hat{P}_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}) = \frac{V_\text{cell}}{\sum\limits_\mathbf{x} w_\textrm{H\textsc{i}}(\mathbf{x})w_\text{g}(\mathbf{x})W_\text{g}(\mathbf{x})}\operatorname{Re}\left\{\tilde{F}_\textrm{H\textsc{i}}(\mathbf{k}){\cdot} \tilde{F}^{*}_\text{g}(\mathbf{k})\right\}\frac{1}{N_\text{g}}\. \end{equation} Similarly, the \textrm{H\textsc{i}}\ and galaxy auto-power spectra, which are needed for the error estimation (see below), are given by \begin{equation}\label{eq:HIautoPk} \hat{P}_\textrm{H\textsc{i}}(\mathbf{k}) = \frac{V_\text{cell}}{\sum_\mathbf{x} w_\textrm{H\textsc{i}}^2(\mathbf{x})}\lvert\tilde{F}_\textrm{H\textsc{i}}(\mathbf{k})\rvert^2\,, \end{equation} \begin{equation} \hat{P}_\text{g}(k) = \frac{V_\text{cell}}{\sum_\mathbf{x} w^2_\text{g}(\mathbf{x})W^2_\text{g}(\mathbf{x})} \left[|\tilde{F}_\text{g}(\mathbf{k})|^2-S_\text{g}\right]\frac{1}{N^2_\text{g}}\,, \end{equation} where\)S\_\(accounts for the shot noise in the galaxy survey, given as \begin{equation} S_\text{g} = N_\text{g}\sum_\mathbf{x} w_\text{g}^2(\mathbf{x})W_\text{g}(\mathbf{x})\. \end{equation} These power spectra are all spherically averaged into bandpowers\) k\(to provide the final 1D power spectra results (the cross-and \textrm{H\textsc{i}}-auto power spectra are also corrected with the\)`<!-- -->`{=html}1/(k)\(as discussed in \secref{sec:TF}). For noise-dominated \textrm{H\textsc{i}}\ intensity maps, the errors for the cross-power can be estimated analytically from \begin{equation}\label{eq:Pkerr} \small \hat{\sigma}_{\textrm{H\textsc{i}},\text{g}}(k)=\frac{1}{\sqrt{2 N_\text{m}(k)}} \sqrt{\hat{P}^2_{\textrm{H\textsc{i}},\text{g}}(k)+\hat{P}_\textrm{H\textsc{i}}(k)\left(\hat{P}_\text{g}(k)+\frac{1}{\bar{n}_\text{g}}\right)}\,, \end{equation} where\)N\_\(is the number of modes in each\)k\(-bin and\)\|n\_ = N\_/(l\_ l\_ l\_)\(is the number density of galaxies. The\)`<!-- -->`{=html}1/\(factor in \autoref{eq:Pkerr} appears because this is the error on a cross-correlation of two fields, so the number of independent pairs available to measure the variance on the mean doubles. We compared these analytical error estimations to ones calculated from cross-correlating the MeerKAT data with the random WiggleZ catalogues used to derive the selection function, finding very good agreement across all scales. Furthermore, we also found good agreement with an internal error estimation which used a jackknife approach. \subsection{Theoretical modelling}\label{sec:Model} We fit a model to the cross-power spectrum which is given by \begin{multline}\label{eq:Pkmodel} P_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}) = \overline{T}_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} b_\text{g} r (1+f\mu^2)^2 \,P_\text{m}(k) \\ \times\, \exp\left[\frac{-(1-\mu^2)k^2 R_\text{beam}^2}{2}\right]\,, \end{multline} where\)\_[H[i]{.smallcaps}]{.roman}\(is the mean \textrm{H\textsc{i}}\ temperature of the field in mK,\)b\_[H[i]{.smallcaps}]{.roman}\(and\)b\_\(are the \textrm{H\textsc{i}}\ and galaxy biases and\)r\(is the cross-correlation coefficient. We account for linear redshift-space distortions (RSD) with the\)(1+f\^2)\^2\(factor, where\)f\(is the growth rate of structure and\]is the cosine of the angle from the line-of-sight.\)P\_\(is the matter power spectrum produced using \texttt{CAMB} with a Planck18 cosmology. The exponential factor approximates the smoothing of perpendicular modes due to the MeerKAT beam, where\)R\_\(is the standard deviation of the Gaussian beam profile in comoving units, taking into account the reconvolution as explained in \secref{sec:Resmooth}, which gives\)R\_ = 13.3 h\^-1\(. To fully account for RSD, the model in \autoref{eq:Pkmodel} should be (not including the beam damping for brevity) \begin{equation} P_{\textrm{H\textsc{i}} ,\text{g}}(\mathbf{k}) = \overline{T}_\textrm{H\textsc{i}}\left[r b_\textrm{H\textsc{i}} b_\text{g} + b_\textrm{H\textsc{i}} f \mu^2 + b_\text{g} f \mu^2 + f^2 \mu^4\right] P_\text{m}(k)\,, \end{equation} which accounts for the biases appearing in cross-terms from the expansion of the two fields in redshift space\)\^\_[H[i]{.smallcaps}]{.roman}() = b\_[H[i]{.smallcaps}]{.roman} \_() + f ()\(and\)\^\_() = b\_ \_() + f ()\(, where\)()\(is the velocity divergence field\footnote{The cross-correlation coefficient \)r\( only enters on cross-correlation between biased density terms.}. However, we are only attempting a fit to the spherically-averaged power spectrum monopole, which is uniformly averaged across\[. This would make\)b\_[H[i]{.smallcaps}]{.roman}\(perfectly degenerate with\)\_[H[i]{.smallcaps}]{.roman} r\(. To break this degeneracy we would need to introduce an anisotropic sensitivity on\]in our analysis, achieved by modelling the quadrupole. This would require a higher\)S/N\(than we have available from the MeerKAT pilot survey data. This is why we follow previous literature and probe the degenerate quantity\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(, but include a matter-only RSD, to avoid biasing the amplitude of the power spectrum by the\)(1+ f\^2) \~ 1.7\(Kaiser term. The model in \autoref{eq:Pkmodel} is discretised onto the same 3D grid of modes as the data and then convolved with the survey window functions \begin{multline}\label{eq:ConvolvedModel} P_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}) \rightarrow P_{\textrm{H\textsc{i}},\text{g}} * W_\textrm{H\textsc{i}} W_\text{g} = \\ \frac{\sum_{i} P_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}_{i}^{\prime}) \operatorname{Re}\left\{\tilde{W}_\text{g}(\mathbf{k}-\mathbf{k}_{i}^{\prime}) \tilde{W}_\textrm{H\textsc{i}}(\mathbf{k}-\mathbf{k}_{i}^{\prime})^{*}\right\}}{\sum w_\text{g}(\mathbf{x}) w_\textrm{H\textsc{i}}(\mathbf{x}) W_\text{g}(\mathbf{x}) W_\textrm{H\textsc{i}}(\mathbf{x})}\. \end{multline} In lieu of a precisely constructed survey selection function\)W\_[H[i]{.smallcaps}]{.roman}\(for the \textrm{H\textsc{i}}\ intensity maps, we use a simple binary window function that is 1 wherever a pixel is filled and 0 otherwise. The convolved model in \autoref{eq:ConvolvedModel} is spherically averaged into the same\)k\(-bins as the data. The mean \textrm{H\textsc{i}}\ temperature\)\_[H[i]{.smallcaps}]{.roman}\(can be recast to the \textrm{H\textsc{i}}\ density fraction (\)\_[H[i]{.smallcaps}]{.roman}\() using \begin{equation}\label{eq:TbarModelEq} \overline{T}_\textrm{H\textsc{i}}(z) = 180\,\Omega_{\textrm{H\textsc{i}}}(z)\,h\,\frac{(1+z)^2}{\sqrt{\Omega_\text{m}(1+z)^3 + \Omega_\Lambda}} \, {\text{mK}} \,, \end{equation} where\)\_\(and\)\_\(are the density fractions for matter and the cosmological constant, respectively. Thus fitting the amplitude of the cross-power spectrum allows us to constrain\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(. When fitting\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(to the power spectrum data using the model in \autoref{eq:Pkmodel}, we fix the galaxy bias (\)b\_\() and growth rate (\)f\(), since they are well constrained from other experiments relative to the other parameters. We assume\)f = 0.737\((based on\)f \~ \_(z)\^\(, where\) = 0.545\() and\)b\_ = 0.911\(at the central redshift of our data (\)z\_ = 0.43\(). \section{Results}\label{sec:Results} \subsection{\textrm{H\textsc{i}}\ auto power spectrum} The auto power spectrum of the cleaned \textrm{H\textsc{i}}\ intensity maps gives some indication of how much foregrounds have been suppressed. In \autoref{fig:HIAutoPk} we show the 2D auto-\textrm{H\textsc{i}}\ power spectrum decomposed into anisotropic\)k\(-bins perpendicular (\)k\_ = \() and parallel (\)k\_ k\_\() to the line-of-sight. The far-left panel shows the original data before foreground removal, demonstrating the dominance of the foregrounds and their concentration on the largest scales, particularly at small-\)k\_\(. Removing just a few principal components reduces the amplitude of the auto power spectrum by several orders of magnitude, as shown in the second panel. In simulation tests, removing\)N\_ \~ 4\(PCA modes is sufficient to remove the majority of the foregrounds, and this should also be the case for a perfectly designed and calibrated experiment. The resulting cleaned maps in this idealised case would contain \textrm{H\textsc{i}}\ signal and Gaussian thermal noise, similar to the model in the fifth panel of \autoref{fig:HIAutoPk}. The thermal noise in the ideal model is large due to the survey's low observing time. This yields an auto power spectrum amplitude of\)P\_ \~ 85 \^2h\^-3\^3\(, which is expected to dominate over the \textrm{H\textsc{i}}\ signal\footnote{We assume \)\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}}\,{=}\,0.85\,{\times}\,10^{-3}\( for the \textrm{H\textsc{i}}\ model and simulation in \autoref{fig:HIAutoPk}.} on most scales, as seen in the \autoref{fig:HIAutoPk} model. For this pilot-survey data, there will be instrumental calibration imperfections, residual RFI and other systematics (see discussions in ). The instrumental response modulates the foregrounds, resulting in additional spectral structure that requires more PCA modes to be removed. Uncleaned modes containing residual foregrounds, RFI and other systematics will positively bias the auto-power spectrum\footnote{This assumes these systematics have not caused the gain to be systematically overestimated.}. This is why for the\)N\_ = 4\(case, we do not reach the level of the \textrm{H\textsc{i}}\ signal plus noise. As\)N\_\(increases in \autoref{fig:HIAutoPk}, the amplitude of the power drops, and this occurs more severely for large modes, particularly at small-\)k\_\(. While it appears that we are reaching below the idealised \textrm{H\textsc{i}}\ + noise model at high-\)k\_\(and low-\)k\_\(in the\)N\_ 20\(cases, we would need to account for the effects of the map reconvolution and the foreground clean, both of which would damp the \textrm{H\textsc{i}}\ + noise model further. In the far-right panel we show a simulated \textrm{H\textsc{i}}\ mock with the same noise level as the idealised model, but include some observational effects. To emulate the beam and reconvolution, we smooth the simulation perpendicular to the line-of-sight. We also emulate signal loss from the foreground clean by projecting out modes based on the same PCA mixing matrix functions derived for the\)N\_ = 30\(data case. Comparison between the\)N\_ = 30\(and far-right panel suggests we have not reached the estimated auto-power spectrum level, indicating that residual RFI, foregrounds, and other systematics are present in the data. Since the additive bias from non-cosmological residuals is unknown, it is difficult to compare auto-correlated data and model and conclude that a cosmological detection has been achieved. Cross-correlating with galaxy surveys avoids these additive biases, serving the motivation for this work. We leave a detailed study into the auto-\textrm{H\textsc{i}}\ power for future work, where we will explore cross-correlating different sets of dishes or observational time-blocks (a method adopted in GBT experiments ). \subsection{Post-foreground clean signal reconstruction} Along with residuals from an imperfect cleaning process, foreground removal will also cause cosmological signal loss in the intensity maps, which requires reconstruction to avoid negatively biasing the measured power spectrum. To do this, we follow previous literature and implement a foreground transfer function. The transfer function is constructed by injecting simulated mock maps with a known signal into the real data, as discussed in \secref{sec:TF}. We plot the Fourier-space transfer function\)(k\_,k\_)\(in \autoref{fig:TF2D} (3 left panels) as well as\)(k)\((far-right panel). As seen in the auto-\textrm{H\textsc{i}}\ power spectrum in \autoref{fig:HIAutoPk}, after foreground cleaning there is still evidence of residual systematics, which rival the amplitude of the \textrm{H\textsc{i}}\ signal. However, as the\) \< 1\(values in \autoref{fig:TF2D} show, the foreground cleaning is causing signal loss, mostly in the small-\)k\_\(modes. Thus, a fine balance is required between reducing foreground residuals and limiting cosmological signal loss. Encouragingly, we found we were still able to obtain a\)\> 4\(cross-correlation detection with the overlapping WiggleZ galaxies even when the transfer function is not used to correct for signal loss, as we will demonstrate in the following section. However, using the transfer function avoids biasing parameter estimates and improves the cross-correlation detection. We therefore still implement it in the cross-correlation results, which we present next. \subsection{Cross-correlation with WiggleZ galaxies} There are 4031 galaxies in the overlapping 11hr field of the WiggleZ galaxy survey. Following the steps outlined in \secref{sec:Pkest}, we compute an estimate for the cross-power spectrum between the WiggleZ galaxies and the MeerKAT intensity maps, foreground cleaned by removing\)N\_ = 30\(PCA modes. We present this power spectrum in \autoref{fig:CrossPkSteve}. The middle panel shows the signal-to-noise ratio, where we find\)S/N \~ 2\(on large scales. We use an analytical method to estimate the errors (discussed in \secref{sec:Pkest}). At smaller scales, the MeerKAT beam, which is significantly larger than previous intensity mapping surveys\footnote{MeerKAT dishes are \)13.5\,\text{m}\( in diameter and for comparison, the full-width-half-maximum for the central lobe from the beam is \){\sim}\,5\( times larger than the GBT (\)100\,\text{m}\( dish diameter) observations at \)z\,{\sim}\,0.8\(.}, is the main reason for the poor\)S/N\(. The model (black-dotted line) in \autoref{fig:CrossPkSteve} is calculated following \secref{sec:Model}. In this model, we fix all parameters to fiducial quantities except for the degenerate quantity\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(for which we assume scale-independence, hence this quantity will only affect the amplitude of the model. Thus, by fitting the amplitude of the cross-power spectrum, we are sensitive to\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\((from the relation in \autoref{eq:TbarModelEq}). We quote the best-fit value for\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(in the top-right of \autoref{fig:CrossPkSteve}, which was fitted to the data using a least-squares method. We discuss the parameter constraints on\)\_[H[i]{.smallcaps}]{.roman}\(in the following section. We find a good agreement (\)\^2\_dof \~ 1\()\footnote{\)\chi^2_{\rm dof}\,{\equiv}\,\chi^2/{\rm dof}\( where \)\text{dof}\,{=}\,24\( are the degrees of freedom which is the number of \)k\(-bins minus 1 for the single parameter we fit.} between the data and model across all scales (\)`<!-- -->`{=html}0.05 \< k \< 0.28 h \^-1\() in \autoref{fig:CrossPkSteve}. Furthermore, we analyse\) \(, the difference between the data's\)\^2\(evaluated using our cross-correlation model, and one using a null model with zero cross-power. This quantifies the statistical significance of the cross-correlation detection. We achieve a\)`<!-- -->`{=html}7.7\(detection, providing strong evidence for the first verification of a cosmological signal with a multi-dish array performing single-dish intensity mapping. We conducted various null tests on the analysis pipeline. The bottom panel of \autoref{fig:CrossPkSteve} shows the results from shuffling the galaxy maps along the line-of-sight, which should destroy the cross-correlation clustering signal. We re-ran the power spectrum estimation pipeline after each shuffle and found a result consistent with zero. We found similar null results when shuffling the cleaned \textrm{H\textsc{i}}\ intensity maps along the line-of-sight and swapping the WiggleZ maps with random mocks from. We did not apply the transfer function in the bottom panel of \autoref{fig:CrossPkSteve}, since scaling the null results would make no difference. We were still able to obtain a detection for the original un-shuffled results where a transfer function has also not been used. This provides a\)\> 4\(detection, which is strong evidence for correlated clustering since this result relied on no signal reconstruction. For the power spectrum in \autoref{fig:CrossPkSteve} we chose\)N\_ = 30\(as it offers the best balance between goodness-of-fit (\)\^2\_\() and detection significance (\[). \autoref{fig:Nfg_dependence} (top panel) shows how varying\)N\_\(influences\)\^2\_\(, which should ideally be close to unity to represent a good model fit to the data with reasonably sized errors. For each\)N\_\(case we recalculate the transfer function and re-fit the free parameter\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\((values shown by the bottom panel), which avoids the\)\^2\_\(improving simply because the amplitude of the power is decreasing into agreement with a pre-selected fiducial\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(. We also show the cross-correlation detection strength, given by\], on the right-hand (red) axis of \autoref{fig:Nfg_dependence} (top panel). At low\)N\_\(, the\)\^2\_\(appears reasonable but this is due to the larger statistical errors on the cross-power spectrum, which is fairly consistent with zero for these\)N\_\(, as identified by the low detection significance in the\[results. The errors are larger for low\)N\_\(because the residual foregrounds contribute significantly more variance to the maps, even though the residuals themselves are expected to correlate out on average. Increasing\)N\_\(from 10 to 20 does little to improve the detection significance and initially worsens the\)\^2\_\(caused by a decrease in error-bar size. At\)N\_ \~ 30\(enough components have been removed that a clear detection starts to manifest along with an improved agreement between data and model, given by the\)\^2\_ \~ 1\(. Going to much higher\)N\_\(starts to over-clean the maps, reducing\)S/N\(and worsening the detection. To justify this explanation for the deterioration in results for\)N\_ \> 30\(, we analysed the cross-correlation for maps constructed using just the principal components between 30 and 40. This map provided a\)\~ 3.4\(detection, indicating that a lot of signal is present in the modes with\)`<!-- -->`{=html}30 \< N\_ \< 40\(, which explains the deterioration in\)\^2\_\(beyond\)N\_ = 30\(, where these modes are gradually removed. \autoref{fig:Nfg_dependence} highlights the sensitivity of results to the foreground clean, and is further evidence that residual foregrounds and systematics are spread throughout the principal components. The ratio between systematics and signal varies among the various components thus some will be more influential on the cross-correlation than others. This causes a variation in the derived parameter\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(, shown by the bottom panel of \autoref{fig:Nfg_dependence}. We estimate a contribution to the error budget of\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(caused by the variance across the different\)N\_\(, discussed further in the following section. \subsection{Constraints on \boldmath\)\Omega_\textrm{H\textsc{i}}\(} Fitting the amplitude of the \textrm{H\textsc{i}}-galaxy cross-power spectrum provides a constraint on\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(as a function of redshift. From our results in \autoref{fig:CrossPkSteve}, we find\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r = \[0.86 0.10 (stat) 0.12 (sys)\] 10\^-3\(. The systematic error accounts for uncertainty from the map calibration and variance in results from the choice of\)N\_\(. Firstly, uncertainty from the map calibration could cause a bias to the overall amplitude of the power spectrum. We address this by studying the residuals relative to the model in our calibration study. We are able to estimate that gain uncertainties should be at a level of\)\~ 2%\(. Secondly, we account for the variance in results from the different number of PCA modes removed, indicative of residual systematics. We do this by evaluating the standard deviation on all\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(fits (see \autoref{fig:Nfg_dependence}, bottom panel) from each reasonable choice of\)N\_ (10 \< N\_ \< 40)\(, which we find to be\)\~ 0.115\(, equating to a\)\~ 13%\(error on\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(. The combination of these two error components added in quadrature yields the systematic error in our final constraint. The \textrm{H\textsc{i}}\ bias is not yet well understood but is expected to have some scale dependence when entering non-linear scales (high-\)k\(). Furthermore, the cross-correlation coefficient\)r\(, included to account for stochasticity between the two fields, will also have some scale dependence. We therefore examined how the constraint on\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(changed as we varied the scales at which it was measured. By cutting small-\)k\(data points, we change the effective scale of the measurement, calculated by\)k\_ = \_i k_i (S/N)\_i\^2/\_i(S/N)\_i\^2\(, where\)(S/N)\_i\(is the signal-to-noise ratio in each\)k_i\(bin, i.e.\)\_[H[i]{.smallcaps}]{.roman},(k_i)/\_[H[i]{.smallcaps}]{.roman},(k_i)\(. The scale dependence on the measurements of\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(is shown in \autoref{fig:OmHI} (red-star points) for\)N\_ = 30\(. The other coloured data points show previous intensity mapping constraints from GBT cross-correlation with galaxy surveys at\)z \~ 0.8\(. The GBT intensity maps had a significantly smaller beam than MeerKAT thus were able to probe higher-\)k\(. The MeerKAT and GBT measurements are at different redshifts so a direct comparison is not possible. Despite this, there still appears to be a trend with\)k\_\(suggesting a detection of scale dependence in\)b\_[H[i]{.smallcaps}]{.roman} r\(. However, the continuity of the trend is affected by the choice of\)N\_\(, as shown by the different grey lines in \autoref{fig:OmHI}. Furthermore, there is a possibility the scale dependence is influenced by systematics, which are mitigated in the scale cuts. The smallest-\)k\(modes are the most affected by the transfer function, with up to 80\% increase in amplitude (see \autoref{fig:TF2D}). Thus, further investigation is needed to disentangle the scale-dependence of\)b\_[H[i]{.smallcaps}]{.roman} r\(from possible scale-dependent systematics. By making some further assumptions on\)b\_[H[i]{.smallcaps}]{.roman}\(and\)r\(, we can isolate the constraint on\)\_[H[i]{.smallcaps}]{.roman}\(. We assume\)r = 0.9 0.1\(and for the bias we interpolate between the hydrodynamic simulations in and use\)b\_[H[i]{.smallcaps}]{.roman} = 1.13 0.10\(. Fitting the power spectrum data in \autoref{fig:CrossPkSteve} across all scales (\)k\_ \~ 0.13 h \^-1\(), provides a constraint of\)\_[H[i]{.smallcaps}]{.roman} = \[0.83 0.15 (stat) 0.11 (sys)\] 10\^-3\(at the redshift of\)z = 0.43\(, which is reasonably consistent with other results in the literature (see comparison in ). \section{Conclusion}\label{sec:Conclusion} \textrm{H\textsc{i}}\ intensity mapping is a novel method for probing large scale cosmic structure and will be a primary objective for the future SKAO. To achieve this, it is necessary for the multi-dish array to operate in single-dish (auto-correlation) mode, as opposed to a conventional interferometer. In this work we have demonstrated, for the first time, the successful detection of cosmological signal using the MeerKAT multi-dish array in single-dish mode. This represents a major milestone in demonstrating the feasibility of this survey mode for SKAO. We achieved this by cross-correlating\)`<!-- -->`{=html}10.5 \(of MeerKAT pilot survey intensity maps with overlapping optical galaxies from the WiggleZ Dark Energy Survey. A measurement of the cross-power spectrum between these fields provided a\)`<!-- -->`{=html}7.7\(detection of a cross-correlation. We relied on an aggressive filtering process, removing 30 modes in a PCA-based foreground clean, necessary due to the presence of systematic contributions in the pilot survey data. This allowed us to obtain a constraint of\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r = \[0.86 0.10 (stat) 0.12 (sys)\] 10\^-3\(from fitting the amplitude of the cross-power spectrum at an effective scale of\)k\_ \~ 0.13 h \^-1\(. Varying the effective scale of the measurement changed the value for\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(, something noted in previous studies. We also found\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(to have a dependence on the number of foreground modes removed, so we included this variance in the systematic error budget of the constraint. The\)\~ 17.8%\(precision represents a competitive\)\_[H[i]{.smallcaps}]{.roman} b\_[H[i]{.smallcaps}]{.roman} r\(constraint relative to other intensity mapping experiments. Furthermore, with additional assumptions on\)b\_[H[i]{.smallcaps}]{.roman}\(and\)r\(, we provided insight into the cosmic \textrm{H\textsc{i}}\ density\)\_[H[i]{.smallcaps}]{.roman}\(, for which measurements at higher redshifts are vital for understanding the evolution of \textrm{H\textsc{i}}. The MeerKAT telescope will continue to conduct \textrm{H\textsc{i}}\ intensity mapping observations in single-dish mode. With enhanced calibration techniques and more observing time, improved constraints will be possible with less aggressive foreground removal. With this we can attempt a detection of the \textrm{H\textsc{i}}\ in auto-correlation, which is yet to be achieved. Observations have now been conducted in MeerKAT's UHF band (\)`<!-- -->`{=html}0.40 \< z \< 1.45\(), opening the possibility of higher redshift probes and for cross-correlating UHF-band data with the L-band data used in this work, with the aim of mitigating systematics. \section*{Acknowledgements} The authors would like to thank Stefano Camera for useful comments and questions during the development of this project. We would also like to thank Sourabh Paul for his comments on the final manuscript. SC is supported by a UK Research and Innovation Future Leaders Fellowship grant [MR/V026437/1] and also acknowledges funding from the UK Research and Innovation Future Leaders Fellowship grant [MR/S016066/1]. MGS, YL and JW acknowledge support from the South African Radio Astronomy Observatory and National Research Foundation (Grant No. 84156). IPC acknowledges support from the `Departments of Excellence 2018-2022' Grant (L.\ 232/2016) awarded by the Italian Ministry of University and Research (\textsc{mur}). AP is a UK Research and Innovation Future Leaders Fellow [grant MR/S016066/1]. LW is a UK Research and Innovation Future Leaders Fellow [grant MR/V026437/1]. PS is supported by the Science and Technology Facilities Council [grant number ST/P006760/1] through the DISCnet Centre for Doctoral Training. This result is part of a project that has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (Grant agreement No. 948764; PB). PB acknowledges support from STFC Grant ST/T000341/1. JF acknowledges support from the Fundação para a Ciência e a Tecnologia (FCT) through the Investigador FCT Contract No. 2020.02633.CEECIND/CP1631/CT0002 and the research grants UIDB/04434/2020 and UIDP/04434/2020. We acknowledge the use of the Ilifu cloud computing facility, through the Inter-University Institute for Data Intensive Astronomy (IDIA). The MeerKAT telescope is operated by the South African Radio Astronomy Observatory, which is a facility of the National Research Foundation, an agency of the Department of Science and Innovation. For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising. \section*{Data Availability} The data underlying this article will be shared upon acceptance of the manuscript. \bibliographystyle{mnras} \section{Introduction} Probing the large scale structure of the Universe is a crucial step towards precision cosmology as we try to constrain the nature of dark energy, non-Gaussian fluctuations in the Universe's primordial density field, and test general relativity. Typically, this is done using galaxy surveys with spectroscopic or photometric redshifts in the optical or near-infrared. At radio wavelengths, we use the redshifted neutral hydrogen (\textrm{H\textsc{i}}) hyperfine transition line, with a rest-frame wavelength of 21cm, to measure redshift. Given the ubiquitous nature of \textrm{H\textsc{i}}\ in the Universe, we can use it to trace the distribution of dark matter at low and high redshifts. The faintness of the \textrm{H\textsc{i}}\ emission line makes it challenging to resolve individual galaxies at higher redshifts over large volumes. However, for cosmology we are interested in the bulk fluctuations on large (Mpc) scales, so we can use the \textrm{H\textsc{i}}\ intensity mapping technique. This technique relaxes the requirement of galaxy detection by integrating all 21cm emission within relatively large spatial voxels. This delivers high survey speeds over large volumes, providing a novel solution to the current challenges of observational cosmology. One of the main challenges in detecting the \textrm{H\textsc{i}}\ intensity mapping signal is the presence of foregrounds that are orders of magnitude brighter. Removing these requires precise instrumental calibration. Cross-correlating with galaxy surveys helps to mitigate residual systematics from foregrounds, Radio Frequency Interference (RFI) and thermal noise. Moreover, it can improve constraints on cosmological parameters and provide insight into the \textrm{H\textsc{i}}\ astrophysics of the correlated galaxies. Currently, the \textrm{H\textsc{i}}\ intensity mapping signal has only been detected in cross-correlation with galaxy surveys. Both MeerKAT and the future SKA Observatory (SKAO) have been put forward as state-of-the-art intensity mapping instruments capable of complementing and extending cosmological measurements at other wavelengths. Using the single-dish data from each element of the array, we can access the large cosmological scales inaccessible by the interferometer due to its lack of very short baselines. In we presented the first calibrated maps of a pilot MeerKAT survey, which utilised this single-dish method on the multi-dish interferometer array. In this article, we use this data to measure the cross-correlation power spectrum between the \textrm{H\textsc{i}}\ signal and overlapping WiggleZ data. With only\)`<!-- -->`{=html}10.5 \(of data for each of the 64 dishes over an effective survey area of\)\~ 200 \^2\(, this detection shows the power of this approach and paves the way towards probing large cosmological scales with much larger surveys on MeerKAT and SKAO. The paper is structured as follows; in \secref{sec:ObsData} we introduce the data products used in this study. \secref{sec:FGcleaning} introduces our approach to foreground cleaning in the MeerKAT intensity maps. The formalism adopted for the power spectrum estimation and modelling is discussed in \secref{sec:Pkest}. We present our main results in \secref{sec:Results} and finally conclude in \secref{sec:Conclusion}. \section{Observational data}\label{sec:ObsData} The MeerKAT \textrm{H\textsc{i}}\ intensity mapping data was collected using the L-band with\)`<!-- -->`{=html}10.5\(hours of observation per dish over six nights between February and July 2019. The L-band has 4096 frequency channels at\)`<!-- -->`{=html}856-1712 \(, but in this work, we only use 199 channels at\)`<!-- -->`{=html}973.2-1014.6 \((\)`<!-- -->`{=html}0.400 \< z \< 0.459\(). Of these channels, a further 32 are removed due to their dominant contributions to the eigenmodes of the principal component analysis (see later discussion in \secref{sec:FGcleaning}) for those particular channels. This aggressive strategy can be seen as a final RFI-flagging stage\footnote{Full details of the initial RFI-flagging are given in.} and was necessary in order for us to obtain a cross-correlation detection. The survey area spans around\)`<!-- -->`{=html}200 deg\^2\(and targeted the WiggleZ 11hr field, covering\)`<!-- -->`{=html}153\^ \< \< 172\^\(and\)-1\^ \< \< 8\^\(. In this analysis, only the auto-correlation of MeerKAT interferometer visibility data, i.e. the single dish mode observation, is used. The time-ordered data for each dish,\)d\(, are projected into the map space via the map-making process, \begin{align} \hat{m} = \left(A^\text{T} N^{-1} A\right)^{-1} A^\text{T} N^{-1} d \end{align} in which\)A\(is the pointing matrix mapping the time-ordered data to the map coordinates and\)N\(is the noise covariance matrix between time stamps. As described in, the time-ordered data (TOD) is calibrated against a noise diode which is fired every\)`<!-- -->`{=html}20\(seconds. This removes long-term noise correlation, so-called\)`<!-- -->`{=html}1/\(noise, due to receiver chain gain variations on time scales longer than 20\,s. On shorter time scales,\)`<!-- -->`{=html}1/\(noise is negligible compared to thermal noise fluctuations. We also subtract the average signal every 220\,s in the TOD which suppresses long-term sloping gain changes. This should reduce the overall variance of the signal but can potentially have the adverse effect of removing \textrm{H\textsc{i}}\ signal over large angular scales. In this work we assume any signal loss from this process is sub-dominant relative to the signal loss from foreground cleaning and thus do not attempt any reconstruction. The noise covariance matrix\)N\(is assumed to be diagonal with constant variance during the observation, but we allow the variance to differ between dishes. The noise covariance is also projected to the map space via, \begin{align} \hat{n} = \left(A^\text{T} N^{-1} A\right)^{-1}, \end{align} where\]is the pixel noise variance. The inverse of the pixel noise variance,\)w\_[H[i]{.smallcaps}]{.roman} = 1/\(, is used as the inverse-variance weight in the analysis\footnote{We adopt the \textrm{H\textsc{i}}\ in the subscript of \)w_\textrm{H\textsc{i}}\( to maintain a consistent notation with later formalism where they require distinguishing from the galaxy weights.}. We use the flat-sky approximation and grid the map into square pixels with a width of\)`<!-- -->`{=html}0.25\^\(. To create the final maps, we average over all individual dish maps (full details of the single-dish calibration are provided in ). The observed intensity maps are shown in \autoref{fig:MKmaps}. The top panel shows the frequency-averaged intensity map before foreground cleaning. The bottom panel shows the residual map after foreground cleaning, a process we outline in the following section. \section{Foreground Cleaning}\label{sec:FGcleaning} Here we discuss the foreground cleaning performed on the MeerKAT intensity maps. We provide detailed descriptions on each stage in the following sub-sections but begin with a summary of the foreground cleaning method we adopt. The foreground cleaning is performed using a blind Principal Component Analysis (PCA) method, which relies on the foregrounds being the dominant signal and correlated in frequency. Thus, by removing the first\)N\_\(principal component modes in frequency from each pixel, the majority of their contribution is suppressed (see \secref{sec:PCA} for further discussion). Before cleaning, the map is resmoothed using a Gaussian window function with kernel size\)`<!-- -->`{=html}1.2\(times the largest beam size within the frequency range (see \secref{sec:Resmooth} for details). Foreground cleaning is imperfect, and the cleaned maps contain residual foreground structure which can be seen by comparing both panels in \autoref{fig:MKmaps}. However, the amplitude of the map has decreased by several orders of magnitude, thus the foreground residuals should dominate less over the \textrm{H\textsc{i}}\ fluctuations. Furthermore, some \textrm{H\textsc{i}}\ signal will be removed, typically on larger scales where modes are most degenerate with the spectrally smooth foregrounds. We aim to reconstruct this lost signal with a foreground transfer function, which we discuss in \secref{sec:TF}. \subsection{Reconvolution of maps}\label{sec:Resmooth} It is understood that a frequency-dependent beam size can cause the foregrounds to leak into a greater number of spectral modes, requiring more aggressive cleaning. A way to potentially mitigate this issue is to convolve all maps to a common resolution before performing the foreground clean, as done in previous experiments. However, recent tests on simulations suggest that a simple Gaussian resmoothing of the data to a common resolution does not improve blind foreground removal techniques, even if the true beam is a perfect Gaussian. For real data though, it is beneficial to resmooth to homogenise some of the systematic contributions from e.g. residual RFI or polarisation leakage. For this reason we perform a weighted resmoothing on the MeerKAT \textrm{H\textsc{i}}\ intensity maps prior to foreground cleaning. An intensity map\)T\^\(which has a frequency dependent beam (denoted by the\)\^\(index) with a full-width-half-maximum\)\_()\(in degrees, and an angular separation between pixels given by\[, is convolved with the following kernel: \begin{equation}\label{eq:ResmoothKernel} K(\Delta \theta,\nu)=\exp \left[-\frac{\Delta \theta^{2}}{2[\gamma \sigma_{\max }^{2}-\sigma^{2}(\nu)]}\right]\,, \end{equation} where\)()=\_() /(2 )\(,\)\_\(is the maximum\)()\(value and\]is a scaling factor which governs how much the final effective resolution is decreased by. In previous Green Bank Telescope (GBT) studies, a choice of\) = 1.4\(was used. In this work, due to the already large MeerKAT beam, we use a smaller value of\) = 1.2\(. We experimented with this choice, finding\) = 1\(produced a noisier power spectrum with worse model agreemen. The higher choice of\) = 1.4\(delivered a similarly good model agreement compared to\) = 1.2\(, but had a slightly lower cross-correlation detection significance due to the increased damping at high-\)k\(. Not performing any reconvolution still delivered a clear detection but resulted in a particularly noisy power spectrum at small-\)k, indicating the presence of residual foreground and systematics which are mitigated by this resmoothing procedure. The kernel in is normalised such that the sum over all pixels is equal to unity, then the weighted convolution used to resmooth the maps is given by \[\delta T(\theta,\nu) = \frac{\left[\delta T^\prime(\theta,\nu)\,w_\textrm{H\textsc{i}}^\prime(\theta,\nu)\right] * K(\Delta\theta,\nu)}{w_\textrm{H\textsc{i}}^\prime(\theta,\nu) * K(\Delta\theta,\nu)}\,,\] where \(w_\textrm{H\textsc{i}}^\prime(\theta)\) is the inverse variance weight. The \(*\) denotes a convolution performed separately in each frequency channel e.g. \(w_\textrm{H\textsc{i}}^\prime(\theta) * K(\Delta\theta) = \Sigma_i w_\textrm{H\textsc{i}}^\prime(\theta_i)K(\theta-\theta_i)\). To ensure the weight field still represents the inverse variance of the new resmoothed field, the weight \(w_\textrm{H\textsc{i}}^\prime\) is convolved according to \[w_\textrm{H\textsc{i}}(\theta,\nu) = \frac{\left[w_\textrm{H\textsc{i}}^\prime(\theta,\nu) * K(\Delta\theta,\nu)\right]^2}{w_\textrm{H\textsc{i}}^\prime(\theta,\nu) * K^2(\Delta\theta,\nu)}\.\]
## PCA foreground cleaning {#sec:PCA}
In this work, a PCA-based blind foreground subtraction method is used. The observed intensity maps can be represented by a matrix \(\mathbf{X}_\text{obs}\) with dimensions \(N_\nu\,{\times}\,N_\theta\) where \(N_\nu\) is the number of frequency channels along the line-of sight and \(N_\theta\) is the number of pixels. The assumption behind blind-foreground cleaning is that the data can be represented by the linear system \(\mathbf{X}_\text{obs}\,{=}\,\mathbf{\hat{A}}\mathbf{S}\,{+}\,\mathbf{R}\), where \(\mathbf{\hat{A}}\) represents the mixing matrix and \(\mathbf{S}\) are the \(N_\text{fg}\) separable source maps identified by projecting the mixing matrix along the data \(\mathbf{S}\,{=}\,\mathbf{\hat{A}}^\text{T}\mathbf{X}_\text{obs}\). In PCA, the mixing matrix is extracted from the eigen-decomposition of the frequency covariance matrix of the mean-centred data, defined by \(\mathbf{C}\,{=}\,(\mathbf{w}\mathbf{X}_\text{obs})^\text{T}(\mathbf{w}\mathbf{X}_\text{obs})/(N_\theta\,{-}\,1)\), where \(\mathbf{w}\) are the inverse variance weights recast into \(N_\nu\,{\times}\,N_\theta\) matrices. The eigen-decomposition is then given as \(\mathbf{C}\mathbf{V}=\mathbf{V}\mathbf{\Lambda}\), where \(\mathbf{\Lambda}\) is the diagonal matrix of eigenvalues ordered by descending magnitude, and \(\mathbf{V}\) are the eigenvectors, the first \(N_\text{fg}\) of which supplies the set of functions used to construct the mixing matrix. We assume the subtraction of \(\mathbf{\hat{A}}\mathbf{S}\) in the linear system will remove dominant foregrounds, leaving behind in the residuals \(\mathbf{R}\) any [H[i]{.smallcaps}]{.roman} signal not removed in the subtraction.
## Transfer function construction {#sec:TF}
We compensate the signal loss due to the foreground cleaning with a foreground transfer function. Following previous literature, the transfer function can be constructed by injecting mock intensity mapping data \(\mathbf{M}_\textrm{H\textsc{i}}\) into the true observed data \(\mathbf{X}_\text{obs}\), which includes foregrounds and observational systematics. By running a PCA clean on this combination, we can measure (and compensate for) the signal loss in the cleaned mock data; \[\label{eq:MockforTF} \mathbf{M}_\text{c} = [\mathbf{M}_\textrm{H\textsc{i}} + \mathbf{X}_\text{obs}]_\text{PCA}-[\mathbf{X}_\text{obs}]_\text{PCA}\.\] The \([\ ]_\text{PCA}\) notation represents performing the PCA clean (outlined in ) on the quantities inside the brackets, treating them as a single combination. For example, the mixing matrix is not determined separately for both mock and data in \([\mathbf{M}_\textrm{H\textsc{i}} + \mathbf{X}_\text{obs}]_\text{PCA}\), but determined for the combination of \(\mathbf{M}_\textrm{H\textsc{i}} + \mathbf{X}_\text{obs}\). We also subtract the PCA clean of the data, \([\mathbf{X}_\text{obs}]_\text{PCA}\), since this only adds uncorrelated variance, thus subtracting it makes convergence to a smooth transfer function more efficient, requiring fewer mock iterations. After calculating \(\mathbf{M}_\text{c}\), we measure the cross-power spectrum with a corresponding mock galaxy map \(\mathbf{M}_\text{g}\), then divide this by a foreground-free equivalent to estimate the signal loss at each mode; \[\label{eq:TF} \mathcal{T}(k) = \left\langle \frac{\mathcal{P}(\mathbf{M}_\text{c}\, ,\, \mathbf{M}_\text{g})}{\mathcal{P}(\mathbf{M}_\textrm{H\textsc{i}}\, ,\, \mathbf{M}_\text{g})} \right\rangle \,.\] \(\mathcal{P}()\) denotes an operator which measures the cross-power spectrum, then spherically averages modes into the same \(k\)-bins as the data. The angled brackets represent an ensemble average over a large number of mocks (we use 100 in this work). The [H[i]{.smallcaps}]{.roman} mocks are generated with the lognormal method, sampled from a model [H[i]{.smallcaps}]{.roman} power spectrum (the same as that used in the power spectrum fitting, see ) with a Gaussian smoothing applied perpendicular to the line-of-sight to approximately emulate the MeerKAT beam. Similarly, for the galaxy mocks, we generate a lognormal density field with the same random seed as the [H[i]{.smallcaps}]{.roman}, then Poisson sample galaxies onto the field with the same number count as the WiggleZ catalogue. These steps ensure the mock fields emulate the amplitudes of real maps as realistically as possible. There is evidence suggesting the transfer function is not overly sensitive to the choice of fiducial cosmology, but further investigation into how much it can be relied on for precision cosmology is required. Any power spectrum measurement we make on the data is divided by \(\mathcal{T}(k)\) to correct for the signal loss (unless clearly stated in demonstrative figures). For the [H[i]{.smallcaps}]{.roman} auto-power spectrum used in the error estimation (discussed later in ), we also multiply through by \(1/\mathcal{T}(k)\). Previous studies have opted to use \(1/\mathcal{T}(k)^2\) as a correction instead, motivated by the assumption that in auto-correlation, signal loss occurs in both maps, so there should be twice the reconstruction of power needed. However, from simulation tests, we found this over-corrected the signal loss. Furthermore, our analytical error estimation on the cross-power spectrum (see ), which uses the auto-[H[i]{.smallcaps}]{.roman} power spectrum, is found to be in good agreement with other approaches of error estimation using the WiggleZ randoms and jackknife tests. \(\chi^2_\text{dof}\) analysis also suggests our errors are not over-estimated in any case. This changes if we opt for the \(1/\mathcal{T}(k)^2\) correction where it becomes clear that the errors have been over-estimated, suggesting that the signal loss in the auto-[H[i]{.smallcaps}]{.roman} power spectrum has been over-corrected. We defer further investigation into signal loss in the [H[i]{.smallcaps}]{.roman} auto-correlation to future work.
# Power spectrum estimation & modelling {#sec:Pkest}
The power spectrum estimation process we adopt is based on the optimal weighting method outlined in (see for applications to [H[i]{.smallcaps}]{.roman} intensity mapping cross-correlations). We define the two Fourier transformed fields of the [H[i]{.smallcaps}]{.roman} temperature fluctuation maps \(\delta T_\textrm{H\textsc{i}}\) and the galaxy count field \(n_\text{g}\) as \[\tilde{F}_\textrm{H\textsc{i}}(\mathbf{k})=\sum_{\mathbf{x}} \delta T_\textrm{H\textsc{i}}(\mathbf{x}) w_\textrm{H\textsc{i}}(\mathbf{x}) \exp (i \mathbf{k}{\cdot}\mathbf{x})\] \[\tilde{F}_\text{g}(\mathbf{k})=\sum_{\mathbf{x}} n_\text{g}(\mathbf{x}) w_\text{g}(\mathbf{x}) \exp (i \mathbf{k}{\cdot}\mathbf{x})-N_\text{g}\tilde{W}_\text{g}(\mathbf{k})\,,\] where \(N_\text{g}\,{=}\,\sum n_\text{g}\) is the total number of galaxies in the optical map and \(\tilde{W}_\text{g}\) is the weighted Fourier transform of the selection function (\(W_\text{g}\)) which is normalised such that \(\sum_\mathbf{x}W_\text{g}(\mathbf{x})\,{=}\,1\); \[\tilde{W}_\text{g}(\mathbf{k})=\sum_{\mathbf{x}} W_\text{g}(\mathbf{x}) w_\text{g}(\mathbf{x}) \exp (i \mathbf{k}{\cdot}\mathbf{x})\.\] The selection function \(W_\text{g}\) accounts for incompleteness in the WiggleZ survey and is constructed by stacking the random catalogues generated in, reproducing the varying target and redshift completeness. The map weights in the above equations are the inverse variance map for the [H[i]{.smallcaps}]{.roman} field[^3], \(w_\textrm{H\textsc{i}}(\mathbf{x})\,{=}\,1/\hat{n}(\mathbf{x})\) (defined in ) and for the galaxies we use the optimal weighting as per \[\label{eq:FKPweight} w_\text{g}(\mathbf{x}) = 1\bigg/\left(1+\frac{W_\text{g}(\mathbf{x})N_\text{g}P_0}{V_\text{cell}}\right)\,,\] where \(P_0\) is a chosen value of the power spectrum at the scale where an optimal measurement is sought. We use \(P_0\,{=}\,5000\,h^{-3}\text{Mpc}^3\), approximately corresponding to \(k\,{=}\,0.1\,h\,\text{Mpc}^{-1}\). \(V_\text{cell}\) is the volume of a grid cell in \(h^{-3}\text{Mpc}^3\). The maps are enclosed onto a cuboid in comoving space, using a fiducial Planck18 cosmology with sides of length \((l_\text{x},l_\text{y},l_\text{z})\,{\sim}\,(812,264,176)\,\text{Mpc}/h\), divided into \((n_\text{x},n_\text{y},n_\text{z})\,{=}\,(151,49,250)\) equal sized cells. Assuming a flat-sky approximation, and using the map pixel's angular coordinates and redshifts, we grid the data using a nearest grid point assignment[^4]. To taper the edges along the line-of-sight, we apply a Blackman window along the z-axis of the galaxy and [H[i]{.smallcaps}]{.roman} intensity maps, as well as their weights and selection functions. The cross-correlation between the [H[i]{.smallcaps}]{.roman} intensity maps and the galaxies is then estimated with \[\small \hat{P}_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}) = \frac{V_\text{cell}}{\sum\limits_\mathbf{x} w_\textrm{H\textsc{i}}(\mathbf{x})w_\text{g}(\mathbf{x})W_\text{g}(\mathbf{x})}\operatorname{Re}\left\{\tilde{F}_\textrm{H\textsc{i}}(\mathbf{k}){\cdot} \tilde{F}^{*}_\text{g}(\mathbf{k})\right\}\frac{1}{N_\text{g}}\.\] Similarly, the [H[i]{.smallcaps}]{.roman} and galaxy auto-power spectra, which are needed for the error estimation (see below), are given by \[\label{eq:HIautoPk} \hat{P}_\textrm{H\textsc{i}}(\mathbf{k}) = \frac{V_\text{cell}}{\sum_\mathbf{x} w_\textrm{H\textsc{i}}^2(\mathbf{x})}\lvert\tilde{F}_\textrm{H\textsc{i}}(\mathbf{k})\rvert^2\,,\] \[\hat{P}_\text{g}(k) = \frac{V_\text{cell}}{\sum_\mathbf{x} w^2_\text{g}(\mathbf{x})W^2_\text{g}(\mathbf{x})} \left[|\tilde{F}_\text{g}(\mathbf{k})|^2-S_\text{g}\right]\frac{1}{N^2_\text{g}}\,,\] where \(S_\text{g}\) accounts for the shot noise in the galaxy survey, given as \[S_\text{g} = N_\text{g}\sum_\mathbf{x} w_\text{g}^2(\mathbf{x})W_\text{g}(\mathbf{x})\.\] These power spectra are all spherically averaged into bandpowers \(\lvert\mathbf{k}\rvert\,{\equiv}\,k\) to provide the final 1D power spectra results (the cross-and [H[i]{.smallcaps}]{.roman}-auto power spectra are also corrected with the \(1/\mathcal{T}(k)\) as discussed in ). For noise-dominated [H[i]{.smallcaps}]{.roman} intensity maps, the errors for the cross-power can be estimated analytically from \[\label{eq:Pkerr} \small \hat{\sigma}_{\textrm{H\textsc{i}},\text{g}}(k)=\frac{1}{\sqrt{2 N_\text{m}(k)}} \sqrt{\hat{P}^2_{\textrm{H\textsc{i}},\text{g}}(k)+\hat{P}_\textrm{H\textsc{i}}(k)\left(\hat{P}_\text{g}(k)+\frac{1}{\bar{n}_\text{g}}\right)}\,,\] where \(N_\text{m}\) is the number of modes in each \(k\)-bin and \(\bar{n}_\text{g}\,{=}\,N_\text{g}/(l_\text{x}\,{\times}\,l_\text{y}\,{\times}\,l_\text{z})\) is the number density of galaxies. The \(1/\sqrt{2}\) factor in appears because this is the error on a cross-correlation of two fields, so the number of independent pairs available to measure the variance on the mean doubles. We compared these analytical error estimations to ones calculated from cross-correlating the MeerKAT data with the random WiggleZ catalogues used to derive the selection function, finding very good agreement across all scales. Furthermore, we also found good agreement with an internal error estimation which used a jackknife approach.
## Theoretical modelling {#sec:Model}
We fit a model to the cross-power spectrum which is given by \[\begin{gathered}
\label{eq:Pkmodel} P_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}) = \overline{T}_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} b_\text{g} r (1+f\mu^2)^2 \,P_\text{m}(k) \\ \times\, \exp\left[\frac{-(1-\mu^2)k^2 R_\text{beam}^2}{2}\right]\,, \end{gathered}\] where \(\overline{T}_\textrm{H\textsc{i}}\) is the mean [H[i]{.smallcaps}]{.roman} temperature of the field in mK, \(b_\textrm{H\textsc{i}}\) and \(b_\text{g}\) are the [H[i]{.smallcaps}]{.roman} and galaxy biases and \(r\) is the cross-correlation coefficient. We account for linear redshift-space distortions (RSD) with the \((1+f\mu^2)^2\) factor, where \(f\) is the growth rate of structure and \(\mu\) is the cosine of the angle from the line-of-sight. \(P_\text{m}\) is the matter power spectrum produced using `CAMB` with a Planck18 cosmology. The exponential factor approximates the smoothing of perpendicular modes due to the MeerKAT beam, where \(R_\text{beam}\) is the standard deviation of the Gaussian beam profile in comoving units, taking into account the reconvolution as explained in , which gives \(R_\text{beam}\,{=}\,13.3\,\text{Mpc}\,h^{-1}\). To fully account for RSD, the model in should be (not including the beam damping for brevity) \[P_{\textrm{H\textsc{i}} ,\text{g}}(\mathbf{k}) = \overline{T}_\textrm{H\textsc{i}}\left[r b_\textrm{H\textsc{i}} b_\text{g} + b_\textrm{H\textsc{i}} f \mu^2 + b_\text{g} f \mu^2 + f^2 \mu^4\right] P_\text{m}(k)\,,\] which accounts for the biases appearing in cross-terms from the expansion of the two fields in redshift space \(\delta^\text{s}_\textrm{H\textsc{i}}(\mathbf{k})\,{=}\,b_\textrm{H\textsc{i}} \delta_\text{m}(\mathbf{k})\,{+}\,f\mu\,\theta(\mathbf{k})\) and \(\delta^\text{s}_\text{g}(\mathbf{k})\,{=}\,b_\text{g} \delta_\text{m}(\mathbf{k})\,{+}\,f\mu\,\theta(\mathbf{k})\), where \(\theta(\mathbf{k})\) is the velocity divergence field[^5]. However, we are only attempting a fit to the spherically-averaged power spectrum monopole, which is uniformly averaged across \(\mu\). This would make \(b_\textrm{H\textsc{i}}\) perfectly degenerate with \(\overline{T}_\textrm{H\textsc{i}} r\). To break this degeneracy we would need to introduce an anisotropic sensitivity on \(\mu\) in our analysis, achieved by modelling the quadrupole. This would require a higher \(S/N\) than we have available from the MeerKAT pilot survey data. This is why we follow previous literature and probe the degenerate quantity \(\overline{T}_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\), but include a matter-only RSD, to avoid biasing the amplitude of the power spectrum by the \((1+ f\mu^2)\,{\sim}\,1.7\) Kaiser term. The model in is discretised onto the same 3D grid of modes as the data and then convolved with the survey window functions \[\begin{gathered}
\label{eq:ConvolvedModel} P_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}) \rightarrow P_{\textrm{H\textsc{i}},\text{g}} * W_\textrm{H\textsc{i}} W_\text{g} = \\ \frac{\sum_{i} P_{\textrm{H\textsc{i}},\text{g}}(\mathbf{k}_{i}^{\prime}) \operatorname{Re}\left\{\tilde{W}_\text{g}(\mathbf{k}-\mathbf{k}_{i}^{\prime}) \tilde{W}_\textrm{H\textsc{i}}(\mathbf{k}-\mathbf{k}_{i}^{\prime})^{*}\right\}}{\sum w_\text{g}(\mathbf{x}) w_\textrm{H\textsc{i}}(\mathbf{x}) W_\text{g}(\mathbf{x}) W_\textrm{H\textsc{i}}(\mathbf{x})}\. \end{gathered}\] In lieu of a precisely constructed survey selection function \(W_\textrm{H\textsc{i}}\) for the [H[i]{.smallcaps}]{.roman} intensity maps, we use a simple binary window function that is 1 wherever a pixel is filled and 0 otherwise. The convolved model in is spherically averaged into the same \(k\)-bins as the data. The mean [H[i]{.smallcaps}]{.roman} temperature \(\overline{T}_\textrm{H\textsc{i}}\) can be recast to the [H[i]{.smallcaps}]{.roman} density fraction (\(\Omega_\textrm{H\textsc{i}}\)) using \[\label{eq:TbarModelEq} \overline{T}_\textrm{H\textsc{i}}(z) = 180\,\Omega_{\textrm{H\textsc{i}}}(z)\,h\,\frac{(1+z)^2}{\sqrt{\Omega_\text{m}(1+z)^3 + \Omega_\Lambda}} \, {\text{mK}} \,,\] where \(\Omega_\text{m}\) and \(\Omega_\Lambda\) are the density fractions for matter and the cosmological constant, respectively. Thus fitting the amplitude of the cross-power spectrum allows us to constrain \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\). When fitting \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) to the power spectrum data using the model in , we fix the galaxy bias (\(b_\text{g}\)) and growth rate (\(f\)), since they are well constrained from other experiments relative to the other parameters. We assume \(f\,{=}\,0.737\) (based on \(f\,{\sim}\,\Omega_\text{m}(z)^\gamma\), where \(\gamma\,{=}\,0.545\) ) and \(b_\text{g}\,{=}\,0.911\) at the central redshift of our data (\(z_\text{eff}\,{=}\,0.43\)).
# Results {#sec:Results}
## [H[i]{.smallcaps}]{.roman} auto power spectrum
The auto power spectrum of the cleaned [H[i]{.smallcaps}]{.roman} intensity maps gives some indication of how much foregrounds have been suppressed. In we show the 2D auto-[H[i]{.smallcaps}]{.roman} power spectrum decomposed into anisotropic \(k\)-bins perpendicular (\(k_\perp\,{=}\,\sqrt{k^2_\text{x} + k^2_\text{y}}\)) and parallel (\(k_\parallel\,{\equiv}\,k_\text{z}\)) to the line-of-sight. The far-left panel shows the original data before foreground removal, demonstrating the dominance of the foregrounds and their concentration on the largest scales, particularly at small-\(k_\parallel\). Removing just a few principal components reduces the amplitude of the auto power spectrum by several orders of magnitude, as shown in the second panel. In simulation tests, removing \(N_\text{fg}\,{\sim}\,4\) PCA modes is sufficient to remove the majority of the foregrounds, and this should also be the case for a perfectly designed and calibrated experiment. The resulting cleaned maps in this idealised case would contain [H[i]{.smallcaps}]{.roman} signal and Gaussian thermal noise, similar to the model in the fifth panel of . The thermal noise in the ideal model is large due to the survey's low observing time. This yields an auto power spectrum amplitude of \(P_\text{noise}\,{\sim}\,85\,\text{mK}^2h^{-3}\text{Mpc}^3\), which is expected to dominate over the [H[i]{.smallcaps}]{.roman} signal[^6] on most scales, as seen in the model. For this pilot-survey data, there will be instrumental calibration imperfections, residual RFI and other systematics (see discussions in ). The instrumental response modulates the foregrounds, resulting in additional spectral structure that requires more PCA modes to be removed. Uncleaned modes containing residual foregrounds, RFI and other systematics will positively bias the auto-power spectrum[^7]. This is why for the \(N_\text{fg}\,{=}\,4\) case, we do not reach the level of the [H[i]{.smallcaps}]{.roman} signal plus noise. As \(N_\text{fg}\) increases in , the amplitude of the power drops, and this occurs more severely for large modes, particularly at small-\(k_\parallel\). While it appears that we are reaching below the idealised [H[i]{.smallcaps}]{.roman} + noise model at high-\(k_\perp\) and low-\(k_\parallel\) in the \(N_\text{fg}\,{\ge}\,20\) cases, we would need to account for the effects of the map reconvolution and the foreground clean, both of which would damp the [H[i]{.smallcaps}]{.roman} + noise model further. In the far-right panel we show a simulated [H[i]{.smallcaps}]{.roman} mock with the same noise level as the idealised model, but include some observational effects. To emulate the beam and reconvolution, we smooth the simulation perpendicular to the line-of-sight. We also emulate signal loss from the foreground clean by projecting out modes based on the same PCA mixing matrix functions derived for the \(N_\text{fg}\,{=}\,30\) data case. Comparison between the \(N_\text{fg}\,{=}\,30\) and far-right panel suggests we have not reached the estimated auto-power spectrum level, indicating that residual RFI, foregrounds, and other systematics are present in the data. Since the additive bias from non-cosmological residuals is unknown, it is difficult to compare auto-correlated data and model and conclude that a cosmological detection has been achieved. Cross-correlating with galaxy surveys avoids these additive biases, serving the motivation for this work. We leave a detailed study into the auto-[H[i]{.smallcaps}]{.roman} power for future work, where we will explore cross-correlating different sets of dishes or observational time-blocks (a method adopted in GBT experiments ).
## Post-foreground clean signal reconstruction
Along with residuals from an imperfect cleaning process, foreground removal will also cause cosmological signal loss in the intensity maps, which requires reconstruction to avoid negatively biasing the measured power spectrum. To do this, we follow previous literature and implement a foreground transfer function. The transfer function is constructed by injecting simulated mock maps with a known signal into the real data, as discussed in . We plot the Fourier-space transfer function \(\mathcal{T}(k_\perp,k_\parallel)\) in (3 left panels) as well as \(\mathcal{T}(k)\) (far-right panel). As seen in the auto-[H[i]{.smallcaps}]{.roman} power spectrum in , after foreground cleaning there is still evidence of residual systematics, which rival the amplitude of the [H[i]{.smallcaps}]{.roman} signal. However, as the \(\mathcal{T}\,{<}\,1\) values in show, the foreground cleaning is causing signal loss, mostly in the small-\(k_\parallel\) modes. Thus, a fine balance is required between reducing foreground residuals and limiting cosmological signal loss. Encouragingly, we found we were still able to obtain a \({>}\,4\sigma\) cross-correlation detection with the overlapping WiggleZ galaxies even when the transfer function is not used to correct for signal loss, as we will demonstrate in the following section. However, using the transfer function avoids biasing parameter estimates and improves the cross-correlation detection. We therefore still implement it in the cross-correlation results, which we present next.
## Cross-correlation with WiggleZ galaxies
There are 4031 galaxies in the overlapping 11hr field of the WiggleZ galaxy survey. Following the steps outlined in , we compute an estimate for the cross-power spectrum between the WiggleZ galaxies and the MeerKAT intensity maps, foreground cleaned by removing \(N_\text{fg}\,{=}\,30\) PCA modes. We present this power spectrum in . The middle panel shows the signal-to-noise ratio, where we find \(S/N\,{\sim}\,2\) on large scales. We use an analytical method to estimate the errors (discussed in ). At smaller scales, the MeerKAT beam, which is significantly larger than previous intensity mapping surveys[^8], is the main reason for the poor \(S/N\). The model (black-dotted line) in is calculated following . In this model, we fix all parameters to fiducial quantities except for the degenerate quantity \(\overline{T}_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) for which we assume scale-independence, hence this quantity will only affect the amplitude of the model. Thus, by fitting the amplitude of the cross-power spectrum, we are sensitive to \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) (from the relation in ). We quote the best-fit value for \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) in the top-right of , which was fitted to the data using a least-squares method. We discuss the parameter constraints on \(\Omega_\textrm{H\textsc{i}}\) in the following section. We find a good agreement (\(\chi^2_{\rm dof}\,{\sim}\,1\))[^9] between the data and model across all scales (\(0.05\,{<}\,k\,{<}\,0.28\,h\,\text{Mpc}^{-1}\)) in . Furthermore, we analyse \(\sqrt{\Delta \chi^2}\,{\equiv}\,\sqrt{\chi^2-\chi^2_{\rm null}}\), the difference between the data's \(\chi^2\) evaluated using our cross-correlation model, and one using a null model with zero cross-power. This quantifies the statistical significance of the cross-correlation detection. We achieve a \(7.7\sigma\) detection, providing strong evidence for the first verification of a cosmological signal with a multi-dish array performing single-dish intensity mapping. We conducted various null tests on the analysis pipeline. The bottom panel of shows the results from shuffling the galaxy maps along the line-of-sight, which should destroy the cross-correlation clustering signal. We re-ran the power spectrum estimation pipeline after each shuffle and found a result consistent with zero. We found similar null results when shuffling the cleaned [H[i]{.smallcaps}]{.roman} intensity maps along the line-of-sight and swapping the WiggleZ maps with random mocks from. We did not apply the transfer function in the bottom panel of , since scaling the null results would make no difference. We were still able to obtain a detection for the original un-shuffled results where a transfer function has also not been used. This provides a \({>}\,4\sigma\) detection, which is strong evidence for correlated clustering since this result relied on no signal reconstruction. For the power spectrum in we chose \(N_\text{fg}\,{=}\,30\) as it offers the best balance between goodness-of-fit (\(\chi^2_\text{dof}\)) and detection significance (\(\sqrt{\Delta\chi^2}\)). (top panel) shows how varying \(N_\text{fg}\) influences \(\chi^2_\text{dof}\), which should ideally be close to unity to represent a good model fit to the data with reasonably sized errors. For each \(N_\text{fg}\) case we recalculate the transfer function and re-fit the free parameter \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) (values shown by the bottom panel), which avoids the \(\chi^2_\text{dof}\) improving simply because the amplitude of the power is decreasing into agreement with a pre-selected fiducial \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\). We also show the cross-correlation detection strength, given by \(\sqrt{\Delta\chi^2}\), on the right-hand (red) axis of (top panel). At low \(N_\text{fg}\), the \(\chi^2_\text{dof}\) appears reasonable but this is due to the larger statistical errors on the cross-power spectrum, which is fairly consistent with zero for these \(N_\text{fg}\), as identified by the low detection significance in the \(\sqrt{\Delta\chi^2}\) results. The errors are larger for low \(N_\text{fg}\) because the residual foregrounds contribute significantly more variance to the maps, even though the residuals themselves are expected to correlate out on average. Increasing \(N_\text{fg}\) from 10 to 20 does little to improve the detection significance and initially worsens the \(\chi^2_\text{dof}\) caused by a decrease in error-bar size. At \(N_\text{fg}\,{\sim}\,30\) enough components have been removed that a clear detection starts to manifest along with an improved agreement between data and model, given by the \(\chi^2_\text{dof}\,{\sim}\,1\). Going to much higher \(N_\text{fg}\) starts to over-clean the maps, reducing \(S/N\) and worsening the detection. To justify this explanation for the deterioration in results for \(N_\text{fg}\,{>}\,30\), we analysed the cross-correlation for maps constructed using just the principal components between 30 and 40. This map provided a \({\sim}\,3.4\sigma\) detection, indicating that a lot of signal is present in the modes with \(30\,{<}\,N_\text{fg}\,{<}\,40\), which explains the deterioration in \(\chi^2_\text{dof}\) beyond \(N_\text{fg}\,{=}\,30\), where these modes are gradually removed. highlights the sensitivity of results to the foreground clean, and is further evidence that residual foregrounds and systematics are spread throughout the principal components. The ratio between systematics and signal varies among the various components thus some will be more influential on the cross-correlation than others. This causes a variation in the derived parameter \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\), shown by the bottom panel of . We estimate a contribution to the error budget of \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) caused by the variance across the different \(N_\text{fg}\), discussed further in the following section.
## Constraints on \(\Omega_\textrm{H\textsc{i}}\)
Fitting the amplitude of the [H[i]{.smallcaps}]{.roman}-galaxy cross-power spectrum provides a constraint on \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) as a function of redshift. From our results in , we find \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\,{=}\,[0.86\,{\pm}\,0.10\,({\rm stat})\,{\pm}\,0.12\,({\rm sys})]\,{\times}\,10^{-3}\). The systematic error accounts for uncertainty from the map calibration and variance in results from the choice of \(N_\text{fg}\). Firstly, uncertainty from the map calibration could cause a bias to the overall amplitude of the power spectrum. We address this by studying the residuals relative to the model in our calibration study. We are able to estimate that gain uncertainties should be at a level of \({\sim}\,2\%\). Secondly, we account for the variance in results from the different number of PCA modes removed, indicative of residual systematics. We do this by evaluating the standard deviation on all \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) fits (see , bottom panel) from each reasonable choice of \(N_\text{fg} \, (10\,{<}\,N_\text{fg}\,{<}\,40)\), which we find to be \({\sim}\,0.115\), equating to a \({\sim}\,13\%\) error on \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\). The combination of these two error components added in quadrature yields the systematic error in our final constraint. The [H[i]{.smallcaps}]{.roman} bias is not yet well understood but is expected to have some scale dependence when entering non-linear scales (high-\(k\)). Furthermore, the cross-correlation coefficient \(r\), included to account for stochasticity between the two fields, will also have some scale dependence. We therefore examined how the constraint on \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) changed as we varied the scales at which it was measured. By cutting small-\(k\) data points, we change the effective scale of the measurement, calculated by \(k_\text{eff}\,{=}\,\sum_i k_i (S/N)_i^2/\sum_i(S/N)_i^2\), where \((S/N)_i\) is the signal-to-noise ratio in each \(k_i\) bin, i.e. \(\hat{P}_{\textrm{H\textsc{i}},\text{g}}(k_i)/\hat{\sigma}_{\textrm{H\textsc{i}},\text{g}}(k_i)\). The scale dependence on the measurements of \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) is shown in (red-star points) for \(N_\text{fg}\,{=}\,30\). The other coloured data points show previous intensity mapping constraints from GBT cross-correlation with galaxy surveys at \(z\,{\sim}\,0.8\). The GBT intensity maps had a significantly smaller beam than MeerKAT thus were able to probe higher-\(k\). The MeerKAT and GBT measurements are at different redshifts so a direct comparison is not possible. Despite this, there still appears to be a trend with \(k_\text{eff}\) suggesting a detection of scale dependence in \(b_\textrm{H\textsc{i}} r\). However, the continuity of the trend is affected by the choice of \(N_\text{fg}\), as shown by the different grey lines in . Furthermore, there is a possibility the scale dependence is influenced by systematics, which are mitigated in the scale cuts. The smallest-\(k\) modes are the most affected by the transfer function, with up to 80% increase in amplitude (see ). Thus, further investigation is needed to disentangle the scale-dependence of \(b_\textrm{H\textsc{i}} r\) from possible scale-dependent systematics. By making some further assumptions on \(b_\textrm{H\textsc{i}}\) and \(r\), we can isolate the constraint on \(\Omega_\textrm{H\textsc{i}}\). We assume \(r\,{=}\,0.9\,{\pm}\,0.1\) and for the bias we interpolate between the hydrodynamic simulations in and use \(b_\textrm{H\textsc{i}}\,{=}\,1.13\,{\pm}\,0.10\). Fitting the power spectrum data in across all scales (\(k_\text{eff}\,{\sim}\,0.13\,h\,\text{Mpc}^{-1}\)), provides a constraint of \(\Omega_\textrm{H\textsc{i}}\,{=}\,[0.83\,{\pm}\,0.15\,({\rm stat})\,{\pm}\,0.11\,({\rm sys})]\,{\times}\,10^{-3}\) at the redshift of \(z\,{=}\,0.43\), which is reasonably consistent with other results in the literature (see comparison in ).
# Conclusion {#sec:Conclusion}
[H[i]{.smallcaps}]{.roman} intensity mapping is a novel method for probing large scale cosmic structure and will be a primary objective for the future SKAO. To achieve this, it is necessary for the multi-dish array to operate in single-dish (auto-correlation) mode, as opposed to a conventional interferometer. In this work we have demonstrated, for the first time, the successful detection of cosmological signal using the MeerKAT multi-dish array in single-dish mode. This represents a major milestone in demonstrating the feasibility of this survey mode for SKAO. We achieved this by cross-correlating \(10.5\,\text{hrs}\) of MeerKAT pilot survey intensity maps with overlapping optical galaxies from the WiggleZ Dark Energy Survey. A measurement of the cross-power spectrum between these fields provided a \(7.7\sigma\) detection of a cross-correlation. We relied on an aggressive filtering process, removing 30 modes in a PCA-based foreground clean, necessary due to the presence of systematic contributions in the pilot survey data. This allowed us to obtain a constraint of \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\,{=}\,[0.86\,{\pm}\,0.10\,({\rm stat})\,{\pm}\,0.12\,({\rm sys})]\,{\times}\,10^{-3}\) from fitting the amplitude of the cross-power spectrum at an effective scale of \(k_\text{eff}\,{\sim}\,0.13\,h\,\text{Mpc}^{-1}\). Varying the effective scale of the measurement changed the value for \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\), something noted in previous studies. We also found \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) to have a dependence on the number of foreground modes removed, so we included this variance in the systematic error budget of the constraint. The \({\sim}\,17.8\%\) precision represents a competitive \(\Omega_\textrm{H\textsc{i}} b_\textrm{H\textsc{i}} r\) constraint relative to other intensity mapping experiments. Furthermore, with additional assumptions on \(b_\textrm{H\textsc{i}}\) and \(r\), we provided insight into the cosmic [H[i]{.smallcaps}]{.roman} density \(\Omega_\textrm{H\textsc{i}}\), for which measurements at higher redshifts are vital for understanding the evolution of [H[i]{.smallcaps}]{.roman}. The MeerKAT telescope will continue to conduct [H[i]{.smallcaps}]{.roman} intensity mapping observations in single-dish mode. With enhanced calibration techniques and more observing time, improved constraints will be possible with less aggressive foreground removal. With this we can attempt a detection of the [H[i]{.smallcaps}]{.roman} in auto-correlation, which is yet to be achieved. Observations have now been conducted in MeerKAT's UHF band (\(0.40\,{<}\,z\,{<}\,1.45\)), opening the possibility of higher redshift probes and for cross-correlating UHF-band data with the L-band data used in this work, with the aim of mitigating systematics. | {'timestamp': '2022-06-06T02:14:49', 'yymm': '2206', 'arxiv_id': '2206.01579', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01579'} |
null | null |
# Introduction {#itm:intro}
Data envelopment analysis (DEA) is a data-supported nonparametric approach that is used to measure the relative performance of decision-making units (DMUs). Charnes, Cooper, and Rhodes developed a model called the CCR DEA Model to measure the performance of DMUs. The efficiency of DMU is defined as the ratio of output to input (efficiency = output/input). The relative efficiency of a DMU is defined as, the ratio of its efficiency to the largest efficiency under consideration. It lies in the interval (0,1\]. A DMU with an efficiency score of 1 is called efficient DMU otherwise inefficient DMU. Education and research play a vital role in the development of any nation. In the current study, we will try to measure the performance efficiencies of IIMs with the help of the network DEA model, particularly the two-stage network DEA. The main reason for using the two-stage network DEA model over the conventional CCR DEA model is to understand the causes behind the inefficiencies of DMUs. In general, DEA works like a black box that does not provide any adequate detail to identify the specific reason for inefficiency in DMUs. The motivation of this study is to analyze the cause of the inefficiency of DMUs in a decision process with the help of a two-stage relational network DEA model. There are many types of research available in literature where the network DEA model is developed and used for the performance evaluation of DMUs. Kao reported a detailed review on network DEA which consists of the DMUs with series, parallel and mixed structure. Application of network DEA approach can be found in several fields like education sector, banking sector, healthcare sector, hotel industries, etc. Tan and Despotis investigated the efficiency of the hotel industry by using a network DEA model. Kao and Hung, Arya and Singh, and Ma and Chen proposed a two-stage parallel series network and provided the application of proposed models in the healthcare sector. Puri et al., Lofti and Ghasemi also used multi-component DEA in banking sector. In past studies, many researchers used conventional DEA in education sector. As Lee and Worthington explained in their work that higher educational institutions are the more complex structure that uses multiple inputs to produce multiple outputs. The research process is also complicated. As, in some stage, some output of research process become the inputs for the next stages to produce final outputs. Many researchers applied network DEA models for performance evaluation in the education sector. Lee and Worthington applied a network DEA model for evaluating the research performance of Australian universities. In this study, they developed a university research production model. The 1st stage uses input (full-time equivalent (FTE) academic staff and capital stock) to produce output as the number of publications. In the 2nd stage, the number of publications is used to produce the final output (grant secured). Our study is also motivated by Lee and Worthington's study. Tavares et al. proposed a multistage network DEA model for performance evaluation of higher educational institutions in Brazil. In this study, they proposed an approach that is used to measure the efficiencies of 45 Brazilian federal universities. Yang et al. also applied a two-stage network DEA model for analyzing the reasons behind the inefficiencies of Chinese research universities. Moreno-Gómez et al. measured efficiencies of 78 Colombian universities across the period 2015--2017 by using a two-stage network DEA model. Monfared and Safi, proposed a set of performance indicators to enable efficient analysis of academic activities and apply network DEA structure to analyze the performance of the academic colleges at Alzahra University in Iran. There are many kinds of research available in the field of network DEA, but the application of network DEA in the field of the education sector is still very limited. In the current study, by considering quantitative attributes to measure the performance of Indian institutes of management (IIMs) by using network DEA, we develop a procedure that captures both quality and quantity. In the current study, we developed a research production model to elaborate the research process in IIMs. With the help of a two-stage relational network DEA model performance analysis is done. The results are compared with the conventional DEA model. In this study, we aim to know the reasons for the inefficiencies of DMUs with the help of the network DEA model. The rest of the paper is organized into 4 sections. Introduction is given in Section [1](#itm:intro){reference-type="ref" reference="itm:intro"}. A two-stage relational network DEA model and supported statements are given in Section [2](#itm:Twostagemodel){reference-type="ref" reference="itm:Twostagemodel"}. The conceptual framework of the proposed two-stage research production model and education sector application with results is discussed in Section [3](#itm:appli){reference-type="ref" reference="itm:appli"}. Finally, some concluding remarks and future scope are given in Section [4](#itm:conclu){reference-type="ref" reference="itm:conclu"}.
# Two-stage relational network DEA model {#itm:Twostagemodel}
Suppose there are \(n\) homogeneous DMUs (\(DMU_{j};j=1, 2, 3..., n\)), which consumes \(m\) inputs, \(x_{ij},\,i=1, 2, 3..., m\) to produce \(s\) outputs \(y_{rj}, \,r=1, 2, 3...,s\). Then the efficiency score of the \(k^{th}\) DMU can be defined with the help of the following CCR DEA model:
Model 1 [\[itm:Model1\]]{#itm:Model1 label="itm:Model1"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
[\[Model1\]]{#Model1 label="Model1"} & E_k= &\
&subject to ; &\
&u\_ik i, v\_rk r, \> 0;&
:::
where \(\varepsilon\) is the non-Archimedean infinitesimal; \(y_{rj}\) is the amount of the \(r^{th}\) output produced by the \(j^{th}\) DMU; \(x_{ij}\) is the amount of the \(i^{th}\) input used by the \(j^{th}\) DMU; \(u_{ik}\) and \(v_{rk}\) are the weights corresponding to the \(i^{th}\) input and \(r^{th}\) output respectively.
```{=html}
<!-- -->
```
**Definition 1:**[\[itm: Def 1\]]{#itm: Def 1 label="itm: Def 1"}
: \(E_{k}\) is called the CCR efficiency of \(DMU_k\). \(E_{k}\) lies in the interval (0,1\]. Let \(E_{k}^{*}\) be the optimal value of \(E_{k}\). Then \(DMU_{k}\) is said to be **CCR-efficient** if \(E_{k}^{*}=1\), an, \(DMU_{k}\) is said to be **CCR-inefficient** if \(E_{k}^{*}<1\).\
Now, we consider a production process composed of two sub-processess as depicted in Fig [\[fig: Fig. 1.\]](#fig: Fig. 1.){reference-type="ref" reference="fig: Fig. 1."}. The whole process consists of two stages. The 1st stage uses m inputs \(x_{ik}, ~ i=1,2...,m\) to produce p intermediate products \(z_{dk},~ d=1,2,3...,p\). The 2nd stage uses p intermediate products \(z_{dk},~ d=1,2,3...,p\) to produce s outputs \(y_{rk},~ r=1,2,3...,s\). The whole process uses m inputs \(x_{ik}, ~ i=1,2...,m\) to produce s outputs \(y_{rk},~ r=1,2,3...,s\). Moreover, the outputs of 1st stage (i.e. \(z_{dk}\)) are the inputs of the 2nd stage. The efficiency of 1st and 2nd stage can be calculated by the conventional CCR DEA model explained in . Suppose the efficiency of \(k^{th}\) DMU in 1st stage is \(E_{k}^{1}\) and in the 2nd stage is \(E_{k}^{2}\). Then the 1st and 2nd stage efficiencies of \(DMU_k\) can be calculated with the help of and as follows :
Model 2[\[itm:Model2\]]{#itm:Model2 label="itm:Model2"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
& E\_k\^1= &\
&subject to ; &\
&u\_ik i, w\_dk p, \> 0.&
:::
```{=html}
<!-- -->
```
Model 3 [\[itm:Model 3\]]{#itm:Model 3 label="itm:Model 3"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
& E\_k\^2= &\
&subject to ; &\
&v\_rk r, w'\_dk p, \> 0.&
:::
Note that the weights \(w_{dk}\) and \(w'_{dk}\) used in and are same. The reason is, input of sub-process 2 to be the expected output of sub-process 1 while calculating the expected output of the whole process. When this concept is generalized to the case of multiple intermediate products, it requires the aggrigated value of intermediate products, which is \(\sum_{d=1}^{p}w_{dk}z_{dk}\) to be the same. Hence can be re-written as follows:
Model 4[\[itm:Model4\]]{#itm:Model4 label="itm:Model4"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
& E\_k\^2= &\
&subject to ; &\
&v\_rk r, w\_dk p, \> 0.&
:::
With the help of , efficiency of the whole process can be calculated independently as the 1st stage efficiency model () and 2nd stage efficiency model () are essentially same. () and () give the 1st and second stage efficiencies of independent network DEA model. To link these two sub-processes with the whole process, a model must be constructed that describe the series relationship of the whole process with the two sub-processes.\
Suppose, for \(DMU_k\); \(u_{ik}^{*}\), \(v_{rk}^{*}\), \(w_{dk}^{*}\) are the optimal weights (multipliers) that are obtained while calculating overall efficiency \(E_k\), sub-process efficiencies \(E_{k}^{1}\), and \(E_{k}^{2}\). Then we have,\
From Eqs. [\[Eq: Eq:9\]](#Eq: Eq:9){reference-type="ref" reference="Eq: Eq:9"},[\[Eq: Eq:10\]](#Eq: Eq:10){reference-type="ref" reference="Eq: Eq:10"} & [\[Eq: Eq:11\]](#Eq: Eq:11){reference-type="ref" reference="Eq: Eq:11"}, we have a relation among \(E_k\), \(E_{k}^{1}\), and \(E_{k}^{2}\) as; \[\textbf{\(E_k= E_{k}^{1} \times E_{k}^{2}\)}. \label{Eq: Eq. (12)}\\ \] Based on this concept, let us incorporate the ratio constraints of the two sub-processes into conventional DEA model i.e., , to get a two-stage relational network DEA model. This will give the overall efficiency \(E_{k}\), to a two-stage relational network DEA model, taking into account the series relationship of two sub-processes as follows:
Model 5 [\[itm:Model5\]]{#itm:Model5 label="itm:Model5"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
& E_k= &\
&subject to ; &\
& ;&\
& ; &\
&u\_ik i, v\_rk r, w\_dk p,\> 0.&
:::
Now, can be written in linear form as follows:
Model 6 [\[itm:Model6\]]{#itm:Model6 label="itm:Model6"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
& E_k= \_r=1\^sy\_rkv\_rk &\
&subject to \_i=1\^mx\_iku\_ik=1;&\
& \_r=1\^sy\_rjv\_rk-\_i=1\^mx\_iju\_ik ; &\
&\_d=1\^pz\_djw\_dk-\_i=1\^mx\_iju\_ik ;&\
&\_r=1\^sy\_rjv\_rk-\_d=1\^pz\_djw\_dk ; &\
&u\_ik i, v\_rk r, w\_dk p,\> 0.&
:::
It is quite possible that the optimal solution (optimal multipliers) obtained after solving are not unique. Therefore, the relation **\(E_k= E_{k}^{1} \times E_{k}^{2}\)** would not be unique. This makes comparison of \(E_{k}^{1}\) and \(E_{k}^{2}\) between all DMUs not the same basis. One solution to this problem is to find a set of multipliers that produce the largest \(E_k^{1}\) while maintaning the \(E_k\) obtained from .
Model 7 [\[itm:Model7\]]{#itm:Model7 label="itm:Model7"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
& E\_k\^1= \_d=1\^pz\_dkw\_dk &\
&subject to \_i=1\^mx\_iku\_ik=1;&\
& \_r=1\^sy\_rkv\_rk-E\_k\_i=1\^mx\_iku\_ik = 0; &\
& \_r=1\^sy\_rjv\_rk-\_i=1\^mx\_iju\_ik ; &\
&\_d=1\^pz\_djw\_dk-\_i=1\^mx\_iju\_ik ;&\
&\_r=1\^sy\_rjv\_rk-\_d=1\^pz\_djw\_dk ; &\
&u\_ik i, v\_rk r, w\_dk p,\> 0.&
:::
After calculating \(E_{k}^{1}\) with the help of , second stage efficiency efficiency \(E_{k}^{2}\) can be calculated easily by using the relation ( Eq. [\[Eq: Eq. (12)\]](#Eq: Eq. (12)){reference-type="ref" reference="Eq: Eq. (12)"}) as \(E_{k}^{2}= E_{k}/ E_{k}^{1}\). Silmilarly, if decision-maker is more concerned for second stage efficiency then, one can formulated model for calculating \(E_{k}^{2}\) by the following model as follows:
Model 8 [\[itm:Model8\]]{#itm:Model8 label="itm:Model8"}
: : For \(k=1,2,3...,n,\)
::: {.flalign}
& E\_k\^2= \_r=1\^sy\_rkv\_rk &\
&subject to \_d=1\^pz\_dkw\_dk=1;&\
& \_r=1\^sy\_rkv\_rk-E\_k\_i=1\^mx\_iku\_ik = 0; &\
& \_r=1\^sy\_rjv\_rk-\_i=1\^mx\_iju\_ik ; &\
&\_d=1\^pz\_djw\_dk-\_i=1\^mx\_iju\_ik ;&\
&\_r=1\^sy\_rjv\_rk-\_d=1\^pz\_djw\_dk ; &\
&u\_ik i, v\_rk r, w\_dk p,\> 0.&
:::
The efficiency of 1st stage \(E_{k}^{1}\) can be calculated easily by using the relation ( Eq. [\[Eq: Eq. (12)\]](#Eq: Eq. (12)){reference-type="ref" reference="Eq: Eq. (12)"}) as \(E_{k}^{1}= E_{k}/ E_{k}^{2}\).
Proof:
: Suppose \((u_{ik}^{*}, v_{rk}^{*},w_{dk}^{*}~ \forall~ i,r,d)\) be an optimal solution of then we have\
\(E_k= {\sum_{r=1}^{s}y_{rk}v_{rk}^*}\),
: and
::: {.flalign}
&\_i=1\^mx\_iku\_ik\^\*=1;&\
& \_r=1\^sy\_rjv\_rk\^\*-\_i=1\^mx\_iju\_ik\^\* ; &\
&\_d=1\^pz\_djw\_dk\^\*-\_i=1\^mx\_iju\_ik\^\* ;&\
&\_r=1\^sy\_rjv\_rk\^\*-\_d=1\^pz\_djw\_dk\^\* ; &\
&u\_ik\^\* i, v\_rk\^\* r, w\_dk\^\* p,\> 0.&
:::
With \(E_{k}=1\), we have, for specific \(DMU_k:\)\
\[1={\sum_{r=1}^{s}y_{rk}v_{rk}^*}\leq {\sum_{d=1}^{p}z_{dk}w_{dk}^*} \leq {\sum_{i=1}^{m}x_{ik}u_{ik}^*} =1 \label{Eq: Eq (38)}\] Therefore, from Eq. ([\[Eq: Eq (38)\]](#Eq: Eq (38)){reference-type="ref" reference="Eq: Eq (38)"}) it is possible that \({\sum_{r=1}^{s}y_{rk}v_{rk}^*}= {\sum_{d=1}^{p}z_{dk}w_{dk}^*}=1\). In other words, the optimal solution \((u_{ik}^{*}, v_{rk}^{*},w_{dk}^{*}~ \forall~ i,r,d)\) can make the objective function values of and are equal to 1. It is easy to varify that \((u_{ik}^{*}, v_{rk}^{*},w_{dk}^{*}~ \forall~ i,r,d)\) can satisfy all the constraints of and simultaneously.\
Thus, we have \(E_{k}^{1}=E_{k}^{2}=1\) for \(DMU_k.\) \(~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\square\)
# Education sector application {#itm:appli}
For the application of the network DEA, we consider the problem of checking the performance efficiency of the Indian Institutes of Management (IIMs) in India. IIMs are administrative and research educational institutions in India. Primarily they offer undergraduate, postgraduate, doctoral, and other additional courses. IIMs have been declared the most important institutions in the country by The Ministry of Education Govt. of India.
## Conceptual framework
Education and research play a vital role in the development of any nation. In the current study, we will try to measure the performance efficiencies of IIMs with the help of the network DEA model, particularly the two-stage network DEA. The main reason for using the two-stage network DEA model over the conventional CCR DEA model is to understand the causes behind the inefficiencies of DMUs. The network DEA model used for the current study is depicted in Fig [\[fig: Fig. 2.\]](#fig: Fig. 2.){reference-type="ref" reference="fig: Fig. 2."}. Here, in this study, we constructed a two-stage research production model of IIMs. The research production model in IIMs is complex. Fig [\[fig: Fig. 2.\]](#fig: Fig. 2.){reference-type="ref" reference="fig: Fig. 2."} depicts our network DEA model for IIMs. This model consists of two stages. In the 1st stage, the model consumes \(x_{ij}\) inputs to produce \(z_{dj}\) intermediate products. Outputs of 1st stage start working as inputs in 2nd stage. In our research production model, we use three inputs (\(x_1\): financial resources, \(x_2\): number of Ph.D. students, \(x_1\): number of faculty members) in 1st stage. In research, these three parameters play a very crucial role. For any Ph.D. student, it is always challenging to research without a guide and financial support. This is the reason why we considered these three parameters as inputs in 1st stage. The output of 1st stage or intermediate product is considered as the 'number of publications' (\(z_1\)). This is the most obvious output for any research production model. In India, Ph.D. students have to publish a few research papers to submit their thesis. These publications validate the choice of inputs as, Ph.D. students, faculties, and financial resources. The output of the 1st stage \(z_1\) is used as an intermediate input in the 2nd stage to produce the final output \(y_1\) (research grant). Since in the present study, we are focusing only on research activities so we are not very much bothered about other parameters. For any IIM, it is very important to win ample research grants so that they can continue their research process. IIMs get their research grant from two types of research projects; first as the sponsored research projects, and second as consultancy research projects. In our study, we consider the research grant index (\(y_1\)) as the sum of research grants obtained from sponsored research projects and consultancy research projects. It is a widespread perception that can be treated as the fact that if any IIMs publishes more research papers then that IIM can win more research grants. Hence the choice of intermediate input \(z_1\) is the best choice for the final output \(y_1\) for our two-stage research production model.\
## Results and discussion
In the current study, IIMs are considered as our DMUs. The data is collected from the NIRF website launched by The Ministry of Education on 29th September 2015. Data is collected for the year 2020-21. Input, intermediate input, and output data for 13 IIMs for the application two-stage network DEA model is given in Table [\[tab:Table-1\]](#tab:Table-1){reference-type="ref" reference="tab:Table-1"}.\
In our research production model, the efficiency of 1st stage measures the performance of IIMs based on publications, while the 2nd-stage measures the performance on the basis of the research grant won by any IIM. By applying , the overall efficiency of two-stage relational network DEA model for 13 IIMs is calculated. The results are shown in Table [\[tab:Table-2\]](#tab:Table-2){reference-type="ref" reference="tab:Table-2"} under the subcolumn \(E_k\) of the column two-stage relational network DEA model. Since, to win an attractive research grant is of more concern for any IIM so that the research process can be performed without any obstacle, we choose to measure \(E_k^{2}\) first with the help of \(E_k\) and after this calculate \(E_k^{1}\) via \(E_k^{1}={E_k}/{E_k^{2}}\). The choice of selecting performance measures as a research grant or a number of publications is completely dependent upon the decision-maker. Here, the research grants are our primary concern, while publications can be the topic of paramount concern for some other decision-maker. This choice is not going to affect the 1st and 2nd stage efficiency scores of a two-stage relational network DEA model.\
Now we will analyze the results obtained from applying , , and on the data given in Table [\[tab:Table-1\]](#tab:Table-1){reference-type="ref" reference="tab:Table-1"}. The results are given under the column named as **'two-stage relational network DEA model'**. Note that in Table [\[tab:Table-2\]](#tab:Table-2){reference-type="ref" reference="tab:Table-2"}, none of the 13 IIMs, performed efficiently in both stages. IIM Bangalore achieved the highest overall efficiency score of 0.4973 while, IIM Ranchi achieved the lowest overall efficiency score of 0.0007. For the 1st stage, there are 3 IIMs named, IIM Udaipur, IIM Shillong, and IIM Ranchi, which perform efficiently; as they have an efficiency score \(E_k^{1}=1.\) In the 2nd stage, IIM Bangalore is the only IIM that is efficient. As we have already derived a relation **\(E_k= E_{k}^{1} \times E_{k}^{2}\)** in Eq. ([\[Eq: Eq. (12)\]](#Eq: Eq. (12)){reference-type="ref" reference="Eq: Eq. (12)"}), indicates that the overall efficiency is the product of 1st and 2nd stage efficiencies. So there is no \(E_k\) greater than its corresponding \(E_{k}^{1} ~or~ E_{k}^{2}\). It is to be noted that IIM Udaipur (\(D_7\)), IIM Shillong (\(D_{11}\)), and IIM Ranchi (\(D_{13}\)) perform efficiently in 1st stage but not efficiently in the 2nd stage. It means that these IIMs are good at publishing research but still not able to win an attractive research grant. Whereas IIM Bangalore (\(D_1\)) does not perform well in the 1st stage but in the 2nd stage it is performing efficiently. This indicates that it is a bit easy for IIM Bangalore to win ample research grants. It is observable from Table [\[tab:Table-2\]](#tab:Table-2){reference-type="ref" reference="tab:Table-2"} that except for IIM Bangalore, the efficiency of the 1st stage is higher than the efficiency of the 2nd stage for all IIMs. This indicates that the low \(E_k\) of the whole process is mainly due to the low \(E_{k}^{2}\) of the 2nd stage (i.e., research grant). Since \(E_k\) is always less than or equal to \(E_{k}^{1}\), and \(E_{k}^{2}\) so it will be more informative to look at the ranks, which are the numbers in brackets in Table [\[tab:Table-2\]](#tab:Table-2){reference-type="ref" reference="tab:Table-2"}, over the efficiency scores. If any IIM has similar ranks in \(E_{k},~E_{k}^{1}, ~and~ E_{k}^{2}\) then it can be said that the particular IIM performs evenly in the whole process and in both the sub-processes. Hence, with the help of a two-stage relational network DEA, we are in the position to understand the reason behind the inefficiency of any IIM.\
Kao and Hung explained logically that \(E_k\) for any IIM should lie either between \(E_{k}^{1}, ~and~ E_{k}^{2}\) or in the neighborhood of \(E_{k}^{1}, ~and~ E_{k}^{2}\). The main reason for this is that the performance of the whole process is the aggregation of two sub-processes. Here in Table [\[tab:Table-2\]](#tab:Table-2){reference-type="ref" reference="tab:Table-2"}, it can be seen that out of 13 IIMs the ranks of 10 IIMs lies between \(E_{k}^{1}, ~and~ E_{k}^{2}\), and the ranks of remaining 3 IIMs lies in the neighborhood of \(E_{k}^{1}, ~or~ E_{k}^{2}\).\
Now, with the help of two-stage relational network DEA model we are some what able to understand the actual reasons of inefficiencies in DMUs. As we know that the DEA works like a black box that does not provide any adequate detail to identify the specific reason for inefficiency in decision-making units (DMUs). After knowing the reasons of inefficiecies, now we are interested in the comparison of the efficiency of the whole research production process with the two-stage relational network DEA model (\(E_k\)) and conventional CCR DEA model (\(E_k^{CCR}\)). Actually, we wanted to understand what will happen if the overall efficiency of DMUs is calculated by taking sub-stages of the model in mind and without bothering about sub-stages. The comparison is done and the results are shown in Table [\[tab:Table-3\]](#tab:Table-3){reference-type="ref" reference="tab:Table-3"}. The results depicted in Table [\[tab:Table-3\]](#tab:Table-3){reference-type="ref" reference="tab:Table-3"} shows that the conventional CCR efficiency (\(E_k^{CCR}\)) is higher than two-stage relational network DEA efficiency (\(E_k\)). But it will be more appropriate to compare the ranks of IIMs rather than their efficiency scores because the relational model has two more sets of constraints than the conventional CCR DEA model. These additional sets of constraints represent the 1st and 2nd stage efficiency scores. The 1st and 2nd stage efficiency score can be affecting factors in the overall efficiency scores of IIMs in the two-stage relational network DEA model as explained earlier. Kao and Hung, mathematically, explained in detail why this deviation occurs in the \(E_k\) and \(E_k^{CCR}\) by constructing the dual model of the relational two-stage network DEA model. A Spearman's rank correlation coefficient is calculated for \(E_k\) and \(E_k^{CCR}\). The correlation coefficient \((\rho=0.91758)\) indicates that efficiencies obtained from both two-stage relational and conventional CCR DEA models are highly correlated. This means that the efficiency scores from both models produce almost the same rankings. This can be easily observed in Table [\[tab:Table-3\]](#tab:Table-3){reference-type="ref" reference="tab:Table-3"}.
# Conclusions {#itm:conclu}
The main objective of the current study is to illustrate the process of research of IIMs with the help of the proposed research production model. This study applies a two-stage relational network DEA model for the performance analysis of IIMs. In our proposed research production model there are two sub-processes. As we know that higher educational institutions are complex structures that use multiple inputs to produce multiple outputs. The research process is also complicated. As, in some stage, some output of research process become the inputs for the next stages to produce final outputs. Our proposed research production model explains the use of inputs and outputs in sub-stages very interestingly. Hence, it is pretty reasonable to consider the research production model for IIM in sub-stages, not as simply one that transforms input to produce outputs. As we know that in general, data envelopment analysis (DEA) works like a black box that does not provide any adequate detail to identify the specific reason for inefficiency in decision-making units (DMUs). Our study shows that the efficiency level of IIMs is higher in the conventional DEA model than the two-stage relational network DEA model. The reason for the deviation in both types of efficiency levels is explained in our study. So it will be more appropriate to compare the ranks of IIMs rather than their efficiency scores. A Spearman's rank correlation coefficient is calculated for \(E_k\) and \(E_k^{CCR}\). The correlation coefficient \((\rho=0.91758)\) indicates that efficiencies obtained from both two-stage relational and conventional CCR DEA models are highly correlated. This means that the efficiency scores from both models produce almost the same rankings. But due to the use of the two-stage relational network DEA model, we are now able to know the reasons behind the inefficiencies of IIMs. We found that many IIMs performed well on the 1st stage but not on the 2nd stage. Based on this, we can conclude that for IIMs it is easier to produce research publications than to win attractive research grants. In a real-world situation, input-output data can be imprecise. The future research direction can lead us to the development and application of a two-stage network DEA model in a fuzzy environment. | {'timestamp': '2022-06-06T02:14:23', 'yymm': '2206', 'arxiv_id': '2206.01561', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01561'} |
null | null |
# Introduction
The main contribution of the present paper is verification of the following elementary inequality conjectured by de Dios Pont, Greenfeld, Ivanisvili, and Madrid.
By taking \(a=b=1\) one clearly sees that formula [\[eq:defofpk\]](#eq:defofpk){reference-type="eqref" reference="eq:defofpk"} gives the smallest possible exponent \(p_k\) such that estimate [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} can hold. The particular case \(k=2\) of this estimate was established by Kane and Tao, while the authors of verified it for all \(k\leq10\) in, with a comment that they also performed verification for \(k\leq100\) with an aid of a computer. The paper calls [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} simply *an elementary inequality*, while also calls it *a subtle inequality for the Legendre polynomials*. Namely, if \(P_k\) denote the Legendre polynomials, \[P_k(z) = \frac{1}{2^k} \sum_{j=0}^k \binom{k}{j}^2 (z+1)^{k-j} (z-1)^{j},\] see, then inequality [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} can be rephrased as \[P_k(z) \leq \bigg( \Big(\frac{z+1}{2}\Big)^{k/p_k} + \Big(\frac{z-1}{2}\Big)^{k/p_k} \bigg)^{p_k} \quad\text{for } z\in[1,\infty);\] see the details in. The importance of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} comes from the following application in additive combinatorics. For a positive integer \(k\) the notion of *\(k\)-additive energy* \(E_k(A)\) of a finite set \(A\subset\mathbb{Z}^d\) was defined in the paper as the number of \(2k\)-tuples \((a_1,\ldots,a_{2k})\in A^{2k}\) such that \(a_1+\cdots+a_{k}=a_{k+1}+\cdots+a_{2k}\). In the particular case \(k=2\) this specializes to the well-known concept of the *additive energy*; see. Let \(|A|\) denote the cardinality of \(A\).
By taking \(A=\{0,1\}^d\) one again sees that \(p_k\) is the smallest possible exponent such that inequality [\[eq:mainadditive\]](#eq:mainadditive){reference-type="eqref" reference="eq:mainadditive"} can hold. This sharp estimate was conjectured by de Dios Pont, Greenfeld, Ivanisvili, and Madrid, who also showed how it can be derived from [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} by using it in the step of the mathematical induction on the dimension \(d\); see. The same deduction was previously performed for \(k=2\) by Kane and Tao. Thus, prior to the present paper, Corollary [\[cor:additive\]](#cor:additive){reference-type="ref" reference="cor:additive"} has only been confirmed for small values of \(k\), namely \(k\leq100\); see and. We use the opportunity to also give a probabilistic reformulation of inequality [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"}, which will not be needed in its proof, but it might be interesting on its own. Suppose that \(X_1,X_2,\ldots\) are independent identically distributed random variables that take values in \(\{-1,0,1\}\) and satisfy \(\mathbb{P}(X_1=0)=1/2\). Let \(S_n=X_1+\cdots+X_n\) denote the associated random walk on \(\mathbb{Z}\) starting at \(0\). Such a process is often called *a lazy simple random walk*, as it only makes a move with probability \(1/2\); see for instance. Note that the distribution of this discrete stochastic process \((S_n)_{n=0}^\infty\) is uniquely determined by a single number, namely the probability of taking a step to the right, \(\mathbb{P}(X_1=1)\in[0,1/2]\).
The number \(\mathbb{P}(S_k=-k)\) (resp.\(\mathbb{P}(S_k=k)\)) can be interpreted as the probability that the first \(k\) steps of the random walk are all made to the left (resp.right). Clearly, \(\mathbb{P}(S_k=0)\) is the probability that, after \(k\) steps, the random walk ends up at its starting point. The proof of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} is given in Section [2](#sec:theorem){reference-type="ref" reference="sec:theorem"}. It only needs basic single-variable calculus. A crucial ingredient is a second-order ordinary differential equation coming from the classical differential equation for the Legendre polynomials, even though we do not work with special polynomials at all. Mathematica is used extensively in two different ways.[^1] First, finitely many inequalities for concrete real numbers are verified as parts of the proofs of Lemmata [\[lm:easyest\]](#lm:easyest){reference-type="ref" reference="lm:easyest"} and [\[lm:cisnegative\]](#lm:cisnegative){reference-type="ref" reference="lm:cisnegative"}, by computing relevant numerical expressions using infinite precision (i.e., tracking the propagation of the numerical error) and sufficient accuracy (which is in all of our cases chosen to be \(10\) accurate digits after the leading zeros). In other words, a numerical expression `e` is always approximated using the command `N[e,{`\(\infty\)`,10}]`. Second, symbolic differentiation via the command `D` and algebraic simplification via the command `Simplify` are used in the proof of Lemma [\[lm:differential\]](#lm:differential){reference-type="ref" reference="lm:differential"}. All these operations are perfectly reliable; see the Wolfram Language Documentation Center. Note that we do not rely on testing infinitely many inequalities for real numbers, or on any sketches of graphs of functions. Corollary [\[cor:probab\]](#cor:probab){reference-type="ref" reference="cor:probab"} is established in Section [3](#sec:corollary){reference-type="ref" reference="sec:corollary"} by showing that [\[eq:probineq\]](#eq:probineq){reference-type="eqref" reference="eq:probineq"} is just a restatement of the elementary inequality [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"}. Finally, Section [4](#sec:means){reference-type="ref" reference="sec:means"} gives yet another reformulation of [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"}, in terms of the means of a pair of numbers, suggested to the author by Jairo Bochi.
# Proof of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} {#sec:theorem}
For each positive integer \(k\) let us define a function \(f_k\colon[0,1]\to[0,\infty)\) by the formula \[f_k(x) := \sum_{j=0}^{k} \dbinom{k}{j}^2 (1-x)^{p_k(k-j)/k} x^{p_k j/k}.\] When \(x=0\) or \(x=1\), the expression \(0^0\) is interpreted as \(1\), as is common in relation with discrete sums. The desired inequality [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} is homogeneous of order \(p_k\) in \(a,b\). Thus, one can additionally assume \(a+b=1\) and then [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} simply reads \[\label{eq:mainineq2} f_k(x)\leq 1 \quad\text{for every } x\in[0,1].\] Note that [\[eq:mainineq2\]](#eq:mainineq2){reference-type="eqref" reference="eq:mainineq2"} is trivial for \(k=1\), since \(f_1\) is identically equal to \(1\). Throughout this section we assume that \(k\geq2\) is a fixed integer. The following exact form of Stirling's formula is shown in: \[\label{eq:Stirling} (2\pi)^{1/2} n^{n+1/2} e^{-n+1/(12n+1)} < n! < (2\pi)^{1/2} n^{n+1/2} e^{-n+1/12n}\] for every positive integer \(n\). From [\[eq:Stirling\]](#eq:Stirling){reference-type="eqref" reference="eq:Stirling"} we get \[\frac{2^{2k}}{\sqrt{\pi k}} e^{-1/6k} < \binom{2k}{k} < \frac{2^{2k}}{\sqrt{\pi k}},\] so taking logarithms and using \((\log_2 e)/6<1/4\) yields \[\label{eq:ineqforpk} 2k-\frac{1}{2}\log_2(\pi k)-\frac{1}{4k} < p_k < 2k-\frac{1}{2}\log_2(\pi k).\] In particular, we certainly have \[\label{eq:ineqforpk2} k < p_k < 2k-1.\]
Now we are in position to give a short proof of the main result.
# Proof of Corollary [\[cor:probab\]](#cor:probab){reference-type="ref" reference="cor:probab"} {#sec:corollary}
Take an arbitrary parameter \(q\in[0,1]\) and let \(Y_1,Z_1,Y_2,Z_2,\ldots\) be independent random variables with distributions \[Y_i \sim \left(\begin{matrix} 0 & 1 \\ 1-q & q \end{matrix}\right), \quad Z_i \sim \left(\begin{matrix} 0 & 1 \\ 1/2 & 1/2 \end{matrix}\right)\] for every index \(i=1,2,\ldots\). Observe that \[Y_i-Z_i \sim \left(\begin{matrix}-1 & 0 & 1 \\ (1-q)/2 & 1/2 & q/2 \end{matrix}\right),\] so by choosing \(q=2\mathbb{P}(X_1=1)\) we achieve that the sequence \((Y_i-Z_i)_{i=1}^{\infty}\) has the same (joint) distribution as \((X_i)_{i=1}^{\infty}\). Inequality [\[eq:probineq\]](#eq:probineq){reference-type="eqref" reference="eq:probineq"} now becomes \[\begin{aligned}
& \bigg(\sum_{j=0}^{k} \mathbb{P}(Y_1+\cdots+Y_k=j,\, Z_1+\cdots+Z_k=j)\bigg)^{1/p_k} \\ & \leq \mathbb{P}(Y_1=\cdots=Y_k=0,\, Z_1=\cdots=Z_k=1)^{1/p_k} \\ & \quad + \mathbb{P}(Y_1=\cdots=Y_k=1,\, Z_1=\cdots=Z_k=0)^{1/p_k}. \end{aligned}\] This simplifies further as \[\label{eq:altprob} \bigg(\sum_{j=0}^{k} \dbinom{k}{j} (1-q)^{k-j} q^j \dbinom{k}{j}\Big(\frac{1}{2}\Big)^k \bigg)^{1/p_k} \leq \Big(\frac{1-q}{2}\Big)^{k/p_k} + \Big(\frac{q}{2}\Big)^{k/p_k}\] for every positive integer \(k\) and every \(q\in[0,1]\), which becomes precisely [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} with \[\label{eq:altparam} a=\Big(\frac{1-q}{2}\Big)^{k/p_k}, \quad b=\Big(\frac{q}{2}\Big)^{k/p_k}.\] Conversely, by the homogeneity of [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} one is allowed to add an additional constraint \(a^{p_k/k}+b^{p_k/k}=1/2\), which allows us to parameterize the pair \((a,b)\) as in [\[eq:altparam\]](#eq:altparam){reference-type="eqref" reference="eq:altparam"} for some \(q\in[0,1]\) and then [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} turns precisely into [\[eq:altprob\]](#eq:altprob){reference-type="eqref" reference="eq:altprob"}. The claim about the optimality of \(p_k\) then also follows from the initial comments on the sharpness of [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"}.
# Comments on number means {#sec:means}
For a positive integer \(k\) the quantity \[\mathfrak{M}_2^{[k,k]}(x,y) := \bigg( \frac{\sum_{j=0}^{k}\binom{k}{j}^2 x^{k-j} y^{j}}{\binom{2k}{k}} \bigg)^{1/k}\] is *the Whiteley mean* with parameters \(k,k\) of the numbers \(x,y\in[0,\infty)\). It can also be understood as *the \(k\)-th elementary symmetric polynomial mean* of the numbers \[\underbrace{x,\ldots,x}_{k},\underbrace{y,\ldots,y}_{k}.\] On the other hand, \[\mathfrak{M}_2^{[r]}(x,y) := \begin{cases} \bigg( {\displaystyle\frac{x^r + y^r}{2}} \bigg)^{1/r} & \text{for } r\in\mathbb{R},\ r\neq0, \\ \sqrt{x y} & \text{for } r=0 \end{cases}\] is the well-known *power mean* with exponent \(r\). Substituting \(x=a^{p_k/k}\), \(y=b^{p_k/k}\) the main inequality of this paper [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"} can be reformulated equivalently as \[\label{eq:meansineq1} \mathfrak{M}_2^{[k,k]}(x,y) \leq \mathfrak{M}_2^{[r_k]}(x,y) \quad\text{for } x,y\in[0,\infty),\] where \[r_k = \frac{k}{\log_2\binom{2k}{k}} \in\Big(\frac{1}{2},1\Big].\] Choosing \(x=1\), \(y=0\) one easily observes that this exponent \(r_k\) is the smallest one such that [\[eq:meansineq1\]](#eq:meansineq1){reference-type="eqref" reference="eq:meansineq1"} can hold. A very special case of a result by Bochi, Iommi, and Ponce showed a weaker inequality, \[\mathfrak{M}_2^{[k,k]}(x,y) \leq 2^{2-1/r_k} \mathfrak{M}_2^{[1/2]}(x,y),\] which would not be sufficient for our intended application to Corollary [\[cor:additive\]](#cor:additive){reference-type="ref" reference="cor:additive"}, but it becomes the same as [\[eq:meansineq1\]](#eq:meansineq1){reference-type="eqref" reference="eq:meansineq1"} in the limit as \(k\to\infty\). Nice observations from this paragraph have all been communicated to the author by Jairo Bochi. In the other direction, Bochi, Iommi, and Ponce also proved \[\mathfrak{M}_2^{[k,k]}(x,y) \geq \mathfrak{M}_2^{[1/2]}(x,y).\] We are in position to give a sharpening of the last estimate: \[\label{eq:meansineq2} \mathfrak{M}_2^{[k,k]}(x,y) \geq \mathfrak{M}_2^{[k/(2k-1)]}(x,y) \quad\text{for } x,y\in[0,\infty)\] and every positive integer \(k\). In fact, the exponent \(r=k/(2k-1)\) is the largest one such that inequality \(\mathfrak{M}_2^{[k,k]}(x,y) \geq \mathfrak{M}_2^{[r]}(x,y)\) can hold; this is easily seen by observing the asymptotic expansions: \[\begin{aligned}
\mathfrak{M}_2^{[k,k]}(1+\varepsilon,1-\varepsilon) & = 1-\frac{k-1}{2(2k-1)}\varepsilon^2 + O(\varepsilon^3), \\ \mathfrak{M}_2^{[r]}(1+\varepsilon,1-\varepsilon) & = 1-\frac{1-r}{2}\varepsilon^2 + O(\varepsilon^3) \end{aligned}\] as \(\varepsilon\to0^+\). Therefore, [\[eq:meansineq1\]](#eq:meansineq1){reference-type="eqref" reference="eq:meansineq1"} and [\[eq:meansineq2\]](#eq:meansineq2){reference-type="eqref" reference="eq:meansineq2"} together give optimal comparisons of the Whiteley mean \(\mathfrak{M}_2^{[k,k]}(x,y)\) with power means. The proof of inequality [\[eq:meansineq2\]](#eq:meansineq2){reference-type="eqref" reference="eq:meansineq2"} is much simpler than that of [\[eq:mainineq\]](#eq:mainineq){reference-type="eqref" reference="eq:mainineq"}. We only need to reuse a few ideas from Section [2](#sec:theorem){reference-type="ref" reference="sec:theorem"}. Substituting \(a=x^{k/(2k-1)}\), \(b=y^{k/(2k-1)}\) we transform [\[eq:meansineq2\]](#eq:meansineq2){reference-type="eqref" reference="eq:meansineq2"} into \[\sum_{j=0}^{k} \dbinom{k}{j}^2 a^{(2k-1)(k-j)/k} b^{(2k-1)j/k} \geq \dbinom{2k}{k} 2^{-2k+1} (a+b)^{2k-1} \quad\text{for } a,b\in[0,\infty).\] By homogeneity we can normalize \(a+b=1\). Thus, we only need to prove \[h_k(x) \geq h_k\Big(\frac{1}{2}\Big) \quad \text{for every } x\in[0,1]\] and a function \(h_k\colon[0,1]\to[0,\infty)\) defined by \[h_k(x) := \sum_{j=0}^{k} \dbinom{k}{j}^2 (1-x)^{(2k-1)(k-j)/k} x^{(2k-1)j/k}.\] Recall that the proof of Lemma [\[lm:differential\]](#lm:differential){reference-type="ref" reference="lm:differential"} did not use the particular value of \(p_k\), so it can be applied with \(p_k\) replaced with \(2k-1\). This observation yields the differential equation \[\label{eq:diffeqhk} \widetilde{a}_k(x) h_k''(x) + \widetilde{b}_k(x) h_k'(x) + (2k-1) \widetilde{c}_k(x) h_k(x) = 0,\] where \[\begin{aligned}
\widetilde{a}_k(x) & := (1-x)^2 x^2 \big((1-x)^{2-1/k}-x^{2-1/k}\big)^2, \\ \widetilde{b}_k(x) & := (1-x) x \big((1-x)^{2-1/k}-x^{2-1/k}\big) \\ & \qquad\times \Big((1-x)^{2-1/k}\big(1+4(k-1)x\big)+x^{2-1/k}\big(1+4(k-1)(1-x)\big)\Big), \\ \widetilde{c}_k(x) & := (1-x)^{4-2/k} x \big(1+2(k-1)x\big) + x^{4-2/k} (1-x) \big(1+2(k-1)(1-x)\big) \\ & \qquad-(1-x)^{2-1/k} x^{2-1/k} \big(2k-1-4(k-1)(1-x)x\big). \end{aligned}\] Note that for \(k\geq2\) and \(x\in(0,1/2)\) we have \(a_k(x)>0\) and \(c_k(x)>0\). Indeed, by substituting \(z=((1-x)/x)^{1/k}>1\), simplifying, and factoring polynomials, positivity of \(c_k\) reduces to \[(z-1) (z^{2k-1}-1) \Big(\sum_{i=1}^{k-1}(z^i + z^{-i}-2)\Big) > 0,\] which clearly holds. It is easy to see \(h_k(1-x)=h_k(x)\) and \(h_k(0)\geq h_k(1/2)\). Now take \(k\geq2\) and suppose that \(h_k\) attains its minimum at some point \(x_{\min}\in(0,1/2)\). Differential equation [\[eq:diffeqhk\]](#eq:diffeqhk){reference-type="eqref" reference="eq:diffeqhk"} gives \[0 = \underbrace{\widetilde{a}_k(x_{\min})}_{>0} \underbrace{h_k''(x_{\min})}_{\geq0} + \widetilde{b}_k(x_{\min}) \underbrace{h_k'(x_{\min})}_{=0} + (2k-1) \underbrace{\widetilde{c}_k(x_{\min})}_{>0} \underbrace{h_k(x_{\min})}_{>0} > 0,\] which is a contradiction. | {'timestamp': '2022-06-14T02:17:43', 'yymm': '2206', 'arxiv_id': '2206.01591', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01591'} |
null | null |
# Introduction
Since 1970 the Bass-Serre theory of groups acting on trees stood out as one of the major advances in classical combinatorial group theory. The main notion of the Bass-Serre theory is the notion of graph of groups. The fundamental group of a graph of groups acts naturally on a standard (universal) tree that allows to describe subgroups of these constructions. This theory raised naturally the question of accessibility: namely whether we can continue to split \(G\) into an amalgamated free product or an HNN-extension forever, or do we reach the situation, after finitely many steps, where we can not split it anymore. In other words accessibility is the question whether splittings of \(G\) as the fundamental group of a graph of groups have natural bound. Accessibility of splittings over finite groups (i.e., as a graph of groups with finite edge groups) was studied by Dunwoody ( and ) who proved that finitely presented groups are accessible but found an example of an inaccessible finitely generated group. This initiated naturally a search for a kind of accessibility that holds for finitely generated groups. The breakthrough in this direction is due to Sela who proved \(k\)-acylindrical accessiblity for any finitely generated group: accessibility provided the stabilizer of any segment of length \(k\) of the group acting on its standard tree is trivial for some \(k\). The profinite version of Bass-Serre theory was developed by Luis Ribes, Oleg Melnikov and the second author. However the pro-\(p\) version of Bass-Serre theory does not give subgroup structure theorems the way it does in the classical Bass-Serre theory: even in the pro-\(p\) case, if \(G\) acts on a pro-\(p\) tree \(T\) then a maximal subtree of the quotient graph \(G\backslash T\) does not always exist and even if it exists it does not always lift to \(T\). Nevertheless, the pro-\(p\) version of the subgroups structure theorem works for pro-\(p\) groups acting on a pro-\(p\) trees that are accessible with respect to splitting over edge stabilizers; see. This shows additional importance of studying accessibility of pro-\(p\) groups. In general finitely generated pro-\(p\) groups are not accessible, as shown by G. Wilkes, and it is an open question whether finitely presented are. Our main result in the direction is the pro-\(p\) version of the celebrated Sela's result (cf. Theorem [\[k-acylindrical accessibility\]](#k-acylindrical accessibility){reference-type="ref" reference="k-acylindrical accessibility"}).
We use our accessibility theorem to establish the Kropholler type JSJ-decomposition for Poincaré duality pro-\(p\) groups. JSJ decompositions first appeared in 3-dimensional topology with the theory of the characteristic submanifold by Jaco-Shalen and Johannson. These topological ideas were carried over to group theory first by Kropholler for some Poincaré duality groups. Later constructions of JSJ decompositions were given in various settings by Sela for torsion-free hyperbolic groups, and in various settings by Rips-Sela, Bowditch, Dunwoody-Sageev, Fujiwara-Papasoglu, Dunwoody-Swenson\... ). This has had a vast influence and range of applications in geometric and combinatorial group theory. The result below can be considered as the first step towards this theory in the category of pro-\(p\) groups. We establish a canonical JSJ-decomposition of Poincaré duality pro-\(p\) groups of dimension \(n\) (i.e., \(PD^n\) pro-\(p\) groups) which is a pro-\(p\) version of the Kropholler. It also can be viewed as a pro-\(p\) version of the torus decomposition theorem for 3-manifolds (cf. Theorem [\[thm:JSJsection\]](#thm:JSJsection){reference-type="ref" reference="thm:JSJsection"}).
Examples of JSJ-decompositions of \(PD^3\) pro-\(p\) groups can be obtained by the pro-\(p\) completion of abstract JSJ-decomposition of some 3-manifolds (see ). The pro-\(p\) completion of \(PD^n\) groups in general were studied in . The structure of the paper is as follows. Section 2 recalls the notions of a pro-\(p\) tree, a pro-\(p\) fundamental group and a graph of pro-\(p\) groups with a special focus on finite graphs of pro-\(p\) groups. Throughout the paper finite graphs of pro-\(p\) groups will be often required to be reduced and proper (see Definitions [\[proper\]](#proper){reference-type="ref" reference="proper"} and [\[reduced\]](#reduced){reference-type="ref" reference="reduced"}) but Remarks [\[remark proper\]](#remark proper){reference-type="ref" reference="remark proper"} and [\[reduced-2\]](#reduced-2){reference-type="ref" reference="reduced-2"} show that such an assumption is not restrictive. Section 3 is devoted to the proof of the pro-\(p\) version of Sela's accessibility which states that every finitely generated pro-\(p\) group is \(k\)-acylindrically accessible. Recall that a profinite graph of pro-\(p\) groups \((\G,\Gamma)\) is *\(k\)-acylindrical* if the action of the pro-\(p\) fundamental group on its standard pro-\(p\) tree is \(k\)-acylindrical (cf. Subsection [2.2](#pro-p tree){reference-type="ref" reference="pro-p tree"}). In this section we also prove the pro-\(p\) version of Karras-Solitar result describing 2-generated subgroups of free products with malnormal amalgamation (see Theorem [\[thm:KS\]](#thm:KS){reference-type="ref" reference="thm:KS"}). Finally Section 4 deals with splittings of \(PD^n\) pro-\(p\) groups and culminates with a JSJ-decomposition for \(PD^n\) pro-\(p\) groups (see Theorem [\[thm:JSJ\]](#thm:JSJ){reference-type="ref" reference="thm:JSJ"}) which is a pro-\(p\) version of the Kropholler. Note that the Kropholler theorem gives also information on vertex groups of a JSJ-splitting that is based on the Kropholler-Roller decomposition theorem that states that a \(PD^n\) group \(G\) having a \(PD^{n-1}\) subgroup \(H\) virtually splits as a free product with amalgamation or HNN-extension over a subgroup commensurable with it if \(cd(H\cap H^g)\neq n-2\) for each \(g\in G\). In fact, by , \(G\) virtually splits over \(H\) if \(H\) is polycyclic. Unfortunately Kropholler-Roller theorems do not hold in the prop-\(p\) case as shown by the following example, which has been constructed in communication with Peter Kropholler during the visit of the second author to the University of Southampton.
In Section 5 we provide the details of the statement written in the example above. Here we just remark that the absence of the Kroholler-Roller splitting result is an obstacle of obtaining information on vertex groups of a JSJ-splitting from Theorem [\[thm:JSJ\]](#thm:JSJ){reference-type="ref" reference="thm:JSJ"}.
# Notation, definitions and basic results {#preliminaries}
We shall denote by \(d(G)\) the number of a minimal set of generators of a pro-\(p\) group \(G\) and by \(\Phi(G)\) its Frattini subgroup. If a pro-\(p\) group \(G\) continuously acts on a profinite space \(X\) we denote by \(G_x\) the stabilizer of \(x\) in \(G\). If \(x\in X\) and \(g\in G\), then \(G_{gx}=gG_xg^{-1}\). We shall use the notation \(h^g=g^{-1} hg\) for conjugation. For a subgroup \(H\) of \(G\), \(H^G\) will stand for the (topological) normal closure of \(H\) in \(G\). If \(G\) is an abstract group \(\widehat G\) will mean the pro-\(p\) completion of \(G\). Throughout the paper, unless otherwise stated, groups are pro-\(p\), subgroups will be closed and morphisms will be continuous. Finite graphs of groups will be proper and reduced (see Definitions [\[proper\]](#proper){reference-type="ref" reference="proper"} and [\[reduced\]](#reduced){reference-type="ref" reference="reduced"}). Actions of a pro-\(p\) group \(G\) on a profinite graph \(\Gamma\) will a priori be supposed to be faithful (i.e., the action has no kernel), unless we consider actions on subgraphs of \(\Gamma\). Next we collect basic definitions, following.
## Profinite graphs
A connected finite graph without circuits is called a *tree*. In the next subsection we shall explain how this notion extends to the pro-\(p\) context. At the moment we shall prove several easy lemmas on finite graphs needed in the paper. The *valency* of a vertex is the number of edges connected to it. Hence, a vertex is pending if it has valency 1. A tree with two pending vertices will be called a *line*.
## Pro-\(p\) trees {#pro-p tree}
If \(T\) is a pro-\(p\) tree, then we say that a pro-\(p\) group \(G\) *acts on \(T\)* if it acts continuously on \(T\) and the action commutes with \(d_0\) and \(d_1\). If \(t\in V(T)\cup E(T)\) we denote by \(G_t\) the stabilizer of \(t\) in \(G\). For a pro-\(p\) group \(G\) acting on a pro-\(p\) tree \(T\) we let \(\tilde{G}\) denote the subgroup generated by all vertex stabilizers. Moreover, for any two vertices \(v\) and \(w\) of \(T\) we let \([v,w]\) denote the geodesic connecting \(v\) to \(w\) in \(T\), i.e., the (unique) smallest pro-\(p\) subtree of \(T\) that contains \(v\) and \(w\). The fundamental group \(\pi_1(\Gamma)\) acts freely on a pro-\(p\) tree \(\widetilde \Gamma\) (universal cover) such that \(\pi_1(\Gamma)\backslash \tilde\Gamma=\Gamma\) (see or for details). An action of a pro-\(p\) group on a pro-\(p\) tree \(T\) is called *\(k\)-acylindrical* if the stabiliser of any geodesic in \(T\) of length greater than \(k\) is trivial. For instance, \(0\)-acylindrical refers to an action with trivial edge stabilisers, and \(1\)-acylindrical implies that edge stabilisers are malnormal in vertex-groups.
## Finite graphs of pro-\(p\) groups
In this subsection we recall the definition of a finite graph of pro-\(p\) groups \((\G,\Gamma)\) and its fundamental pro-\(p\) group \(\Pi_1(\G, \Gamma)\). When we say that \({\cal G}\) is a finite graph of pro-\(p\) groups we mean that it contains the data of the underlying finite graph, the edge pro-\(p\) groups, the vertex pro-\(p\) groups and the attaching continuous maps. More precisely,
The main examples of \(\Pi_1(\G,\Gamma)\) are an amalgamated free pro-\(p\) product \(G_1\amalg_H G_2\) and an HNN-extension \({\rm HNN}(G,H,t)\) that correspond to the cases of \(\Gamma\) having one edge and either two vertices or only one vertex, respectively.
If \((\G,\Gamma)\) is a finite graph of finitely generated pro-\(p\) groups, then by a theorem of J-P. Serre (stating that every finite index subgroup of a finitely generated pro-\(p\) group is open, cf. ) the fundamental pro-\(p\) group \(G=\Pi_1(\G,\Gamma)\) of \((\G,\Gamma)\) is the pro-\(p\) completion of the usual fundamental group \(\pi_1(\G,\Gamma)\) (cf. ). Note that \((\G,\Gamma)\) is proper if and only if \(\pi_1(\G,\Gamma)\) is residually \(p\). In particular, edge and vertex groups will be subgroups of \(\Pi_1(\G,\Gamma)\).
Let \((\G, \Gamma)\) be a profinite graph of pro-\(p\) groups and \(\Delta\) a subgraph of \(\Gamma\). Then by \((\G, \Delta)\) we shall denote the graph of groups restricted to \(\Delta\). We shall often use the following
# Acylindrical accessibility
In this section we shall prove a pro-\(p\) version of Sela's accessibility. Note that Sela used \({\mathbb R}\)-trees for the proof; later Weidmann found another proof using Nielsen method and established a bound. Both methods are not available in the pro-\(p\) case. We shall start with two auxiliary results on free amalgamated product and its generalization for abstract groups.
We finish the section with a pro-\(p\) version of the Karras-Solitar but we start with the lemma below where generation symbols \(\langle\rangle\) mean abstract generation unlike in the rest of the paper where \(\langle\rangle\) means topological generation.
# Decomposing \(PD^n\) pro-\(p\) groups
## Pro-\(p\) \(PD^n\)-pairs
In Wilkes defined the profinite version of group pairs but we shall need only a simple version of it. A pro-\(p\) group pair \((G,\mathcal S)\) consists of a pro-\(p\) group \(G\) and a finite family \(\mathcal S\) of closed subgroups \(S_x\) of \(G\) indexed over a set (we allow repetitions in this family). Given a closed subgroup \(H\) of \(G\), let \(\mathcal S^H\) denote the family of subgroups \[\label{eq:SH} \{H\cap \sigma(y)S_x\sigma(y)^{-1}\mid x\in X, y\in H\backslash G/S_x\},\] indexed over \[H\backslash G/\mathcal S:=\bigsqcup_{x\in X} H\backslash G/S_x,\] where \(\sigma\colon G/H\to G\) is a section of the quotient map \(G\to G/H\)[^2]. In the author develops the theory of the cohomology of a profinite group relative to a collection of closed subgroups and defines profinite Poincaré duality pairs (or \(PD^n\)-pairs for short) and the reader is referred to for rigorous definitions and basic results. A pro-\(p\) group pair \((G,\mathcal S)\) is a pro-\(p\) \(PD^n\)-pair, for some \(n\in\mathbb N\), if the double of \(G\) over the groups in \(\mathcal S\) is a pro-\(p\) \(PD^n\)-group. Here the double of \(G\) over \(\mathcal S\) refers to the fundamental group of a graph of groups with two vertices and \(|\mathcal S|\) edges where a copy of \(G\) is over each vertex and groups of \(\mathcal S\) are over the edges, with natural boudary maps.
We say that a pro-\(p\) \(PD^n\)-pair \((G,\mathcal S)\) splits as an amalgamated free pro-\(p\) product \(G=G_1\amalg_H G_2\) (resp. as HNN-extension \(HNN(G_1,H,t)\)) if each \(S_i\) is conjugate to either \(G_1\) or \(G_2\) (resp. \(G_1\)). The next proposition was communicated to us by G. Wilkes.
The proof relies on the following
## Splitting over polycyclic subgroups
Here we collect some results that will be used later in the proof of the main theorem. We say that a pro-\(p\) group \(G\) admits a *\(k\)-acylidrical splitting* if \(G\) is isomorphic to the fundamental pro-\(p\) group \(\Pi_1(\mathcal G,\Gamma)\) of a \(k\)-acylindrical proper reduced finite graph of pro-\(p\) groups.
The next theorem establishes a pro-\(p\) version JSJ-decomposition for \(PD^n\) pro-\(p\) groups analogous of one from.
The uniqueness of \(\mathcal T\) in Theorem [\[thm:JSJ\]](#thm:JSJ){reference-type="ref" reference="thm:JSJ"} induces an action on it by the automorphism group \(Aut(G)\). This gives a splitting sturcture on \(Aut(G)\) if \(\mathcal T\) is non-trivial. We state this as a
# Example
The next example shows that both Kropholler-Roller theorems do not hold in the pro-\(p\) case, i.e. neither of the statements of the theorems above.
[^1]: It is called a *quasimorphism* in.
[^2]: A different section only affects the family \(\mathcal S^H\) by changing its members by conjugacy in \(H\). | {'timestamp': '2022-06-06T02:14:33', 'yymm': '2206', 'arxiv_id': '2206.01569', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01569'} |
null | null |
# Introduction
[\[Introduction\]]{#Introduction label="Introduction"} Turbulent premixed combustion is characterized by non-linear and multi-scale interactions between chemistry, molecular transport, and fluid motion, leading to modification of both the flame structure and turbulence . Coupling between chemical energy release and kinetic energy occurs due to thermal expansion (dilatation), typically at length scales associated with the flame (*i.e.* the laminar flame thermal thickness \(\delta_\mathrm{L}^0\)). Increasing evidence suggests that the presence of the flame can lead to qualitative deviations from the direct forward energy cascade of constant density turbulence at scales around \(\delta_\mathrm{L}^0\), potentially resulting in mean up-scale kinetic energy transfer (back-scatter) within the flame. Using direct numerical simulations (DNS) of planar flames in homogeneous isotropic turbulence, Towery et al. observed reversal of advective energy transfer at scales around the laminar flame thickness. O'Brien identified combustion-induced mean back-scatter when data were conditioned on the reaction progress variable \(c\). Kim et al. reported an increase in the kinetic energy at small scales inside the flame, with the small scales becoming the donor of energy to larger scales. Experimentally, we recently quantified mean back-scatter in swirl-stabilized premixed flames across a range of Karlovitz numbers, up to \(\mathrm{Ka} = 50\). These experiments confirmed the DNS observations and also demonstrated that mean back-scatter can occur in practical configurations. While our previous study investigated the influence of the local turbulence on cross-scale energy transfer via the local strain-rate field, we did not investigate the role of large-scale coherent flow structures; swirl flames often exhibit such structures, e.g. helical shear layer vortices . These flow structures contain a significant portion of the fluctuating kinetic energy and are responsible for large-scale stirring (and hence flame corrugations, increased reaction rate, etc.). Consequently, there is a potentially complicated coupling between large-scale structures, heat release rate, and kinetic energy dynamics. Indeed, even without a flame, the presence of coherent structures may disrupt the local equilibrium between turbulence generation by the mean flow and dissipation at smaller scales . In this paper, we extend our previous analysis to consider the simultaneous influence of large-scale flow structures and the flame on the kinetic energy cascade at scales around \(\delta_\mathrm{L}^0\) in a premixed swirl flame. Following the physical-space analysis of O'Brien et al. , the transport equation for the kinetic energy of the filtered flow (\(k \equiv \frac{1}{2} \Tilde{u}_i \Tilde{u}_i\)) is \[\frac{\partial k}{\partial t} + \Tilde{u}_i \frac{\partial k}{\partial x_i} = \underbrace{-\frac{\Tilde{u}_i}{\Bar{\rho}} \frac{\partial \Bar{p}}{\partial x_i}}_{\alpha_\mathrm{p}} + \underbrace{\frac{\Tilde{u}_i}{\Bar{\rho}} \frac{\partial \Bar{\tau}_{ij}}{\partial x_j}}_{\alpha_\nu} \underbrace{-\frac{\Tilde{u}_i}{\Bar{\rho}} \frac{\partial \mathcal{T}_{ij}}{\partial x_j}}_{{\asfs}}, \label{e:k}\] where \(\alpha_\mathrm{p}\), \(\alpha_\nu\), and represent the contributions to resolved kinetic energy through work by resolved pressure, resolved viscous stresses, and sub-filter-scale stresses, respectively. The rate of kinetic energy transfer across the filter scale is identified as \[{\asfs} =-\frac{\Tilde{u}_i}{\Bar{\rho}} \frac{\partial \mathcal{T}_{ij}}{\partial x_j}. \label{e:asfs}\] \({\asfs}>0\) represents the local transfer of kinetic energy from scales smaller than \(\Delta\) to scales larger than \(\Delta\) (*i.e.* back-scatter); downscale energy transfer (*i.e.* forward-scatter) is indicated by \({\asfs}<0\). also appears in the equation for the sub-filter scale (SFS) kinetic energy (\(k_\mathrm{sfs} \equiv \frac{1}{2}(\widetilde{u_i u_i}-\Tilde{u}_i \Tilde{u}_i)\)) but with the opposite sign, demonstrating the two-way coupling between the resolved and SFS kinetic energies . Hence, this study experimentally evaluates the simultaneous effects of heat release and large-scale vortex structures on \({\asfs}\) in a turbulent premixed swirl flame using tomographic particle image velocimetry (TPIV) and formaldehyde () planar laser induced fluorescence (PLIF). Large-scale vortices are identified using the -criterion and proper orthogonal decomposition (POD) is used to isolate a coherent periodic vortex. We then analyze the conditional statistics of , , and the flame, both on an instantaneous basis and resolved with respect to the periodic vortex motion. We find that vortex-flame interactions lead to increased back-scatter compared to other regions of the flame.
# Experimental Setup
The experimental setup is identical to that described in Refs. and only a brief summary is given here. The model gas turbine combustor (Fig. [\[f:setup\]](#f:setup){reference-type="ref" reference="f:setup"}(a)) was idential to that originally described by Meier et al. , but was operated without the combustion chamber to prevent window contamination from flow tracer particles (see below). The combustor features a 12-vane radial swirler and a 27.85 mm diameter nozzle with a conical bluff body along the centerline. The measured swirl number is \(S=0.55\) at the exit plane of the nozzle. Flow rates of ambient temperature methane and air were controlled using electromechanical mass flow controllers (Brooks, 1% full-scale uncertainty) and premixed far upstream of the combustor plenum. Two test cases are considered here: a premixed flame at an equivalence ratio of \(\phi=0.85\) and bulk flow rate of \(U=25\) m/s, and a non-reacting air flow at the same bulk flow rate (\(U\) is the ratio of the total volumetric flow rate and the nozzle exit area). The corresponding turbulence Reynolds number, Karloviz number, and Damköhler number were \(\mathrm{Re_T}=u^\prime\ell/\nu_\mathrm{r} = 3300\), \(\mathrm{Ka} = (u^\prime/s_\mathrm{L}^0)^{3/2}(\delta_\mathrm{L}^0/\ell)^{1/2} = 50\), and \(\mathrm{Da}=\big(\mathrm{Re_T}/\mathrm{Ka}^2 \big)^{1/2} = 1.1\). Here, \(\nu_\mathrm{r}\) is the kinematic viscosity in the reactants, \(u^\prime\) is the root-mean-squared velocity fluctuations in the shear layer between the inflowing reactants and central recirculation zone for the non-reacting flow and \(\ell\) is the integral length scale, which was taken to be the full-width-half-maximum thickness of the shear layer. The laminar flame speed \(s_\mathrm{L}^0\) and flame thickness \(\delta_\mathrm{L}^0\), were calculated using the freely propagating flame model in Cantera with GRI3.0 chemical mechanism . Measurements were made of the 3D velocity field and planar distribution using simultaneous TPIV and PLIF, respectively. The TPIV measurement region was a \(16 \times 16 \times 1.3\) mm rectangular volume located directly above the nozzle, focused on the one branch of the axisymmetric swirl flame (see Fig. [\[f:setup\]](#f:setup){reference-type="ref" reference="f:setup"}(a)). The PLIF field-of-view was along the centerline plane of the TPIV volume. The diagnostics configuration is shown in Fig. [\[f:setup\]](#f:setup){reference-type="ref" reference="f:setup"}(b). A total of 1,000 simultaneous TPIV and PLIF measurements were made at each test condition. For the TPIV, a series of lenses and two knife blades were used to shape the output of a dual-head Nd:YAG laser (Quantel Evergreen 200, 532 nm, 10 Hz, 200 mJ/pulse, \(3~\mu\)s between pulses) into a \(1.3\) mm thick collimated rectangular slab. The air flow was seeded with alumina particles having a nominal diameter of \(0.3\) prior to mixing with the methane. The Stokes number of particles -- based on a conservative estimate of the Kolmogorov time scale -- was below \(0.03\). Mie scattering from the seed particles was captured by four sCMOS cameras (Andor Zyla 5.5, \(2048\times2048\) pixels, 6.5 pixel pitch), which were positioned on either side of the laser sheet at angles between \(20^\circ\) and \(30^\circ\) relative to laser propagation direction. A long-distance microscope (Infinity K2, CF-1/B objectives, \(f/ \# =38\)) was mounted to each camera via a Scheimpflug adapter (LaVision) to enable off-axis imaging. Commercial software (LaVision DaVis 8.4) was used to reconstruct the particle tomograms and compute the 3D velocity vectors using direct spatial correlation. The final velocity resolution was 250 with a 50% vector overlap (125 vector spacing). As described in our previous work , significant effort was made to optimize the spatial resolution and accuracy of the TPIV diagnostics; the results presented are converged with respect to the interrogation box size, indicating that the measurements are well resolved. The PLIF system consisted of a sCMOS camera (Andor NEO 5.5, \(2048\times2048\) pixels, 6.5 pixel pitch) with a macro lens (Tamron, \(f=180\) mm, \(f/ \# =2.8\)), image intensifier (LaVision IRO, gate = 100 ns), and an Nd:YAG laser (Quanta-Ray INDI-40-10, 355 nm, 10 Hz, 70 mJ/pulse). The laser sheet was formed using a set of cylindrical lenses and transmitted along the center of the TPIV volume (*i.e.* \(z=0\) mm plane). The laser excited the \(A-X_0^1\) transition of around \(355\) nm; fluoresence in the range of \(370-480\) nm was isolated using a specialized filter (Semrock FF01-CH2O-50.8-D).
# Processing and data analysis
## Determination of
Determination of in this work uses the same methodology as our previous paper, which also includes a detailed assessment of the robustness and uncertainty of the method . While there are several approximations necessary in our empirical calculation, this previous study demonstrated that the results are quantitatively robust to these approximations. A brief summary of the methodology is provided here for completeness. Calculation of requires filtering the instantaneous velocity and density fields. The 3D velocity fields are provided by the TPIV and we estimate the 2D density field from the PLIF. In methane/air flames, is formed at early stages of fuel breakdown in a relatively inert "preaheat zone" and rapidly consumed during formation in the thin region of rapid exothermic reactions; serves as a good indicator of the instantaneous thermal width \(\delta_\mathrm{L}^0\). Here, we use the -containing regions to identify the instantaneous flame boundaries. We then define a progress-variable-like parameter \(c=(d-d_r)/(d_p-d_r)\), where \(d_r\) and \(d_p\) are the shortest in-plane distances to the reactant and product facing boundaries of the -containing region; our \(c\) maps the distance of a location through the instantaneous flame surface. The local density and temperature are then approximated to be the same as a laminar flame with the same reactants at the corresponding \(c\). We highlight and acknowledge that the -to-\(\rho\) mapping is not exact. However, we have demonstrated that the calculation of --the primary quantity of interest--does not depend on the detailed density distribution due to the spatial filtering inherent in the analysis. Indeed, even a simple bimodal density approximation (jumping between the density of the reactants and products) provides qualitatively equivalent results. Calculating via Eq. [\[e:asfs\]](#e:asfs){reference-type="ref" reference="e:asfs"} requires filtered velocity and density fields, along with their spatial derivatives. We consider a top-hat filter with \(\Delta/\delta_\mathrm{L}^0 \approx 1.5\) (\(\Delta = 6\) vectors), capturing the cross-scale transfer of kinetic energy between the flame scales and the larger scales. While the 3D measured velocity fields enable computation of the required derivatives, density is only available in the central TPIV plane. Hence, we restrict our analysis to this plane and use conservation of mass to approximate the out-of-plane density gradient needed for computation of as \[\begin{gathered}
\frac{\partial \bar{\rho}}{\partial x_3} = \frac{1}{\Tilde{u}_3} \left[-\bar{\rho} \left(\frac{\partial \Tilde{u}_1}{\partial x_1} + \frac{\partial \Tilde{u}_2}{\partial x_2} + \frac{\partial \Tilde{u}_3}{\partial x_3} \right)\right.\\-\left.\Tilde{u}_1 \frac{\partial \bar{\rho}}{\partial x_1}-\Tilde{u}_2 \frac{\partial \bar{\rho}}{\partial x_2}-\frac{\partial \bar{\rho}}{\partial t}\right] \label{eq:drho_dz} \end{gathered}\] We present our results in terms of the ensemble averaged (denoted \(\langle \asfs \rangle\)) and hence the contribution from the unsteady density term is negligible in the statistically stationary turbulent flow studied here.
## Identification of coherent structures
We use the -criterion to identify coherent swirling eddies in this variable density flow; eddies are spatial regions in which the velocity gradient tensor has a complex eigenvalue pair (\(\lambda_\mathrm{cr} \pm i{\lci}\)). The imaginary part of the complex eigenvalue pair\--quantifies the strength of the local swirling motion. Physically, the local flow is either stretched or compressed along the axis of the real eigenvector (\(\boldsymbol{\nu_r}\)), while the flow is swirling in the plane spanned by the real and imaginary parts of the complex eigenvector (\(\boldsymbol{\nu_{cr}}\) and \(\boldsymbol{\nu_{ci}}\)). We use fields to identify the position and size of the largest scale coherent structures in the flow. Hence, for the calculation, we first spatially smooth the velocity fields using a 21 vector box filter. This smoothed velocity field retains large scale eddies while suppressing smaller structures. To ensure that this smoothing does not affect our conclusions, we re-performed the analysis using 11-vector and 6-vector filters. While the magnitude of peak decreased with increasing filter size, the largest eddies retained their size and shape. The eigenvalues and eigenvectors were computed on the central plane of the TPIV volume. Locations with three real eigenvalues were assigned a value of \({\lci}=0\) (*i.e.* no swirl); the magnitude of was retained at all other locations to quantify the swirl. The data were further reduced by eliminating locations where the axis of rotation was predominantly in the \(x\)-\(y\) plane. To demonstrate the distribution of rotation axis orientations, we consider the normal vector to the plane of rotation, defined as \(\boldsymbol{\Hat{n}}=\boldsymbol{\nu_\mathrm{cr}} \times \boldsymbol{\nu_\mathrm{ci}}\). Figure [\[f:pdf_vecnorm\]](#f:pdf_vecnorm){reference-type="ref" reference="f:pdf_vecnorm"} shows the joint probablity density function (PDF) of \(\boldsymbol{\Hat{n}_z}\) and the polar angle \(\phi\) of \(\boldsymbol{\Hat{n}}\) in the \(x\)-\(y\) plane. The locations with \(\phi \approx 0-25\) degrees were found to correspond to the large scale swirl induced by the radial swirl vanes. However, the locations with large \(\boldsymbol{\Hat{n}_z}\) correspond to coherent structures that rotate in plane. We therefore set \({\lci}=0\) at locations with \(\boldsymbol{\Hat{n}_z} < 1/\sqrt{2}\) to isolate the structures of interest.
## Proper orthogonal decomposition
We use the method of snapshots POD to extract the dominant coherent structures, from which we phase-sort individual measurements based on phase angle of the structure motion. Given a set of \(N\) measurements of the variable \(\begin{equation} \end{equation} where\)\_j()\(and\)a_j(t)\(represent the spatial modes and their temporal coefficients, respectively. POD modes\)\_j\(are given by the eigenvectors of the covariance matrix\) = XX\^T\(, where\)X\(represents the\)N\(snapshots in a matrix form. The eigenvalues (\)\_j\() are the mode energies, which describe how much of the variance in the original data is contained in a particular mode. We use the measured 2D swirling strength fields on the center plane of the TPIV volume as the variable\) Previously, the dominant dynamic flow structure in this burner was identified as either a helical vortex core or a toroidal vortex, depending on the operating conditions. The structure identified here corresponds to one of these shear layer vortex modes, though the exact mode cannot be identified due to the field of view, which only covers one half of the burner. Based on the operating conditions and results in Ref. , we expect this to be a helical mode. The POD on the non-reacting test case showed a similar set of modes and relative energies as the reacting flow, again consistent with the helical modes observed in all non-reacting cases in Ref. . Regardless, the exact nature of the flow structure is not critical for the analysis presented here. We use time coefficients for modes 1 and 2 to obtain phase information of the periodic flow structure. This is done by phase-sorting individual measurements into nine equal "bins" based on the angle of the measurement in the \(a_1(t)\) and \(a_2(t)\) plane, e.g. . Phase-averaged quantities of interest \(q\) are calculated as \[\langle q|\theta_i \rangle = \frac{1}{N_{\theta_i}} \sum_{k=1}^{N_{\theta_i}} q(t_k), \label{e:ph_avg}\] where \(\theta_i\) is the \(i\)-th range of phase angles and \(N_{\theta_i}\) is the ensemble size. We consider , , and \(c\) as the primary quantities of interest in the analysis below. Figure [\[f:modes\]](#f:modes){reference-type="ref" reference="f:modes"} shows the profiles of \(\langle {\lci} | \theta_i \rangle\) at different phases of propagation in a reacting flow. Every second phase is shown for brevity and the solid lines denote the phase-conditioned flame brush, i.e. \(0.1<\langle c|\theta_i\rangle<0.9\). The vortex moves downstream along the inner shear layer between the flame and the recirculating products. The size of the vortex grows and the peak decreases due to diffusion and flow dilatation as the vortex moves downstream. The structure and shape of the flame brush also changes in response to the presence of the vortex.
# Results and Discussion
## Instantaneous Data and Conditional Statistics
We first examine the relationship between , the flame, and on an instantaneous basis. Figure [\[f:single_snap\]](#f:single_snap){reference-type="ref" reference="f:single_snap"}(a) shows a representative snapshot of , overlaid with \(c=0.1\) and \(c=0.9\) contours as markers of the flame. A set of large-scale coherent swirling structures (eddies) intersects with the instantaneous flame. The flame generally is thicker and more curved at locations of flame/vortex interaction. Figure [\[f:single_snap\]](#f:single_snap){reference-type="ref" reference="f:single_snap"}(b) shows the corresponding field with contours of the flame and eddies shown in magenta and green, respectively. Energy transfer occurs in both the up-scale (back-scatter, red) and down-scale (forward-scatter, blue) directions across the filter scale \(\Delta\), with regions of energy back-scatter predominantly occur in the vicinity of the flame. The magnitude of is substantially lower in the products due to suppression of small-scale motion as a result of flow dilatation and increasing viscosity across the flame. Since it is difficult to ascertain correspondence between the eddies and cross-scale energy transfer from this individual image, the conditional PDFs of are plotted in Fig. [\[f:single_snap_cond\]](#f:single_snap_cond){reference-type="ref" reference="f:single_snap_cond"}(a); the PDFs are conditioned on moderate swirl (\(1000<\lci<2000\)) and the reacting data are conditioned on the pure reactants (\(c=0\)), internal to the flame (\(0<c<1\)) and in the burnt products (\(c = 1\)). In the reactants and in non-reacting flow, the PDFs are nearly symmetric, indicating nearly equal forward-and back-scatter across \(\Delta\). Similarly, the PDF in the products is nearly symmetric, but with reduced range of due to attenuation of turbulence at elevated temperature. Inside the flame, the PDF is positively skewed, indicating an increased probability of back-scatter. To further articulate the simultaneous effects of the flame and eddies on , Fig. [\[f:single_snap_cond\]](#f:single_snap_cond){reference-type="ref" reference="f:single_snap_cond"}(b) shows the conditional mean \(\langle\asfs|{\lci}, c\rangle\) for both the reacting and non-reacting flows (the PDFs in Fig. [\[f:single_snap_cond\]](#f:single_snap_cond){reference-type="ref" reference="f:single_snap_cond"}(a) are from the data comprising the mean at \(\lci = 1500\) in Fig. [\[f:single_snap_cond\]](#f:single_snap_cond){reference-type="ref" reference="f:single_snap_cond"}(b)). The error bars indicate the expected uncertainty in the mean based on the sample standard deviation and number of samples, approximating the statistics as normal. For the non-reacting case, \(\langle\asfs\rangle\) is negative for most values of , obtaining a slightly positive value at locations of high swirl. The data in the reactants and products of the reacting case are qualitatively similar to the non-reacting case. Most interesting are the data within the flame, which clearly demonstrate the flame-induced mean back-scatter and that the back-scatter magnitude increases with the local swirling strength. Regions of locally high swirling strength are expected to correspond to local increases in reaction rate due to flame wrinkling and increased scalar gradients. Hence, the data are indicative of a positive correlation between local heat release rate and back-scatter.
## Phase-Conditioned Analysis
The above results demonstrate how the presence of swirling eddies and flame affect the local cross-scale kinetic energy transfer. We now isolate the impact of the single most dominant coherent structure on the energy dynamics, namely the periodic vortex identified from the POD (see Fig. [\[f:modes\]](#f:modes){reference-type="ref" reference="f:modes"}). To do so, we examine the phase-averaged , , and \(c\) profiles generated using POD modes 1 and 2 in the reacting and non-reacting flows. Phase averaging is performed based on Eq. [\[e:ph_avg\]](#e:ph_avg){reference-type="ref" reference="e:ph_avg"}, and we further condition on specific spatial regions of interest. To demonstrate this, Fig. [\[f:regions\]](#f:regions){reference-type="ref" reference="f:regions"} shows an instantaneous schematic of and \(c\) in phase 7, showing the location of the phase-averaged vortex (\(\langle \lci | \theta_i \rangle\)-red contour), the location of the instantaneous vortex (-green contour), and the region occupied by the flame (\(0<c<1\)-magenta contour). The "\(\mathrm{flame}~\cap~\mathrm{vortex}\)" region (shaded green) represents the region where the flame, , and \(\langle \lci | \theta_i \rangle\) intersect; the instantaneous vortex that lies outside the \(\langle \lci | \theta_i \rangle\) contour is not part of the periodic vortex and, hence, is not considered in this analysis. The "\(\mathrm{flame}~\cap~\overline{\mathrm{vortex}}\)" region (shaded purple) represents the locations inside the flame but external to the instantaneous vortex inside the \(\langle \lci | \theta_i \rangle\) contour. We also consider "\(\mathrm{vortex}\)" and "\(\overline{\mathrm{vortex}}\)" which represent overlapping regions of and \(\langle \lci | \theta_i \rangle\) and its complement set, respectively. Finally, the "\(\mathrm{reactants}\)" region (shaded light blue) represents regions of \(c=0\). We do not present data in the products (shaded orange), as generally is low in this region due to the attenuated turbulence. Figure [\[f:modes_avg\]](#f:modes_avg){reference-type="ref" reference="f:modes_avg"}(a) shows the phase-averaged in different regions of the flow as a function of phase. In the reactants and in the non-reacting flow, \(\langle \asfs | \theta_i\rangle\) is negative and no significant phase dependence is observed. For the non-reacting flow, the increased scatter for regions containing the periodic vortex likely is attributed to uncertainty and convergence. Broadly speaking, the non-reacting flow and non-reacting regions of the reacting flow both show mean forward-scatter across the filter scale. The regions containing the flame show qualitative differences in compared to the non-reacting flows, with further qualitative differences between regions with and without the dominant periodic vortex. In the "\(\mathrm{flame}~\cap~\overline{\mathrm{vortex}}\)" region, \(\langle \asfs | \theta_i, \mathrm{flame}~\cap~\overline{\mathrm{vortex}}\rangle\) is positive. This is consistent with our previous experimental findings and other DNS studies , demonstrating that the flame leads to mean back-scatter at scales around the laminar flame thickness. Importantly, regions where the flame and periodic vortex intersect exhibit higher mean back-scatter compared to regions of the flame without the vortex present, *i.e.* \(\langle \asfs | \theta_i, \mathrm{flame}~\cap~{\mathrm{vortex}}\rangle> \langle \asfs | \theta_i, \mathrm{flame}~\cap~\overline{\mathrm{vortex}}\rangle\). The magnitude of \(\langle \asfs | \theta_i, \mathrm{flame}~\cap~{\mathrm{vortex}} \rangle\) decreases with phase. Figure [\[f:modes_avg\]](#f:modes_avg){reference-type="ref" reference="f:modes_avg"}(b) shows that the magnitude of the swirl, *i.e.* \(\langle {\lci} | \theta_i, \mathrm{flame}~\cap~\mathrm{vortex} \rangle\), also decreases with phase as the vortex moves downstream. Hence, there is a correspondence between the swirling strength of the dominant periodic vortex and the mean back-scatter magnitude at locations of flame/vortex interaction. This result is consistent with the instantaneous data in Fig. [\[f:single_snap_cond\]](#f:single_snap_cond){reference-type="ref" reference="f:single_snap_cond"}(b) and further articulates the simultaneous and complimentary effects of combustion induced heat release rate and large-scale coherent vortices on kinetic energy back-scatter.
# Conclusion
This study has experimentally investigated the cross-scale energy transfer in turbulent premixed swirl flames in the presence of coherent structures using high-resolution TPIV and PLIF. Coherent structures were identified using the -criterion and POD was performed on -fields to isolate the most dominant periodic coherent vortex structure. Cross-scale energy transfer was computed across a filter scale \(\Delta/\delta_\mathrm{L}^0=1.5\) and conditional statistics of on and the flame were compiled. The main conclusions are that (i) the non-reacting case and non-reacting regions of the the reacting case (pure reactants and products) exhibited mean down-scale transfer of kinetic energy, *i.e.* forward-scatter, with nearly symmetric PDFs about \(\asfs = 0\); (ii) reacting regions of the flow (inside the instantaneous flame surface) exhibited mean up-scale transfer of kinetic energy, *i.e.* back-scatter, with slightly positively skewed PDFs; (iii) the mean magnitude of the back-scatter inside the flame was increased in regions of flame/vortex interaction, as demonstrated both through the instantaneous data and statistics conditioned on the phase of the periodic vortex structure; and (iv) there was a correspondence between the magnitude of the local swirling strength and the mean back-scatter magnitude. It is possible that the observed increased back-scatter in regions of flame/vortex interaction is due to greater local heat release rates, though further measurements would be needed to confirm this. Regardless, this study articulates the complicated influences of the flame and flow/flame interactions on turbulence dynamics around the flame scale, which should be considered in turbulence modeling. | {'timestamp': '2022-06-06T02:16:03', 'yymm': '2206', 'arxiv_id': '2206.01621', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01621'} |
{'timestamp': '2022-06-06T02:16:17', 'yymm': '2206', 'arxiv_id': '2206.01631', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01631'} |
|
null | null |
# Introduction {#sec:intro}
Minkowski summation is a basic and ubiquitous operation on sets. Indeed, the Minkowski sum \(A+B = \{a+b: a \in A, b \in B\}\) of sets \(A\) and \(B\) makes sense as long as \(A\) and \(B\) are subsets of an ambient set in which a closed binary operation denoted by \(+\) is defined. In particular, this notion makes sense in any group, and *additive combinatorics* (which arose out of exploring the additive structure of sets of integers, but then expanded to the consideration of additive structure in more general groups) is a field of mathematics that is preoccupied with studying what exactly this operation does in a quantitative way. "Sumset estimates" are a collection of inequalities developed in additive combinatorics that provide bounds on the cardinality of sumsets of finite sets in a group. In this paper, we use \(\#(A)\) to denote the cardinality of a countable set \(A\), and \(|A|\) to denote the volume (i.e., \(n\)-dimensional Lebesgue measure) of \(A\) when \(A\) is a measurable subset of \({\bf R}^n\). The simplest sumset estimate is the two-sided inequality \(\#(A) \#(B) \geq \#(A+B) \geq \#(A)+\#(B)-1\), which holds for finite subsets \(A, B\) of the integers; equality in the second inequality holds only for arithmetic progressions. A much more sophisticated sumset estimate is Kneser's theorem (cf.,, ), which asserts that for finite, nonempty subsets \(A, B\) in any abelian group \(G\), \(\#(A+B) \geq \#(A+H) +\#(B+H)-\#(H)\), where \(H\) is the stabilizer of \(A+B\), i.e., \(H=\{g\in G: A+B+g=A+B\}\). Kneser's theorem contains, for example, the Cauchy-Davenport inequality that provides a sharp lower bound on sumset cardinality in \(\mathbb{Z}/p\mathbb{Z}\). In the reverse direction of finding upper bounds on cardinality of sumsets, there are the so-called Plünnecke-Ruzsa inequalities. One example of the latter states that if \(\#(A+B)\leq \alpha \# A\), then \(\#(A+k\cdot B)\leq \alpha^k \# A\), where \(k\cdot B\) refers to the sum of \(k\) copies of \(B\). Such sumset estimates form an essential part of the toolkit of additive combinatorics. In the context of the Euclidean space \({\mathbb R}^n\), inequalities for the volume of Minkowski sums of convex sets, and more generally Borel sets, play a central role in geometry and functional analysis. For example, the well known Brunn-Minkowski inequality can be used to deduce the Euclidean isoperimetric inequality, which identifies the Euclidean ball as the set of any given volume with minimal surface area. Therefore, it is somewhat surprising that in the literature, there has only been rather limited exploration of geometric analogies of sumset estimates. We work towards correcting that oversight in this contribution. The goal of this paper is to explore a variety of new inequalities for volumes of Minkowski sums of convex sets, which have a combinatorial flavor and are inspired by known inequalities in the discrete setting of additive combinatorics. These inequalities are related to the notion of supermodularity: we say that a set function \(F:2^{[n]}\ra{\mathbb R}\) is *supermodular* if \(F(s\cupt)+F(s\capt) \geq F(s) + F(t)\) for all subsets \(s, t\) of \([n]\), and that \(F\) is *submodular* if \(-F\) is supermodular. Our study is motivated by two relatively recent observations. The first observation motivating this paper, due to (Theorem 4.5), states that given convex bodies \(A, B_1, B_2\) in \({\mathbb R}^{n}\), \(|A+B_1+B_2|+|A|\geq |A+B_1|+|A+B_2|\). This inequality has a form similar to that of Kneser's theorem--indeed, observe that the latter can be written as \(\#(A+B+H)+\#(H) \geq \#(A+H) +\#(B+H)\), since adding the stabilizer to \(A+B\) does not change it. Furthermore, it implies that the function \(v: 2^{[n]} \to {\mathbb R}\) defined, for given convex bodies \(B_1, \dots, B_k\) in \({\mathbb R}^{n}\), by \(v(s)=\left|\sum_{i\in s} B_i \right|\) is supermodular. Foldes and Hammer defined the notion of higher order supermodularity for set functions. In Section [3](#sec:super){reference-type="ref" reference="sec:super"}, we generalize their definition and main characterization theorem from to functions defined on \({\mathbb R}_+^n\), and apply it to show that volumes and mixed volumes satisfy this higher order supermodularity. The second observation motivating this paper is due to Bobkov and the second named author , who proved that given convex bodies \(A, B_1, B_2\) in \({\mathbb R}^{n}\), \[\label{eqBM3} |A+B_1+B_2 | |A| \leq 3^n |A+B_1| |A+B_2|.\] The above inequality is inspired by an inequality in information theory analogous to the Plünnecke-Ruzsa inequality (the most general version of which was proved by Ruzsa for compact sets in, and is discussed in Section [2.2](#secPR){reference-type="ref" reference="secPR"} below). If not for the multiplicative factor of \(3^n\) in ([\[eqBM3\]](#eqBM3){reference-type="ref" reference="eqBM3"}), this inequality would imply that the *logarithm* of the volume of the Minkowski sum of convex sets is submodular. In this sense, it goes in the reverse direction to the supermodularity of volume and thus complements it. However, the constant \(3^n\) obtained by is rather loose. We take up the question of tightening this constant in Section [4](#sec:Pl\"unnecke){reference-type="ref" reference="sec:Pl\\\"unnecke"}. Specifically, we obtain both upper and lower bounds for the optimal constant \[\begin{aligned}
\label{cn-def} c_n=\sup \frac{|A+B+C | |A| }{|A+B| |A+B|}, \end{aligned}\] where the supremum is taken over all convex bodies \(A, B, C\) in \({\mathbb R}^n\), in general dimension \(n\). We get an upper bound of \(c_n\leq \varphi^{n}\) in Section [4.2](#ss:gen-ub){reference-type="ref" reference="ss:gen-ub"}, where \(\varphi=(1+\sqrt{5})/2\) is the golden ratio, and an asymptotic lower bound of \(c_n\geq (4/3+o(1))^n\) in Section [4.4](#ss:PR-LB){reference-type="ref" reference="ss:PR-LB"}. In Section [4.3](#ss:ub34){reference-type="ref" reference="ss:ub34"}, we show that the optimal constant is \(1\) in dimension \(2\) and \(\frac{4}{3}\) in dimension 3 (i.e., \(c_2=1\) and \(c_3=4/3\)), and also that \(c_4\leq 2\). In Section [4.5](#ss:improved){reference-type="ref" reference="ss:improved"}, we improve inequality ([\[eqBM3\]](#eqBM3){reference-type="ref" reference="eqBM3"}) in the special case where \(A\) is an ellipsoid, \(B_1\) is a zonoid, and \(B_2\) is any convex body: in this case, the optimal constant is \(1\). This result partially answers a question of Courtade, who asked (motivated by an analogous inequality in information theory) if \(|A+B_1+B_2|\,|A|\le |A+B_1|\,|A+B_2|\) holds when \(A\) is the Euclidean ball and \(B_1, B_2\) are arbitrary convex bodies. Finally, in Section [4.6](#ss:compact){reference-type="ref" reference="ss:compact"}, we prove that ([\[eqBM3\]](#eqBM3){reference-type="ref" reference="eqBM3"}) cannot possibly hold in the more general setting of compact sets with any absolute constant, which signifies a sharp difference between the proof of this inequality compared with the tools used by Ruzsa in . The last section of the paper is dedicated to questions surrounding Ruzsa's triangle inequality: if \(A, B,\) and \(C\) are finite subsets of an abelian group, then \(\#(A)\#(B-C)\leq \#(A-B)\#(A-C)\). The inequality is also known to be true for volume of compact sets in \({\mathbb R}^n\): \(|A|\,|B-C|\le |A-B|\,|A-C|\). We investigate the best constant \(c\) such that the inequality \[\label{triang3} |A|\,|A+B+C|\leq c |A-B|\,|A-C|.\] is true for all convex sets \(A,B\) and \(C\) in \({\mathbb R}^n\). For example, in the plane, we observe that it holds with the sharp constant \(c=\frac{3}{2}\). Again, it is interesting to note that ([\[triang3\]](#triang3){reference-type="ref" reference="triang3"}) is different from Ruzsa's triangle inequality, and it is not true, with any absolute constant \(c\), if one omits the assumption of convexity. In a companion paper, we explore the question of reducing the constant in the Plünnecke-Ruzsa inequality for volumes from \(\varphi^{n}\), when we restrict attention to the subclass of convex bodies known as zonoids. In another companion paper, we explore measure-theoretic extensions of the preceding results for convex bodies, in the category of \(\log\)-concave and in particular Gaussian measures. We also mention that there are probabilistic or entropic analogs of many of the inequalities in this paper. For example, the afore-mentioned observation due to, that a Plünnecke-Ruzsa inequality for convex bodies holds with a constant \(3^n\), emerges as a consequence of Rényi entropy comparisons for convex measures on the one hand, and the submodularity of entropy of convolutions on the other. The submodularity of entropy of convolutions refers to the inequality \(h(X)+h(X+Y+Z)\leq h(X+Y)+h(X+Z)\), where \(h\) denotes entropy, and \(X, Y, Z\) are independent \({\mathbb R}^n\)-valued random variables, and may be thought of as an entropic analogue of the Plünnecke-Ruzsa inequality. This latter inequality was obtained in as part of an attempt to develop an additive combinatorics of probability measures where cardinality or volume is replaced by entropy. A number of works have explored this avenue, starting with for discrete probability measures on groups (e.g., when the random variables take values in finite groups or the integers), and for probability measures on \({\mathbb R}^n\) and more general locally compact abelian groups. Piotr Nayar and Tomasz Tkocz independently obtained upper and lower bounds on the optimal constants in the Plünnecke-Ruzsa inequality for volumes (versions of Theorems [\[thm:ub\]](#thm:ub){reference-type="ref" reference="thm:ub"} and [\[thm:lb\]](#thm:lb){reference-type="ref" reference="thm:lb"}, though with weaker bounds obtained using different methods); we are grateful to them for communicating their work. We are indebted to Ramon Van Handel for pointing to us the original work of W. Fenchel on the local version of Alexandrov's inequality, to Daniel Hug for suggesting that we consider equality cases in Theorem [\[lmcool2\]](#lmcool2){reference-type="ref" reference="lmcool2"}, and to Mathieu Meyer and Dylan Langharst for a number of valuable discussions and suggestions.
# Preliminaries {#sec:prelim}
## Mixed Volumes
In this section, we introduce basic notation and collect essential facts and definition from Convex Geometry that are used in the paper. As a general reference on the theory we use. We write \(x \cdot y\) for the inner product of vectors \(x\) and \(y\) in \({\mathbb R}^n\) and by \(|x|\) the length of a vector \(x \in {\mathbb R}^n\). The closed unit ball in \({\mathbb R}^n\) is denoted by \(B_2^n\), and its boundary by \(S^{n-1}\). We will also denote by \(e_1, \dots, e_n\) the standard orthonormal basis in \({\mathbb R}^n\). Moreover, for any set in \(A \subset {\mathbb R}^n\), we denote its boundary by \(\partial A\). A convex body is a convex, compact set with nonempty interior. We write \(|K|_m\) for the \(m\)-dimensional Lebesgue measure (volume) of a measurable set \(K \subset {\mathbb R}^n\), where \(m = 1,..., n\) is the dimension of the minimal affine space containing \(K\), we will often use the shorten notation \(|K|\) for \(n\)-dimensional volume. A polytope which is the Minkowski sum of finitely many line segments is called a zonotope. Limits of zonotopes in the Hausdorff metric are called zonoids, see, Section 3.2, for details. From , for any compact convex sets \(K_1,\dots K_r\) in \({\mathbb R}^n\) and any non-negative numbers \(t_1, \dots, t_r\) one has \[\begin{aligned}
\label{eq:mvf} \left|t_1K_1+\cdots+t_rK_r\right|= \sum_{i_1,\dots,i_n=1}^rt_{i_1}\cdots t_{i_n}V(K_{i_1},\dots K_{i_n}), \end{aligned}\] for some non-negative numbers \(V(K_{i_1},\dots K_{i_n})\), which are called the mixed volumes of \(K_1,\dots K_r\). One readily sees that the mixed volumes satisfy \(V(K,\dots,K)=|K|\), moreover, they satisfy a number of properties which are crucial for our study (see ) including the fact that a mixed volume is symmetric in its argument; it is multilinear, i.e. for any \(\lambda, \mu \ge 0\) we have \(V(\lambda K + \mu L, K_2, \dots, K_n)=\lambda V(K,K_2, \dots, K_n)+ \mu V(L,K_2, \dots, K_n).\) Mixed volume is translation invariant, i.e. \(V(K+a,K_2, \dots K_n)= V(K,K_2, \dots, K_n),\) for \(a \in {\mathbb R}^n\) and satisfy a monotonicity property, i.e \(V(K,K_2, K_3, \dots, K_n) \le V(L,K_2, K_3, \dots, K_n)\), for \(K \subset L\). We will also often use a two body version of ([\[eq:mvf\]](#eq:mvf){reference-type="ref" reference="eq:mvf"})--the Steiner formula: \[\begin{aligned}
\label{eq:ste} \left|A+tB\right|= \sum_{k=0}^n {n \choose k} t^k V(A[n-k],B[k]), \end{aligned}\] for any \(t>0\) and compact, convex sets \(A, B\) in \({\mathbb R}^n\), where for simplicity we use notation \(A[m]\) for a convex set \(A\) repeated \(m\) times. Mixed volumes are also very useful for studying the volume of orthogonal projections of convex bodies. Let \(P_H A\) be an orthogonal projection of a convex body \(A\) onto \(m\) dimensional subspace \(H\) of \({\mathbb R}^n\), then \[\begin{aligned}
\label{eq:proj} |P_HA|_m|U|_{n-m}={{n}\choose{m}}V(A[m], U[n-m]), \end{aligned}\] where \(U\) is any convex body of volume one in the subspace \(H^\perp\) orthogonal to \(H\). For example, if we denote by \(\theta^\perp = \{x \in {\mathbb R}^n: x \cdot \theta =0\}\) a hyperplane orthogonal to \(\theta\in S^{n-1}\), we obtain \[\begin{aligned}
\label{eq:proj1} |P_{\theta^\perp} A|_{n-1}= n V(A[n-1], [0, \theta]). \end{aligned}\] Yet another useful formula is connected with computation of surface area and mixed volumes: \[\begin{aligned}
\label{surface} |\partial A|= nV(A[n-1], B_2^n), \end{aligned}\] where by \(|\partial A|\) we denote the surface area of the compact set \(A\) in \({\mathbb R}^n\). Mixed volumes satisfy a number of extremely useful inequalities. The first one is the Brunn-Minkowski inequality \[\begin{aligned}
\label{BM} |A+B|^{1/n}\ge |A|^{1/n}+|B|^{1/n}, \end{aligned}\] whenever \(A,B\) and \(A+B\) are measurable. The most powerful inequality for mixed volumes is the Alexandrov--Fenchel inequality: \[\begin{aligned}
\label{AF} V(K_1,K_2, K_3, \dots, K_n) \ge \sqrt{V(K_1,K_1, K_3, \dots, K_n)V(K_2,K_2, K_3, \dots, K_n)}, \end{aligned}\] for any compact convex sets \(K_1,\dots K_r\) in \({\bf R}^n\). We will also use the following classical local version of Alexandrov-Fenchel's inequality that was proved by W. Fenchel (see and also ) and further generalized in \[\label{alexloc} |A|V(A[n-2], B, C)\le 2 V(A[n-1],B) V(A[n-1],C),\] for any convex compact sets \(A,B,C\) in \({\mathbb R}^n\), moreover it was noticed in that ([\[alexloc\]](#alexloc){reference-type="ref" reference="alexloc"}) is true with constant one instead of two in the case when \(A\) is a simplex. The inequality turned out to be a part of rich class Bezout inequalities proposed in . The core tool of our work is the following inequality of J. Xiao (Theorem 1.1 and Lemma 3.3 in ) \[\begin{aligned}
\label{eq:jxiao} |A|V(A[n-j-m], &B[j], C[m]) \nonumber \\&\le \min\left( {n \choose j}, {n \choose m} \right)V(A[n-j],B[j])V(A[n-m],C[m]). \end{aligned}\]
## Plünnecke-Ruzsa inequality. {#secPR}
Plünnecke-Ruzsa inequalities (see for example ) is an important class of inequalities in the field of additive combinatorics. These were introduced by Plünnecke and generalized by Ruzsa, and a simpler proof was given by Petridis; a more recent generalization is proved in, and entropic versions are developed in. For illustration, the form of Plünnecke's inequality developed in states that, if \(A, B_1,\ldots ,B_m\) are finite sets in a commutative group, then there exists an \(X \subset A, X \neq \emptyset\), such that \[\begin{aligned}
\#(A)^m\#(X+B_1+\ldots +B_m) \leq \#(X) \prod_{i=1}^m\#(A + B_i). \end{aligned}\] In, Ruzsa generalized the above inequality to the case of compact sets on a locally compact commutative group, written additively, with the Haar measure. The volume case of this deep theorem is one of our main inspirations: for any compact sets \(A,B_1, \ldots ,B_m\) in \({\bf R}^n,\) with \(|A|>0\) and for every \(\varepsilon >0\) there exists a compact set \(A' \subset A\) such that \[\label{eq:ruzvol} |A|^m|A'+B_1+\ldots +B_m| \le (1+\varepsilon)|A'| \prod_{i=1}^m|A + B_i|.\] It immediately follows that for any compact sets \(A,B_1, \ldots ,B_m\) in \({\bf R}^n,\) \[\label{eq:ruzvol1} |A|^{m-1}|B_1+\ldots +B_m| \le \prod_{i=1}^m|A + B_i|.\]
## Submodularity and supermodularity {#sec:smod-prelim}
Let us first recall the notion of a supermodular set function.
One says that a set function \(F\) is submodular if \(-F\) is supermodular. Submodularity is closely related to a partial ordering on hypergraphs as we will see below. This relationship is frequently attributed to Bollobas and Leader (cf. ), where they introduced the related notion of "compressions". However, it seems to go back much longer--it is more or less explicitly discussed in a 1975 paper of Emerson, where he says it is "well known". To present this relationship, let us introduce some notation. Let \(\mbox{\){\cal M}\(}(n,m)\) be the following family of (multi)hypergraphs: each consists of non-empty (ordinary) subsets \(s_i\) of \([n]\), \(s_i=s_j\) is allowed, and \(\sum_i |s_i|= m\). Consider a given multiset \(\mathcal{C} = \{s_1, \dots, s_l\} \in \mbox{\){\cal M}\(}(n,m)\). The idea is to consider an operation that takes two sets in \(\mathcal{C}\) and replaces them by their union and intersection; however, note that (i) if \(s_i\) and \(s_j\) are nested (i.e., either \(s_i \subset s_j\) or \(s_j \subset s_i\)), then replacing \((s_i,s_j)\) by \((s_i \cap s_j ,s_i \cup s_j)\) does not change \(\mathcal{C}\), and (ii) if \(s_i \cap s_j = \emptyset\), the empty set may enter the collection, which would be undesirable. Thus, take any pair of non-nested sets \(\{s_i,s_j\}\subset \mathcal{C}\) and let \(\mathcal{C}' = \mathcal{C}(i,j)\) be obtained from \(\mathcal{C}\) by replacing \(s_i\) and \(s_j\) by \(s_i \cap s_j\) and \(s_i \cup s_j\), keeping only \(s_i \cup s_j\) if \(s_i \cap s_j = \emptyset\). \(\mathcal{C}'\) is called an *elementary compression* of \(\mathcal{C}\). The result of a sequence of elementary compressions is called a *compression*. Define a partial order on \(\mbox{\){\cal M}\(}(n,m)\) by setting \(\mathcal{A} > \mathcal{A}'\) if \(\mathcal{A}'\) is a compression of \(\mathcal{A}\). To check that this is indeed a partial order, one needs to rule out the possibility of cycles, which can be done by noting that if \(\mathcal{A}'\) is an elementary compression of \(\mathcal{A}\) then \[\begin{aligned}
\sum_{s\in\mathcal{A}} |s|^2 < \sum_{s\in\mathcal{A}'} |s|^2. \end{aligned}\]
Note that for every multiset \(\mathcal{A} \in \mbox{\){\cal M}\(}(n,m)\) there is a unique minimal multiset \(\mathcal{A}^{\#}\) dominated by \(\mathcal{A}\), i.e. \(\mathcal{A}^{\#}<\mathcal{A},\) consisting of the sets \(s^{\#}_j = \{i \in [n]: i \text{ lies in at least } j \text{ of the sets } s \in \mathcal{A}\}\). Thus a particularly nice instance of Theorem [\[thm:compr\]](#thm:compr){reference-type="ref" reference="thm:compr"} is for the special case of \(\mathcal{B}=\mathcal{A}^{\#}\) (we refer to page 132 for further discussion). We also have a notion of supermodularity on the positive orthant of the Euclidean space.
We note that Definition [\[def:supermod-Rmaxmin\]](#def:supermod-Rmaxmin){reference-type="ref" reference="def:supermod-Rmaxmin"} can be viewed as an extension of Definition [\[def:supermod\]](#def:supermod){reference-type="ref" reference="def:supermod"} if one consider set functions on \(2^{[n]}\) as a function on \(\{0;1\}^n\). Indeed, if \(f:{\mathbb R}_{+}^{n} \ra{\mathbb R}\) is supermodular then we define the function \(F:2^{[n]}\to{\mathbb R}\), for \(s\subset[n]\), by \(F(s)=f(e(s))\), where \(e(s)_i=1\) if \(i\in s\) and \(e(s)_i=0\) if \(i\notin s\). Then, the set function \(F\) is supermodular:
The fact that supermodular functions are closely related to functions with increasing differences is classical (see, e.g., or , which describes more general results involving arbitrary lattices). We will denote by \(\partial_i f\) the partial derivative of function \(f\) with the respect to the \(i\)'s coordinate and by \(\partial^m_{i_1, \dots, i_m}\) the mixed derivative with respect to coordinates \(i_1, \dots i_m\).
We will prove Proposition [\[prop:diff\]](#prop:diff){reference-type="ref" reference="prop:diff"} as part of a more general statement on the mixed derivatives of the supermodular functions of higher order (Theorem [\[thm:diff\]](#thm:diff){reference-type="ref" reference="thm:diff"} below).
# Higher order supermodularity of mixed volumes {#sec:super}
## Local characterization of higher order supermodularity
We now present analogues of the above development for higher-order supermodularity. Let us notice that a set function \(F:2^{[n]}\to{\mathbb R}\) is supermodular if and only if for any \(s_0,s_1, s_2\in 2^{[n]}\) with \(s_1\cap s_2=\emptyset\) one has \[F(s_0\cup s_1)+F(s_0\cup s_2)\le F(s_0\cup s_1\cup s_2)+F(s_0).\] Generalizing this property, Foldes and Hammer defined the notion of higher order supermodularity. In this section, we will adapt their definition and study the following property:
Note that for \(m=2\) in the above definition, we recover a supermodular set function. We also introduce the notion of higher-order supermodularity for functions defined on the positive orthant of a Euclidean space.
For \(m=1\) in the above definition, we obtain that \(f\) is \(1\)-supermodular if and only if it is non-decreasing in each coordinate. For \(m=2\), we recover a supermodular function on the orthant as we prove in the following lemma.
The next theorem generalizes Proposition [\[prop:diff\]](#prop:diff){reference-type="ref" reference="prop:diff"} to higher order supermodularity.
As an example, which will help to understand the connection of supermodularity to Minkowski sum of sets, let \(\varphi: {\mathbb R}_+\to{\mathbb R}\) be a convex function. Then for every \(a_0,a_1,a_2\in{\mathbb R}_+\) one has \[\begin{aligned}
\varphi(a_0+a_1)+\varphi(a_0+a_2)\le \varphi(a_0+a_1+a_2)+\varphi(a_0). \end{aligned}\] This property can be seen again as the supermodularity of the function \(\Phi:2^{[2]}\to{\mathbb R}\) defined by \(\Phi(s)=\varphi(a_0+e(s)_1a_1+e(s)_2a_2)\), for any \(s\in 2^{[2]}\), where \(e(s)\) is defined above Lemma [\[lem:set-rn\]](#lem:set-rn){reference-type="ref" reference="lem:set-rn"}. We remark in passing that the positivity of mixed partial derivatives and its global manifestation also arises in the theory of copulas in probability (see, e.g., ). In particular, it is well known there that for smooth functions \(C:[0,1]^m\ra [0,1]\), the condition \(\partial^{m}_{1,2,\ldots,m} C\geq 0\) is equivalent to the condition that \(\sum_{z\in \{x_i, y_i\}^m} (-1)^{N(z)} C(z)\geq 0\) for every box \(\prod_{i=1}^m [x_i,y_i] \subset [0,1]^m\), where \(N(z)=\#\{k:z_k=x_k\}\).
## Higher order supermodularity of volume {#ss:vol-super}
We notice that in Theorem [\[thm:supmodvol\]](#thm:supmodvol){reference-type="ref" reference="thm:supmodvol"} the convexity assumption is essential. Indeed, as was observed in, for \(k=3\), there exists non convex sets \(B_1, B_2, B_3\) such that the function \(v\) defined above is not supermodular. We will discuss this issue in more details in Section [3.3](#sec:notconv){reference-type="ref" reference="sec:notconv"} below. Using Theorem [\[thm:supmodvol\]](#thm:supmodvol){reference-type="ref" reference="thm:supmodvol"}, Remark [\[rk:mixed-vol-sm\]](#rk:mixed-vol-sm){reference-type="ref" reference="rk:mixed-vol-sm"}, Remark [\[rk:setsuper\]](#rk:setsuper){reference-type="ref" reference="rk:setsuper"} and Theorem [\[thm:compr\]](#thm:compr){reference-type="ref" reference="thm:compr"} we deduce the following corollary.
Let us note that the above \(m\)-supermodularity of the function \(\bar{v}\) is equivalent to the fact that for any convex bodies \(B_0, B_1, \dots, B_{k}, C_1, \dots, C_l\) in \({\mathbb R}^n\) \[\sum_{s\subset[k]}(-1)^{k-|s|}V\left(\left(B_0+\sum_{i\in s}B_i\right)[n-l], C_1, \dots, C_l\right)\ge0.\] Applying the previous theorem to \(l=0\), we get \[\label{eq_B_0} \sum_{s\subset[k]}(-1)^{k-|s|}\left|B_0+\sum_{i\ins} B_i\right| \geq 0.\] The above inequality for \(k=n\) and \(B_0=\{0\}\) follows also directly from the following classical formula (see Lemma 5.1.4 in ) \[\sum_{s\subset[n]}(-1)^{n-|s|}\left|\sum_{i\ins} B_i\right| = n!V(B_1, \dots, B_n).\] In the same way, we can also give another proof of the general case of ([\[eq_B\_0\]](#eq_B_0){reference-type="ref" reference="eq_B_0"}).
Thanks to the fact that supermodular set functions taking the value 0 at the empty set are fractionally superadditive (see, e.g., ), we can immediately deduce the following inequality. Let \(n\ge1\), \(k\ge2\) be integers and let \(A_1, \dots, A_k\) be \(k\) convex sets in \({\bf R}^n\). Then, for any fractional partition \(\beta\) using a hypergraph \(\mathcal{C}\) on \([k]\), \[\begin{aligned}
\label{eq:fsa} \left|\sum_{i=1}^kA_i\right| \ge \sum_{s\in\mathcal{C}} \beta(s)\left|\sum_{j\in s}A_j\right|. \end{aligned}\] It was shown in that [\[eq:fsa\]](#eq:fsa){reference-type="eqref" reference="eq:fsa"} actually extends to all compact sets in \({\mathbb R}^n\), but supermodularity does not extend to compact sets as discussed in the next section.
## Going beyond convex bodies {#sec:notconv}
Consider sets \(A, B \subset {\bf R}^n\), such that \(0 \in B\). Define \(\Delta_B(A)=(A+B)\setminus A\), and note that \(A+B\) is always a superset of \(A\) because of the assumption that \(0\in B\). The supermodularity of volume is also saying something about set increments. Indeed, for any sets \(A, B, C\) consider \[\begin{aligned}
\Delta_C\Delta_B(A) =\Delta_C \big((A+B)\setminus A\big) =\bigg(\big((A+B)\setminus A\big)+C \bigg)\setminus \big((A+B)\setminus A\big). \end{aligned}\] We have, if \(0 \in B \cap C\): \[\begin{aligned}
\label{difference} \left|\Delta_C\Delta_B(A)\right| &=\left|\big((A+B)\setminus A\big)+C\right|-\left|(A+B)\setminus A\right| \nonumber\\ &\geq \left|(A+B+C)\setminus (A+C)\right|-\left|A+B\right|+\left|A \right|\nonumber\\ &= |A+B+C|-|A+C|-|A+B|+ |A|, \end{aligned}\] where the inequality follows from the general fact that \((K+C)\setminus (L+C)\subset (K\setminus L)+C\). Moreover, if \(A,B,C\) are convex, compact sets then the estimate is non-trivial, i.e., using Theorem [\[sch\]](#sch){reference-type="ref" reference="sch"} we get that the right hand side of the above quantity is non-negative. It is interesting to note that the \(\Delta\) operation is not commutative, i.e. \(\Delta_C\Delta_B(A) \not = \Delta_A\Delta_C(A)\); this can be seen, for example, in \({\mathbb R}^2\) by taking \(A\) to be a square, \(B\) to be a segment, and \(C\) to be a Euclidean ball. It is natural to ask if the higher-order analog of this observation remains true.
The inequality ([\[difference\]](#difference){reference-type="ref" reference="difference"}) gives a positive answer to the above question in the case \(m=2\). We also observe that if \(B_0, B_1, \dots, B_m\) are convex, then the right hand side is non-negative thanks to Theorem [\[thm:supmodvol\]](#thm:supmodvol){reference-type="ref" reference="thm:supmodvol"}. We note that it was observed in, by considering \(A=\{0,1\}\) and \(B=C=[0,1]\) in \({\mathbb R}^1\), that the volume of Minkowski sums cannot be supermodular (even in dimension 1) if the convexity assumption on the set \(A\) is removed. Nonetheless observed that if \(A, B, C\subset {\bf R}\) are compact, then \[\begin{aligned}
|A+B+C| +|\mathrm{conv}(A)|\ge |A+B| + |A+C|; \end{aligned}\] it is unknown if this extends to higher dimension. In particular, we do not know if the following conjecture is true for \(n\ge2\).
We can confirm Conjecture [\[submodulnotconv\]](#submodulnotconv){reference-type="ref" reference="submodulnotconv"} under the assumption that \(B\) is a zonoid.
# Plünnecke-Ruzsa inequalities for convex bodies {#sec:Pl\\\"unnecke}
## Existing Plünnecke-Ruzsa inequality for convex bodies {#ss:plun}
Bobkov and Madiman developed a technique for going from entropy to volume estimates, by using certain reverse Hölder inequalities that hold for convex measures. Specifically, shows that if \(X_i\) are independent random variables with \(X_i\) uniformly distributed on a convex body \(K_i \subset {\mathbb R}^n\) for each \(i=1,\ldots, m\), then \(h(X_1+\ldots+X_m)\geq \log |K_1+\ldots+K_m|-n\log m\), where the entropy of a random variable \(X\) with density \(f\) on \({\mathbb R}^n\) is defined by \[\label{eq:entropy} h(X)=-\int f(x)\log f(x) dx.\] This is a reverse Hölder inequality in the sense that \(h(X_1+\ldots+X_m)\leq \log |K_1+\ldots+K_m|\) may be seen by applying Hölder's inequality and then taking a limit. More general sharp inequalities relating Rényi entropies of various orders for measures having convexity properties are described in (see also ). Applied to the submodularity of entropy of sums discovered in, they use this technique to demonstrate the following inequality.
In particular, by choosing \(k=1\), one already obtains an interesting inequality for volumes of Minkowski sums.
Thus, one may think of Corollary [\[cor:Pl\\\"unnecke\]](#cor:Pl\"unnecke){reference-type="ref" reference="cor:Pl\\\"unnecke"} as providing yet another continuous analogues of the Plünnecke-Ruzsa inequalities in the context of volumes of convex bodies in Euclidean spaces (compare with ([\[eq:ruzvol\]](#eq:ruzvol){reference-type="ref" reference="eq:ruzvol"})), where going from the discrete to the continuous incurs the extra factor of \((1+m)\), but one does not need to bother with taking subsets of the set \(A\). In particular, with \(m=2\), one gets "log-submodularity of volume up to an additive term" on convex bodies.
Unfortunately the dimension-dependent additive term is a hindrance that one would like to remove or improve, which is the purpose of the next section.
## Improved upper bounds in general dimension {#ss:gen-ub}
In this section, we will present an improvements in the constant \(3^n\) in the three body inequality from Corollary [\[cor:Pl\\\"unnecke2\]](#cor:Pl\"unnecke2){reference-type="ref" reference="cor:Pl\\\"unnecke2"}. We define the constant \(c_n\) by [\[cn-def\]](#cn-def){reference-type="eqref" reference="cn-def"}, or equivalently as the infimum of the constants \(c>0\) such that, for every convex compact sets \(A,B,C\) in \({\mathbb R}^n\), \[\begin{aligned}
|A|\,|A+B+C| \le c |A+B|\,|A+C|. \end{aligned}\] We recall that \(\varphi=(1+\sqrt{5})/2\) denotes the golden ratio.
The next proposition gives a different proof of ([\[eq:ruzvol1\]](#eq:ruzvol1){reference-type="ref" reference="eq:ruzvol1"}) in the special case of convex sets and, we hope, gives yet another example of how the methods of mixed volumes as well as the Bézout type inequality ([\[eq:jxiao\]](#eq:jxiao){reference-type="ref" reference="eq:jxiao"}) can be applied in this context.
## Improved constants in dimensions 3 and 4 {#ss:ub34}
Theorem [\[thm:ub\]](#thm:ub){reference-type="ref" reference="thm:ub"} gives an optimal bound of \(1\) for the three body inequality in dimension \(2\). Next, we will show how we can get better bounds for \(c_n\) in dimension \(3\) and \(4\).
## Lower bounds in general dimension {#ss:PR-LB}
In this section, we provide a lower bound for the Plünnecke-Ruzsa inequality for convex bodies. A weaker lower bound was also independently obtained by Nayar and Tkocz. We first observe that the best constant \(c_n\) in the Plünnecke-Ruzsa inequality \[\label{PR2} |A|\,|A+B+C| \le c_n |A+B|\,|A+C|,\] satisfies \(c_{n+m} \ge c_n c_m\). Indeed, this follows immediately by considering critical examples of \(A_1, B_1, C_1\) in \({\mathbb R}^n\) and \(A_2, B_2, C_2\) in \({\mathbb R}^m\) together with their direct products \(A_1 \times A_2, B_1\times B_2, C_1\times C_2\) in \({\mathbb R}^{n+m}\). Next we notice that if ([\[PR2\]](#PR2){reference-type="ref" reference="PR2"}) is true in a class of convex bodies closed by linear transformations, then \[\label{projcontr} |P_{E \cap H}K| |K| \le c_n |P_E K|\,|P_H K|,\] for any \(K\) in this class and any subspaces \(E,H\) of \({\mathbb R}^n,\) such that \(\dim E =i,\) \(\dim H=j,\) \(i+j \ge n+1\) and \(E^\perp \subset H\). To see this consider \(B=U\), with \(\dim U = n-i,\) \(|U|=1\) and \(C=V\), with \(\dim V = n-j,\) \(|V|=1\) and \(U, V\) belong to orthogonal subspaces of \({\bf R}^n\). Let \(A=tK\), where \(t > 0\) and set \(k=n-(n-i)-(n-j)=i+j-n\). Then ([\[PR2\]](#PR2){reference-type="ref" reference="PR2"}) yield together with ([\[eq:mvf\]](#eq:mvf){reference-type="ref" reference="eq:mvf"}) and ([\[eq:proj\]](#eq:proj){reference-type="ref" reference="eq:proj"}) \[\begin{aligned}
t^n|K|\big(\sum_{m=k}^n &{{n}\choose{m}}V(K[m], (U+V)[n-m])t^m \big) \\ &\le c_n\big(\sum_{m=i}^n {{n}\choose{m}}V(K[m], U[n-m])t^m \big)\big(\sum_{m=j}^n {{n}\choose{m}}V(K[m], V[n-m])t^m \big). \end{aligned}\] Dividing the above inequality by \(t^{n+k}\) and taking \(t= 0\) we get \[|K|{{n}\choose{k}}V(K[k], (U+V)[n-k]) \le c_n {{n}\choose{i}}V(K[i], U[n-i]) {{n}\choose{j}}V(K[j], V[n-j]).\] Finally, using ([\[eq:proj\]](#eq:proj){reference-type="ref" reference="eq:proj"}), we get ([\[projcontr\]](#projcontr){reference-type="ref" reference="projcontr"}). It was proved in that \[\label{eqGPH} |P_{\{u,v\}^\perp}K| |K| \le \frac{2(n-1)}{n} |P_{u^\perp} K|\,|P_{v^\perp}K|,\] for any convex body \(K \subset {\mathbb R}^n\) and a pair of orthogonal vectors \(u,v \in S^{n-1}\). It was also shown in that the constant \(2(n-1)/n\) is optimal. Thus \(c_n \ge 2-\frac{2}{n}\) and this estimate gives a sharp constant in \({\bf R}^3\): \(c_3 = 4/3.\) In the case when \(n=4\), we get \(c_4 \ge 3/2\). The inequalities analogous to ([\[eqGPH\]](#eqGPH){reference-type="ref" reference="eqGPH"}) and ([\[projcontr\]](#projcontr){reference-type="ref" reference="projcontr"}) were studied in many other works, including. In particular, it was proved in that ([\[projcontr\]](#projcontr){reference-type="ref" reference="projcontr"}) is sharp with \[c_n \ge c_n(i,j,k)=\frac{{{i}\choose{k}}{{j}\choose{k}}}{{{n}\choose{k}}}.\] Thus to find a lower bound on \(c_n\) one may maximize over \(c_n(i,j,k)\) with restriction that \(i+j \ge n+1\) and \(k=i+j-n\). One may use Stirling's approximation, with \(i=j=2n /3\) and \(k=n/3\) (when \(n\) is a multiple of 3, with minor modifications if not) to obtain the following theorem.
## Improved upper bound for subclasses of convex bodies {#ss:improved}
The goal of this section is to prove the following theorem
Theorem [\[thm:zonoid-ellipsoid\]](#thm:zonoid-ellipsoid){reference-type="ref" reference="thm:zonoid-ellipsoid"} motivates us to pose the following conjecture.
A detailed study of this conjecture is undertaken in the forthcoming paper. Before proving Theorem [\[thm:zonoid-ellipsoid\]](#thm:zonoid-ellipsoid){reference-type="ref" reference="thm:zonoid-ellipsoid"} we will prove a theorem which would help us to verify Plünnecke-Ruzsa inequalities for convex bodies for a fixed body \(A\).
Next we will prove that \(B_2^n\) satisfies the conditions of theorem [\[thm:zonoid-projratio\]](#thm:zonoid-projratio){reference-type="ref" reference="thm:zonoid-projratio"}.
Let \(T\) be the affine transform such that \(B=T(B_2^n)\). If \(B\) lives in an hyperplane then \(|B|=0\) and the inequality holds. If not, then \(T\) is invertible and since the affine image of a zonoid is a zonoid by applying \(T^{-1}\) we may assume that \(B=B_2^n\). Now the theorem follows immediately from Theorems [\[thm:projection-ball\]](#thm:projection-ball){reference-type="ref" reference="thm:projection-ball"} and [\[thm:zonoid-projratio\]](#thm:zonoid-projratio){reference-type="ref" reference="thm:zonoid-projratio"}.
## The case of compact sets {#ss:compact}
Let us note that inequality, \[\begin{aligned}
\label{eq:cnonconv} |A|\,|A+B+C|\le |A+B|\,|A+C|. \end{aligned}\] is valid, when \(A,B\) are intervals and \(C\) is any compact set in \({\mathbb R}\). Indeed, by approximation we may assume that \(C\) is a finite union of closed intervals, \(A=[0,a]\) and \(B=[0,b],\) for some \(a, b \ge 0\). Then we may assume that \(A+C=\cup_{i=1}^m [\alpha_i, \beta_i+a]\) where intervals \([\alpha_i, \beta_i+a]\) are mutually disjoint. Then \[|A+C| = ma+\sum\limits_{i=1}^m (\beta_i-\alpha_i),\] \[|A+B+C| \le \sum \limits_{i=1}^m (\beta_i-\alpha_i +a+b)=m(a+b)+\sum\limits_{i=1}^m (\beta_i-\alpha_i)\] and ([\[eq:cnonconv\]](#eq:cnonconv){reference-type="ref" reference="eq:cnonconv"}) follows from \[a (m(a+b)+\sum\limits_{i=1}^m (\beta_i-\alpha_i))\le (ma+\sum\limits_{i=1}^m (\beta_i-\alpha_i))(a+b).\] We also note that, as we discussed before, inequality ([\[eq:ruzvol\]](#eq:ruzvol){reference-type="ref" reference="eq:ruzvol"}) as well as inequality ([\[eq:threesimple\]](#eq:threesimple){reference-type="ref" reference="eq:threesimple"}) is valid without additional convexity assumptions (as well as Theorem [\[thm:projection-ball\]](#thm:projection-ball){reference-type="ref" reference="thm:projection-ball"} from above). Still, we will show that there is a sharp difference to those inequalities the convexity assumption in Theorem [\[thm:ub\]](#thm:ub){reference-type="ref" reference="thm:ub"} can not be removed. The construction is inspired by the proof of Theorem 7.1 from:
It turns out (see for example ) that some sumsets estimates can still be proved if the convexity assumption is relaxed by an assumption that the body is star-shaped. The next lemma shows that it is still not the case for Theorem [\[thm:ub\]](#thm:ub){reference-type="ref" reference="thm:ub"}.
# On Ruzsa's triangle inequality {#sec:diff}
In additive combinatorics, the Ruzsa distance is defined by \(d(A,B)=\log \frac{\#(A-B)}{\sqrt{\#(A) \#(B)}},\) where \(A\) and \(B\) are subsets of an abelian group. We refer to for more information and properties of this object, which is useful even though it is *not* a metric (since typically \(d(A,A)>0\)). The Ruzsa distance satisfies the triangle inequality which is equivalent to \(\#(C-B)\cdot \# A \leq \#(A-C) \#(B-A).\) An analogue of Ruzsa's triangle inequality holds for compact sets in \({\bf R}^n\).
This inequality has a short proof that we provide here for the sake of completeness. Indeed \[|B-A| |A-C|=\int_{{\bf R}^n} 1_{B-A}*1_{A-C}(x)dx\ge \int_{B-C} 1_{B-A}*1_{A-C}(x)dx,\] where \(1_M\) is a characteristic function of a set \(M \subset {\bf R}^n\) and \(f*g\) is the convolution of functions \(f,g: {\bf R}^n \to {\bf R}\). Now let \(x\in B-C\), there is \(b\in B\) and \(c\in C\) such that \(x=b-c\). Thus, changing variable, one has \[\begin{aligned}
1_{B-A}*1_{A-C}(b-c)&= \int_{{\bf R}^n}1_{B-A}(z)1_{A-C}(b-c-z)dz=\int_{{\bf R}^n}1_{B-A}(b-y)1_{A-C}(y-c)dy\\ &\ge\int_{A}1_{B-A}(b-y)1_{A-C}(y-c)dy =|A|. \end{aligned}\] In view of Ruzsa's triangle inequality, it is natural to try to generalize Theorem [\[thm:ub\]](#thm:ub){reference-type="ref" reference="thm:ub"} to the case of the difference of convex bodies. We recall that \(\varphi=(1+\sqrt{5})/2\) denotes the golden ratio and for \(n \ge 2\) the constant \(c_n\) was defined in Theorem [\[thm:ub\]](#thm:ub){reference-type="ref" reference="thm:ub"} satisfying \(1=c_2\le c_n\le \varphi^n\).
Next we would like to discuss an improvement of Theorem [\[thm:ruzsa\]](#thm:ruzsa){reference-type="ref" reference="thm:ruzsa"} via an improvement of Litvak's inequality ([\[eq:lit\]](#eq:lit){reference-type="ref" reference="eq:lit"}) in \({\bf R}^2\).
The inequality [\[eqprs3\]](#eqprs3){reference-type="eqref" reference="eqprs3"} is an intriguing improvement (in dimension 2) of Litvak's observation. To see this, observe that by the Brunn-Minkowski inequality, \(|A+C|\geq (\sqrt{|A|}+\sqrt{|C|})^2 = |A|+|C|+2\sqrt{|A|\,|C|} = (\sqrt{|A|}-\sqrt{|C|})^2+4\sqrt{|A|\,|C|}\), and hence the right hand side of [\[eqprs3\]](#eqprs3){reference-type="eqref" reference="eqprs3"} is bounded by \(\frac{3}{2}|A+C|\). Thus, in dimension 2, since \(c_2=1\), we obtain \[\begin{aligned}
\begin{split} |A|\,|A+B+C| &\le |A+B|\,|A+C| \\ &\le (|A-B|+2\sqrt{|A|\,|B|})\,|A+C|\\ &\le \bigg[\frac{3}{2}|A-B|-\frac{1}{2}(\sqrt{|A|}-\sqrt{|B|})^2\bigg] |A+C|, \end{split}\end{aligned}\] which is an improvement of Theorem [\[thm:ruzsa\]](#thm:ruzsa){reference-type="ref" reference="thm:ruzsa"} in dimension 2. Let us define the *additive asymmetry* of the pair \((A,C)\) by \[\text{asym}(A,C)=\bigg| \frac{|A+C|}{|A-C|}-1\bigg|,\] and note that \(\text{asym}(A,C)\) is trivially 0 if either \(A\) or \(C\) is symmetric. Observe that inequality [\[eqprs3\]](#eqprs3){reference-type="eqref" reference="eqprs3"} may be rewritten as \[\begin{aligned}
\label{asym} \text{asym}(A,C) \leq 2e^{-d(A,C)}, \end{aligned}\] where \(d(A,C)=\log \frac{|A-C|}{\sqrt{|A|\cdot |C|}}\) is the Euclidean analogue of the Ruzsa distance defined at the beginning of this section. One wonders if this inequality extends to dimension higher than 2. Finally we note that comparison of cardinality of \(A+C\) and \(A-C\) has also been of interest in additive combinatorics (see, e.g., ), and there are also results comparing the entropies of sums and differences of independent random variables in different abelian groups (see, e.g., ). | {'timestamp': '2022-06-06T02:14:28', 'yymm': '2206', 'arxiv_id': '2206.01565', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01565'} |
null | null |
null | null |
null | null |
# Introduction
The no-cloning theorem is a cornerstone of quantum information theory. On one side it implies that information cannot be intercepted without detection, which is the basis of the widely extolled power of quantum cryptography, on the other side it hampers the reproduction of information stored in quantum states for multiple use, i.e. the use of quantum memories. While, initially quantum computation was hailed mainly for the speed-up with respect to its classical counterpart (for a review see ), it was soon realized that quantum mechanics also leads to an exponential increase in storage capacity for associative retrieval of information and an exponential decrease in memory calls for quantum random access memories (RAM). The prototype associative memory is the Hopfield neural network model. Due to the cross-talk phenomenon, the maximum number of memories that can be classically stored and retrieved is linear in the number \(n\) of neurons (for a review see ). The quantization of the Hopfield model mapping neurons to qubits does not improve on this limit. The exponential gain in capacity is obtained when the patterns are stored in quantum states. The original idea is to store a number \(p\) of binary patterns of \(n\) qubits in a quantum superposition and use this state as the memory. Two different algorithms for retrieval were proposed, one associative, i.e. content-addressable, the other based on the Grover algorithm using a subset of the \(n\) qubits as the memory address. The latter is essentially a quantum RAM memory since it cannot correct corrupted inputs. Various variants of the quantum associative memories have subsequently been proposed. Recently, the whole field of machine learning is being extended to the quantum domain. In an ideal situation one would simply clone the quantum memory state whenever needed for use and keep a master copy for later re-use. However, the no-cloning theorem prevents this. In this paper we point out that, for quantum memories, perfect cloning of a single state is not necessary, it is sufficient to clone the memory state up to a global NOT operation that transforms it into its mirror image. This *mirror modular cloning* can be performed by a (2 \(\times\) 2) unitary transformation that depends on a single parameter of the state and represents thus an easy and efficient "perfect\" cloning of the memory state. We then introduce an improved version of the quantum associative recall and show that this is exponentially faster than the address-based Grover retrieval. This is particularly important in view of recent proposals to use quantum memories for fast associative data triggering in large throughput high-energy experiments at LHC. Let us start from a quantum memory state \(|M \rangle\) encoding \(p\) binary patterns \(|p^i \rangle = |p^i_1 \dots p^i_n\rangle\) of \(n\) qubits, \[|M\rangle = {1\over \sqrt{p}} \sum_{i=1}^p |p^i \rangle \. \label{memory}\] An efficient algorithm to load the patterns to form this state is described in, here we will not repeat the procedure but rather take this state as our starting point. The same algorithm can be used to construct the mirror modular state \[|\overline M\rangle = {1\over \sqrt{p}} \sum_{i=1}^p | \overline p^i \rangle \, \label{mirrorstate}\] where \(\overline p^i\) is the string in which each qubit is reversed compared to \(p^i\). Finally, we measure the scalar product \(\langle M | \overline M \rangle\) (see for e.g. ) of these two states. We now add a normalized ancillary register of \(n\) qubits prepared in state \(|\Sigma \rangle\) and a further ancilla qubit in state \(|0\rangle\) and we borrow a technique from probabilistic quantum cloning to posit a \(2 \times 2\) unitary transformation such that \[\begin{aligned}
U\left( |M\rangle |\Sigma \rangle |0\rangle \right) &&= \sqrt{\gamma} |M\rangle |M\rangle |0\rangle + \sqrt{\overline \gamma} |M\rangle |\overline M\rangle |1\rangle \, \nonumber \\ U\left( |\overline M\rangle |\Sigma \rangle |0\rangle \right) &&= \sqrt{\overline \gamma} |\overline M\rangle |\overline M\rangle |0\rangle + \sqrt{ \gamma} |\overline M\rangle |M\rangle |1\rangle \. \label{posited} \end{aligned}\] This unitary transformation, if it exists, perfectly clones the memory state up to a mirror modular transformation, outputting \(|M\rangle\) with probability \(\gamma\) or \(|\overline M\rangle\) with probability \(\overline \gamma\), the two results being distinguished by the value of the ancilla qubit. As we will show below, this mirror modular cloning is perfectly sufficient for quantum associative retrieval. Before doing that, however, we must show that \(U\) exists. To do so we use the theorem stating that there exists a unitary transformation \(U|\psi_i \rangle = |\phi_i\rangle\), \(i = 1 \dots m\), if the two sets of states \(|\psi_i\rangle\) and \(|\phi_i\rangle\) satisfy \(\langle \psi_i | \psi_j \rangle = \langle \phi_i | \phi_j \rangle\) for all \(i, j = 1\dots m\). The \(2\times 2\) matrix of all scalar products of initial states in ([\[posited\]](#posited){reference-type="ref" reference="posited"}) is \[\begin{pmatrix} 1 & \langle M | \overline M \rangle \\ \langle \overline M | M\rangle & 1 \end{pmatrix} \label{inmatrix}\] The corresponding matrix of scalar products of unitary transformed states is \[\begin{pmatrix} \gamma + \overline \gamma & \sqrt{\gamma \overline \gamma} (\langle M | \overline M \rangle + \langle \overline M| M \rangle) \langle M| \overline M\rangle \\ \sqrt{\gamma \overline \gamma} (\langle \overline M | M \rangle + \langle M| \overline M \rangle) \langle \overline M|M\rangle & \gamma + \overline \gamma \end{pmatrix} \label{outmatrix}\] This shows that the unitary transformation \(U\) exists if the efficiencies \(\gamma\) and \(\overline \gamma\) are chosen to satifsy \[\begin{aligned}
\gamma + \overline \gamma &&= 1 \, \nonumber \\ \sqrt{\gamma \overline \gamma} &&= {1\over \langle M | \overline M \rangle + \langle \overline M| M \rangle} \. \label{eff} \end{aligned}\] It can then be realized with a quantum circuit along the lines detailed in. After separating out the original master copy of the memory state we add an \(n\) qubit register with an input pattern \(|I\rangle = |i_1 \dots i_n \rangle\) and we add \(b\) control qubits prepared in state \(|O\rangle = |c_1 \dots c_b \rangle = |0,\dots 0\rangle\). This is the initial state for the associative retrieval algorithm, \[|\psi_0 \rangle = \sqrt{\gamma} |I\rangle |M\rangle |O\rangle |0\rangle + \sqrt{\overline \gamma}|I\rangle |\overline M\rangle |O\rangle |1\rangle \. \label{initial}\] Before describing retrieval in detail let us focus on the elementary quantum gates (for a review see ) used in this algorithm. First of all there are the single-qubit NOT gate, represented by the first Pauli matrix \(\sigma_1\), and Hadamard gate H, with the matrix representation \[H = {1\over \sqrt{2}} \begin{pmatrix} 1 & 1\\ 1 &-1 \end{pmatrix} \. \label{adda}\] Moreover, we will use the two-qbit XOR (exclusive OR) gate, which performs a NOT operation on the second qubit if and only if the first one is in state \(|1\rangle\). In matrix notation this gate is represented as \({\rm XOR} = {\rm diag} \left( I, \sigma_1 \right)\), where \(I\) denotes a two-dimensional identity matrix. For all these gates we shall indicate by subscripts the qubits on which they are applied, the control qubits coming always first. We start the retrieval algorithm by generating the state \[\begin{aligned}
|\psi _1\rangle &&= \prod_{k=1}^n \ {\rm NOT}_{m_k} \ {\rm XOR}_{i_k m_k} |\psi _0\rangle \nonumber \\ &&= \sqrt{\gamma \over p} \sum_{k=1}^p |I\rangle |D^k\rangle |O\rangle |0\rangle + \sqrt{\overline \gamma\over p} \sum_{k=1}^p |I\rangle |\overline D^k\rangle|O\rangle |1\rangle \, \label{second} \end{aligned}\] where \(|D^k\rangle = |d^k_1 \dots d^k_n \rangle\) and \(|\overline D^k\rangle = |\overline d^k_1 \dots \overline d^k_n \rangle\) with \(d^k_j=1\) iff \(p^k_j = i_j\) and \(\overline d^k_j=1\) iff \(p^k_j \ne i_j\). We now consider the unitary transformation \[U = \prod_{i=1}^b H_{c_i} \ {\rm e}^{ i\pi {\cal H}_i/2n} \ H_{c_i} \, \label{uham}\] where "\(c_i\)\" refers to the \({\rm i}^{\rm th}\) control qubit and \[\begin{aligned}
{\cal H}_i &&= \left( d_H \right)_m \otimes \left( \sigma_3 \right)_{c_i} \, \nonumber \\ \left( d_H \right)_m && = \sum_{k=1}^n \left( {\sigma_3 + 1\over 2} \right) _{m_k}\, \label{h} \end{aligned}\] with \(\sigma _3\) the third Pauli matrix. For each control qubit \({\cal H}_i\) measures the number of 0's in the memory register, now in state \(D^K\) or \(\overline D^k\), with a plus sign if \(c_i\) is in state \(|0\rangle\) and a minus sign if \(c_i\) is in state \(|1\rangle\). When the memory register is in state \(|D^k\rangle\) (first term in ([\[second\]](#second){reference-type="ref" reference="second"})) this is the number of qubits which are different in the input and memory registers. This quantity is called the *Hamming distance* \(d_H\left (i, p^k \right)\) between the stored pattern \(p^k\) under consideration and the input \(i\). When the memory register is in state \(|\overline D^k\rangle\) (second term in ([\[second\]](#second){reference-type="ref" reference="second"})), instead, it represents the number of qubits which are equal in the input and the memory registers, which we denote by \(\overline d_H\left (i, p^k \right)\) when the memory register contains pattern \(p^k\). The two Hadamard gates sandwiching the Hamming distance operator in ([\[uham\]](#uham){reference-type="ref" reference="uham"}) turn every control qubit into the combination \[{\rm cos} {\pi \over 2n} d_H \left( i, p^k \right) |0\rangle + i {\rm sin} {\pi \over 2n} d_H \left( i, p^k \right) |1\rangle \, \label{cossin}\] in the term corresponding to distance pattern \(|D^k\rangle\), with an analogous expression in terms of \(\overline d_H\) for distance patterns \(|\overline D^k\rangle\). After restoring the memory register to its original state with the inverse transformation to ([\[second\]](#second){reference-type="ref" reference="second"}), we obtain the state \[\begin{aligned}
|\psi_2\rangle &&= \sqrt{\gamma \over p} \sum_{k=1}^p \sum_{l=0}^b \ {\rm cos}^{b-l} \left( {\pi\over 2n} d_H\left( i, p^k \right)\right) \times \nonumber \\ &&({\rm i \ sin})^l \left( {\pi\over 2n} d_H\left( i, p^k \right)\right) \ \sum_{\left\{ J^l \right\}} |I\rangle |p^k \rangle |J^l\rangle |0\rangle \nonumber \\ &&+\sqrt{\overline \gamma \over p} \sum_{k=1}^p \sum_{l=0}^b \ {\rm cos}^{b-l} \left( {\pi\over 2n} \overline d_H\left( i, p^k \right)\right) \times \nonumber \\ &&({\rm i \ sin})^l \left( {\pi\over 2n} \overline d_H\left( i, p^k \right)\right) \ \sum_{\left\{ J^l \right\}} |I\rangle |\overline p^k \rangle |J^l\rangle |1\rangle \, \label{al} \end{aligned}\] where \(\left\{ J^l \right\}\) denotes the set of all binary numbers of \(b\) bits with exactly \(l\) bits 1 and \((b-l)\) bits 0. At this point we measure the original ancilla qubit: if \(|0\rangle\) is obtained we use amplitude amplification to rotate the remaining state onto the "good\" subspace with all control qubits in state \(|0_1 \dots 0_b\rangle\), if \(|1\rangle\) is obtained we use the same technique to rotate the remaining state onto the "good\" subspace with all control qubits in state \(|1_1 \dots 1_b\rangle\). After these rotations we can finally measure the memory register. We either obtain the closest (in Hamming distance) pattern \(p^k\) (if \(|0\rangle\) was measured) or the most distant pattern \(\overline p^k\) (if \(|1\rangle\) was measured) with a probability distribution \[P \left (i, p^k \right) = {1\over p} {\rm cos}^{2b} \left( {\pi \over 2n} d_H \left(i, p^k \right) \right) \. \label{prob}\] The probability distribution is the same since, for mirror patterns, \(\overline d_H = n-d_H\) and \({\sin} \left( (\pi/2) (1-d_H/n) \right) = {\rm cos} \left( \pi d_H/2n \right)\). This shows that, as anticipated, it does not matter if one obtains the closest or the most distant pattern: if the original ancilla qubit is measured in state \(|1\rangle\) one needs only to mirror-invert the measured pattern. This is why memory collapse and the difficulty of cloning do not pose any problem for quantum associative memories, differently to what is sometimes discussed in the literature. Mirror modular cloning is a simple and efficient technique that depends on only one parameter of the state. Finally, the complexity of the retrieval algorithm is mainly influenced by the amplitude amplification step. The number of applications \(C\) of the basic amplitude amplification rotation is given by the square root of the inverse probability of measuring the "good\" subspace, i.e. \[C = \sqrt{{p \over \sum_{k=1}^p {\rm cos}^{2b} \left( {\pi \over 2n} d_H \left(i, p^k \right) \right)}} \. \label{compl1}\] This is not a known quantity since it depends on the input under consideration. So, either we need C repetitions of the algorithm, we use amplitude estimation or we record an estimate of C together with the initial memory storage by computing it for some typical inputs, after which we can repeat amplitude amplification a few times varying the number of oracle calls around this initial estimate. Compared to the complexity \(\sqrt{N}\) to retrieve one pattern with the address-oriented Grover algorithm, where \(N\) is the total number of computational basis states, \(N=2^n\), the present complexity not only does not depend on \(N\), but it does not even depend on the pattern number \(p\) since it is determined only by the accuracy parameter \(b\) and the distribution of patterns in the memory. For an approximately uniform pattern distribution it can be estimated as follows. Let us substitute each cosine in the sum by its average, determined using \[{2\over \pi} \int_0^{\pi/2} {\rm cos}^{2b} (x) \ dx = {1\over 2^{2b}} {2b \choose b} \, \label{average}\] and let us approximate the factorials by the Stirling formula \(b! \approx \sqrt{2\pi b} (b/e)^b\) for large \(b\). With this we obtain \[C \approx (\pi b)^{1/4} \, \label{compl2}\] which shows that the present quantum associative retrieval offers an exponential speed-up with respect to the Grover-based one, contrary to what asserted in. Of course there is a trade-off between complexity and accuracy: the higher \(b\) the more the retrieval probability ([\[prob\]](#prob){reference-type="ref" reference="prob"}) is peaked on the correct pattern but the higher the computational complexity. In conclusion, the quantum associative memory originally proposed in, with the improvements proposed here, is not limited by the no-cloning theorem, can correct corrupted inputs on top of completing partial ones and offers an exponential speed-up with respect to the Grover-based alternative. | {'timestamp': '2022-06-06T02:16:38', 'yymm': '2206', 'arxiv_id': '2206.01644', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01644'} |
# [\[sec:level1\]]{#sec:level1 label="sec:level1"}Introduction
Near a continuous phase transition, microscopically different systems may display low-energy, long-distance behaviour that is remarkably similar. Such universality is explained by saying that those systems flow to the same fixed point of the renormalization group (RG). In 1+1 dimensions, the universality class of a (massless) RG fixed point is often described by a conformal field theory (CFT) which, in turn, is completely characterized by a set of parameters known as the conformal data. In this work we are concerned with critical quantum spin chains, as specified in terms of a local lattice Hamiltonian in one spatial dimension. Given one such Hamiltonian, a natural goal is to numerically compute the conformal data of the emergent 1+1 CFT that describes the behaviour of the spin chain at long distances and low energies. In other words, to numerically characterize the emergent universal behaviour of the phase transition realized by the quantum spin chain. For this purpose, one may follow an approach, initiated in the 1980's, that is based on the *operator-state correspondence*. Given a 1+1 CFT on the circle, this correspondence establishes that each simultaneous eigenstate \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) of the Hamiltonian \(H^{\textsl{\tiny CFT}}\) and momentum \(P^{\textsl{\tiny CFT}}\) of the CFT on the circle corresponds to a scaling operator \(\psi^{\textsl{\tiny CFT}}_\alpha\) (an operator that transforms covariantly under scale transformations). Moreover, the energy and momentum of the state \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) relate to the universal scaling properties of the operator \(\psi^{\textsl{\tiny CFT}}_\alpha\). Importantly, the low energy states of a critical quantum spin chain on the circle are organized as in the emergent CFT. Accordingly, Cardy and others proposed that one could extract the conformal data by simply studying the low energy states of the critical quantum spin chain. Over the years, several other authors have contributed additional insights into this strategy. A crucial step was to identify operators on the lattice with operators of the CFT. In particular, certain lattice operators can be identified with Virasoro generators of the conformal symmetry in the CFT. This means that such lattice operators act on the low-energy states of the critical quantum spin chain approximately (up to finite size corrections) in the same way as the Virasoro generators act on the corresponding states in the CFT. The lattice Virasoro generators allow us to see the emergence of conformal symmetry directly on the lattice, offering a way to systematically identify primary operators and conformal towers in the low-energy spectrum of a quantum critical spin chain. In some CFTs, conformal symmetry is enhanced to a larger symmetry, which further relates different Virasoro primary operators and conformal towers. Extended symmetries exist in many CFTs that describe lattice models, such as quantum spin chains, classical statistical-mechanics models, and edge modes of topological orders. One important implication of the extended symmetry is that it reduces the number of independent primary operators in the conformal data. In the cases known as rational CFTs, the number of primary operators with respect to the extended symmetry is finite, while the number of Virasoro primary fields can be infinite. This makes the extraction of complete conformal data possible. A prominent example is when conformal symmetry is enhanced by a global symmetry with Lie group \(G\). In this case, scaling operators are organized by an extension of the Virasoro algebra, the Kac--Moody algebra, denoted by \(\mathfrak{g}_k\), where \(\mathfrak{g}\) is the Lie algebra of the the group \(G\) and \(k\in \mathbb{Z}\) is the *level constant*. A remarkable consequence of the extension is that the global symmetry of the CFT becomes \(G\times G\), which acts independently on left and right moving fields. In this paper we investigate how the Kac--Moody symmetry emerges from critical quantum spin chains with a Lie group symmetry \(G\). We construct lattice operators that correspond to the generators of the Kac--Moody algebra. We numerically confirm that the global symmetry is enhanced to \(G\times G\) in the low-energy subspace and that the eigenstates can be organized into Kac--Moody towers. We test our proposal with the XXZ model for the \(G=U(1)\) case and the Heisenberg model with next-to-nearest-neighbor interactions for the \(G=SU(2)\) case. In both cases we find that the proposed lattice Kac--Moody generators act on the low-energy states as their CFT counterparts in the thermodynamic limit. In particular, we can extract the level constant \(k\) by computing the commutators of lattice Kac--Moody generators in the low-energy subspace. In previous work by some of the authors, lattice operators that correspond to generators of another form of extended symmetry, namely superconformal symmetry, were found numerically. Here, in contrast, we provide an analytical ansatz for the lattice Kac--Moody generators. For the specific case of the XXZ model, which is an integrable model, the lattice Kac--Moody generators can also be found via bosonization (as reviewed in Appendix [\[appendix:a\]](#appendix:a){reference-type="ref" reference="appendix:a"}), and we can use this previous result to further validate our proposal. In contrast, no previous results appear to be known for the Heisenberg model with next-to-near-neighbor interactions. The performance of our method, which treats all models on the same footing, is seen to not rely on integrablity. The rest of the paper is organized as follows. In Sec. [8](#sec:review){reference-type="ref" reference="sec:review"}, we review the conformal symmetry and the Virasoro algebra in a CFT, as well as the approximate lattice version of Virasoro generators. In Sec. [9](#sec:u1){reference-type="ref" reference="sec:u1"}, we consider the Kac--Moody algebra for the Abelian group \(G=U(1)\). We first review the free boson CFT and the action of the Kac--Moody algebra, and then propose our method to build an approximate lattice version of the Kac--Moody generators. We also present numerical tests of the lattice Kac--Moody generators in the XXZ quantum spin model. In Sec. [10](#sec:su2){reference-type="ref" reference="sec:su2"} we generalize the formalism to general non-Abelian group \(G\). We will focus on \(G=SU(2)\) and construct an approximate lattice version of the Kac--Moody generators for the Heisenberg model with next-to-nearest-neighbor interactions. In Sec. [11](#sec:discussion){reference-type="ref" reference="sec:discussion"} we conclude with discussions and future directions.
# Conformal symmetry and lattice Virasoro generators {#sec:review}
In this section we review conformal symmetry, the Virasoro algebra and its approximate realization in the low-energy states of a quantum critical spin chain, see Ref. and Ref. for an introduction. Throughout this paper, we use superscript \(^\textsl{\tiny CFT}\) to denote objects in a CFT, e.g. \(T^\textsl{\tiny CFT}\), \(L_n^\textsl{\tiny CFT}\), \(\J{n}{\alpha}\), and in this way distinguish them from the corresponding objects in a lattice, e.g. \(T\), \(L_n\), \(J_n^{\alpha}\).
## Virasoro algebra
Consider a conformal field theory in 1+1 dimensions. Conformal transformations on the plane are generated by Virasoro generators \(L^{\textsl{\tiny CFT}}_n,\bar{L}^{\textsl{\tiny CFT}}_n~(n\in\mathbb{Z})\) that satisfy the Virasoro algebra, \[\begin{aligned}
\label{VirasoroCFT} &[L^{\textsl{\tiny CFT}}_n,L^{\textsl{\tiny CFT}}_m]=(n-m) L^{\textsl{\tiny CFT}}_{n+m} + \frac{c^{\textsl{\tiny CFT}}}{12}n (n^2-1) \delta_{n+m,0} \\ &[\bar{L}^{\textsl{\tiny CFT}}_n,\bar{L}^{\textsl{\tiny CFT}}_m]=(n-m) \bar{L}^{\textsl{\tiny CFT}}_{n+m} + \frac{c^{\textsl{\tiny CFT}}}{12}n (n^2-1) \delta_{n+m,0} \\ &[L^{\textsl{\tiny CFT}}_n,\bar{L}^{\textsl{\tiny CFT}}_m]=0. \end{aligned}\] In particular, dilations and rotations of the plane are generated by \(D^{\textsl{\tiny CFT}}=L^{\textsl{\tiny CFT}}_0+\bar{L}^{\textsl{\tiny CFT}}_0\) and \(R^{\textsl{\tiny CFT}}=L^{\textsl{\tiny CFT}}_0-\bar{L}^{\textsl{\tiny CFT}}_0\), respectively. The Hilbert space of the CFT is supported on circles around the origin. Any state in the Hilbert space can be spanned by simultaneous eigenstates \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) of \(L^{\textsl{\tiny CFT}}_0\) and \(\bar{L}^{\textsl{\tiny CFT}}_0\), \[L^{\textsl{\tiny CFT}}_0|\psi^{\textsl{\tiny CFT}}_\alpha\rangle=h^{\textsl{\tiny CFT}}_\alpha|\psi^{\textsl{\tiny CFT}}_\alpha\rangle, ~~ \bar{L}^{\textsl{\tiny CFT}}_0|\psi^{\textsl{\tiny CFT}}_\alpha\rangle=\bar{h}^{\textsl{\tiny CFT}}_\alpha|\psi^{\textsl{\tiny CFT}}_\alpha\rangle,\] where \((h^{\textsl{\tiny CFT}}_\alpha,\bar{h}^{\textsl{\tiny CFT}}_\alpha)\) are the holomorphic and anti-holomorphic conformal dimensions of the state. The eigenvalues of \(D^{\textsl{\tiny CFT}}\) and \(R^{\textsl{\tiny CFT}}\) are \(\Delta^{\textsl{\tiny CFT}}_\alpha=h^{\textsl{\tiny CFT}}_\alpha+\bar{h}^{\textsl{\tiny CFT}}_\alpha\) and \(s^{\textsl{\tiny CFT}}_\alpha=h^{\textsl{\tiny CFT}}_\alpha-\bar{h}^{\textsl{\tiny CFT}}_\alpha\), known as scaling dimensions and conformal spins, respectively. Setting \(n=0\) in Eq. [\[VirasoroCFT\]](#VirasoroCFT){reference-type="eqref" reference="VirasoroCFT"}, we see that acting with \(L^{\textsl{\tiny CFT}}_m\) on a state changes the holomorphic dimension by \(-m\). Therefore, \(L^{\textsl{\tiny CFT}}_m\) lowers the scaling dimensions and the conformal spins by \(m\). Similarly, \(\bar{L}^{\textsl{\tiny CFT}}_m\) lowers the scaling dimensions by \(m\) and increases the conformal spins by \(m\). A Virasoro primary state \(|\phi^{\textsl{\tiny CFT}}_\alpha\rangle\) is defined such that its scaling dimension cannot be lowered, \[\label{eq:primarydef} L^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \bar{L}^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \forall n>0.\] Each primary state is associated with a Virasoro conformal tower, which contains that primary state as well as all its Virasoro descendant states. The descendant states are obtained by acting successively with raising operators \(L^{\textsl{\tiny CFT}}_m,\bar{L}^{\textsl{\tiny CFT}}_m~(m<0)\) on the primary state.
## CFT on the cylinder
In 1+1 dimensions one can use a conformal transformation to map the CFT from the plane to a cylinder \(S^{1}\times\mathbb R\), where the axial direction is the imaginary time direction with coordinate \(\tau\in(-\infty,\infty)\), and the angular direction is the spatial direction with coordinate \(x\in [0,L)\). The Hilbert space is supported on the equal-time slices. We will focus on the \(\tau=0\) slice, and denote the fields with its spatial coordinate \(x\). In any CFT, there exist the holomorphic and anti-holomorphic energy-momentum tensors \(T^\textsl{\tiny CFT}\) and \(\bar{T}^\textsl{\tiny CFT}\), with conformal dimensions \((2,0)\) and \((0,2)\), respectively. The Hamiltonian and momentum can be expressed as an integral of the stress tensor, \[\begin{aligned}
H^{\textsl{\tiny CFT}} &=& \int_0^{L} dx\, h^\textsl{\tiny CFT}(x) \\ P^{\textsl{\tiny CFT}} &=& \int_0^{L} dx\, p^\textsl{\tiny CFT}(x), \end{aligned}\] where \[\begin{aligned}
h^\textsl{\tiny CFT}(x) &=& \frac{1}{2\pi}(T^\textsl{\tiny CFT}(x)+\bar{T}^\textsl{\tiny CFT}(x)) \\ p^\textsl{\tiny CFT}(x) &=& \frac{1}{2\pi}(T^\textsl{\tiny CFT}(x)-\bar{T}^\textsl{\tiny CFT}(x)). \end{aligned}\] The Virasoro generators are the Fourier modes of \(T^\textsl{\tiny CFT}\) and \(\bar{T}^\textsl{\tiny CFT}\), \[\begin{aligned} \label{eq:LnCFT} L^{\textsl{\tiny CFT}}_n = \frac{L}{(2\pi)^2} \int_{0}^{L} dx e^{+ i n x \frac{2\pi}{L}} T^\textsl{\tiny CFT} (x) + \frac{c^{\textsl{\tiny CFT}}}{24} \delta_{n,0}, \\ \bar{L}^{\textsl{\tiny CFT}}_n = \frac{L}{(2\pi)^2} \int_{0}^{L} dx e^{-i n x \frac{2\pi}{L}} \bar{T}^\textsl{\tiny CFT} (x) + \frac{c^{\textsl{\tiny CFT}}}{24} \delta_{n,0}. \end{aligned}\] We may express the Hamiltonian and momentum as \[\begin{aligned}
\label{HCFT} H^{\textsl{\tiny CFT}}&=&\frac{2\pi}{L}\left(L^{\textsl{\tiny CFT}}_0+\bar{L}^{\textsl{\tiny CFT}}_0-\frac{c^{\textsl{\tiny CFT}}}{12}\right) \\ \label{PCFT} P^{\textsl{\tiny CFT}}&=& \frac{2\pi}{L}\left(L^{\textsl{\tiny CFT}}_0-\bar{L}^{\textsl{\tiny CFT}}_0\right). \end{aligned}\] We see that each simultaneous eigenstate \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) of \(H^{\textsl{\tiny CFT}}\) and \(P^{\textsl{\tiny CFT}}\) has energy and momentum \[\begin{aligned}
\label{eq:ECFT} E^{\textsl{\tiny CFT}}_\alpha&=&\frac{2\pi}{L}\left(\Delta^{\textsl{\tiny CFT}}_\alpha-\frac{c^{\textsl{\tiny CFT}}}{12}\right) \\ \label{eq:PCFT} P^{\textsl{\tiny CFT}}_\alpha&=& \frac{2\pi}{L} s^{\textsl{\tiny CFT}}_\alpha. \end{aligned}\] It is also useful to express the Fourier mode of the Hamiltonian density as a linear combination of Virasoro generators \(L_n^{\textsl{\tiny CFT}}\) and \(\bar{L}_{-n}^{\textsl{\tiny CFT}}\), \[L^{\textsl{\tiny CFT}}_n+\bar{L}^{\textsl{\tiny CFT}}_{-n}= H^{\textsl{\tiny CFT}}_n \equiv \frac{L}{2\pi} \int_{0}^{L} dx e^{+ i n x \frac{2\pi}{L}} h^\textsl{\tiny CFT} (x).\] These operators transform states within the same conformal tower of the CFT.
## Lattice Virasoro generators
Given a critical quantum spin chain on a circle with Hamiltonian \[H=\sum_{j} h_j,\] its low-energy eigenstates \(|\psi_\alpha\rangle\) are in one-to-one correspondence with CFT states \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\). The energies and momenta of \(|\psi_\alpha\rangle\) are related to the scaling dimensions and conformal spins by \[\begin{aligned}
\label{eq:Elat} E_\alpha &=& A+B\frac{2\pi}{N}\left(\Delta^{\textsl{\tiny CFT}}_\alpha-\frac{c^{\textsl{\tiny CFT}}}{12}\right)+O(N^{-x}) \\ \label{eq:Plat} P_\alpha &=& \frac{2\pi}{N}s^{\textsl{\tiny CFT}}_\alpha, \end{aligned}\] where the \(O(N^{-x})\) with \(x>1\) is the non-universal finite-size correction. These relations are direct lattice versions of expressions Eqs. [\[eq:ECFT\]](#eq:ECFT){reference-type="eqref" reference="eq:ECFT"}-[\[eq:PCFT\]](#eq:PCFT){reference-type="eqref" reference="eq:PCFT"} for the CFT, in that upon identifying the circle size \(L\) with the size \(N\) of the spin chain, we could rewrite them as \(E_{\alpha} = A + BE_{\alpha} + O(N^{-x})\) and \(P_{\alpha} = P_{\alpha}^{\textsl{\tiny CFT}}\). From Eqs. [\[eq:Elat\]](#eq:Elat){reference-type="eqref" reference="eq:Elat"}-[\[eq:Plat\]](#eq:Plat){reference-type="eqref" reference="eq:Plat"}, one may extract approximate scaling dimensions \(\Delta_\alpha\) and exact conformal spins \(s_\alpha\) from the low-energy spectrum of the spin chain. In particular, the constants \(A,B\) (which depend on how the lattice Hamiltonian is normalized) may be determined by demanding that the scaling dimensions of the identity operator and the stress tensor be \(\Delta_{\mathbf{1}}=\Delta^{\textsl{\tiny CFT}}_{\mathbf{1}}=0\) and \(\Delta_{T}=\Delta^{\textsl{\tiny CFT}}_T=2\). Notice that the states \(\ket{\mathbf{1}^{\textsl{\tiny CFT}}}\) and \(\ket{T^{\textsl{\tiny CFT}}}\) are present in any CFT, so the above procedure to determine constants \(A\) and \(B\) can always be applied. The lattice operator \(H_n\) that corresponds to \(H^{\textsl{\tiny CFT}}_n\) is the Fourier mode of the lattice Hamiltonian density \(h_j\), \[H_n=\frac{N}{B}\sum_{j} e^{inj \frac{2\pi}{N}} h_j \sim H^{\textsl{\tiny CFT}}_n.\] There is both analytical and numerical evidence that \(H_n\) acts on the low-energy eigenstates of the lattice Hamiltonian as \(H^{\textsl{\tiny CFT}}_n\) acts on the corresponding CFT states. We stress that the lattice Virasoro generators \(H_n\) only (approximately) satisfy the algebra obeyed by the CFT operators \(H^{\textsl{\tiny CFT}}_n\) when projected onto the low-energy eigenstates of the quantum spin chains. At the level of lattice operators, it is easily checked (e.g. numerically) that the \(H_n\) operators do not satisfy the Virasoro algebra. Similar to the \(O(N^{-x})\) term in Eq. [\[eq:Elat\]](#eq:Elat){reference-type="eqref" reference="eq:Elat"}, there are also non-universal finite-size corrections in the matrix elements of \(H_n\), which can be reduced by an extrapolation to the thermodynamic limit (that is, the limit of an inifnitely large spin chain, \(N \rightarrow \infty\)). Since the lattice Virasoro generators \(H_n\) connect low-energy eigenstates within the same conformal tower, they can be used to identify Virasoro primary states and their conformal towers on the lattice.
# Lattice realization of Kac--Moody algebra: the \(U(1)\) case {#sec:u1}
In this section we first review the \(U(1)\) Kac--Moody algebra and its manifestation in the free compactified boson CFT. We then consider a critical quantum spin chain with \(U(1)\) symmetry and construct a lattice version of the Kac--Moody generators. Finally we numerically verify the actions of the proposed lattice Kac--Moody generators in the XXZ model.
## \(U(1)\) Kac--Moody algebra
We consider a CFT with a global \(U(1)\) symmetry, with conserved \(U(1)\) charge \(Q^{\textsl{\tiny CFT}}\), which is an integral over space of a conserved local current \(q^{\textsl{\tiny CFT}}(x)\), \[\label{eq:U1QCFT} Q^{\textsl{\tiny CFT}}=\int dx\, q^{\textsl{\tiny CFT}}(x).\] In a CFT, it turns out that the conserved current can be further divided into holomorphic and anti-holomorphic parts which are separately conserved, \[\label{eq:U1qCFT} q^{\textsl{\tiny CFT}}(x)=J^{\textsl{\tiny CFT}}(x)+\bar{J}^{\textsl{\tiny CFT}}(x),\] where \(J^{\textsl{\tiny CFT}}\) and \(\bar{J}^{\textsl{\tiny CFT}}\) are *current operators* with conformal dimensions \((1,0)\) and \((0,1)\), respectively. Since both current operators satisfy conservation laws, namely \(\bar{\partial} J^{\textsl{\tiny CFT}}=0\) and \(\partial \bar{J}^{\textsl{\tiny CFT}}=0\), there is an additional conserved charge \[\label{eq:U1MCFT} M^{\textsl{\tiny CFT}}=\int dx\, m^{\textsl{\tiny CFT}}(x),\] where \[\label{eq:U1mCFT} m^{\textsl{\tiny CFT}}(x)=J^{\textsl{\tiny CFT}}(x)-\bar{J}^{\textsl{\tiny CFT}}(x).\] The global symmetry is therefore \(U(1)\times U(1)\). For later reference, let us combine Eqs. [\[eq:U1qCFT\]](#eq:U1qCFT){reference-type="eqref" reference="eq:U1qCFT"}-[\[eq:U1mCFT\]](#eq:U1mCFT){reference-type="eqref" reference="eq:U1mCFT"} and the conservation relations above, to obtain \[-i\partial_\tau q^{\textsl{\tiny CFT}}(x)=\partial_x m^{\textsl{\tiny CFT}}(x).\] Then, upon substituting the time derivative with the commutator by the Hamiltonian we further obtain \[\label{eq:U1qmCFT} i [H^{\textsl{\tiny CFT}},q^{\textsl{\tiny CFT}}(x)]=\partial_x m^{\textsl{\tiny CFT}}(x).\] Similar to Eq. [\[eq:LnCFT\]](#eq:LnCFT){reference-type="eqref" reference="eq:LnCFT"}, which defined the Virasoro generators, we may define the generators of the Kac--Moody algebra by \[\begin{aligned}
\label{eq:U1JnCFT} J^{\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{+inx\frac{2\pi}{L}} J^{\textsl{\tiny CFT}}(x) \\ \bar{J}^{\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{-inx\frac{2\pi}{L}} \bar{J}^{\textsl{\tiny CFT}}(x).\label{eq:U1JnCFTb} \end{aligned}\] They satisfy the \(\mathfrak{u}(1)_k\) Kac--Moody algebra for some value \(k\in \mathbb{Z}\), known as the level constant, \[\label{eq:JmJn} \begin{aligned} &[J^{\textsl{\tiny CFT}}_m, J^{\textsl{\tiny CFT}}_n] = km \delta_{m+n,0},\\ &[\bar{J}^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = km \delta_{m+n,0},\\ &[J^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = 0. \end{aligned}\] The Virasoro generators and Kac--Moody generators satisfy the commutation relations \[\label{eq:LmJn} \begin{aligned} [L^{\textsl{\tiny CFT}}_m, J^{\textsl{\tiny CFT}}_n] =-n J^{\textsl{\tiny CFT}}_{m+n}, \\ [\bar{L}^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] =-n \bar{J}^{\textsl{\tiny CFT}}_{m+n}, \end{aligned}\] which is compatible with the fact that \(J^{\textsl{\tiny CFT}}\) is a Virasoro primary operator with conformal dimensions \((1,0)\). Setting \(m=0\) in Eq. [\[eq:LmJn\]](#eq:LmJn){reference-type="eqref" reference="eq:LmJn"}, we see that \(J^{\textsl{\tiny CFT}}_m\) changes the holomorphic dimension by \(-m\), as what \(L^{\textsl{\tiny CFT}}_m\) does. Therefore, \(J^{\textsl{\tiny CFT}}_{m}\) is a raising operator with negative \(m\), and a lowering operator with positive \(m\). When \(m=0\), \(J^{\textsl{\tiny CFT}}_{0}\) commutes with \(L^{\textsl{\tiny CFT}}_0\) and they have the same eigenstates. This reflects the fact that \(J^{\textsl{\tiny CFT}}(x)\) is a conserved current. In analogy with the definition of Virasoro primary states in Eq. [\[eq:primarydef\]](#eq:primarydef){reference-type="eqref" reference="eq:primarydef"}, we define Kac--Moody primary states, \[J^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \bar{J}^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \forall n>0.\] A Kac--Moody tower consists of a Kac--Moody primary state \(|\phi^{\textsl{\tiny CFT}}_\alpha\rangle\) with scaling dimension \(\Delta_\alpha\) and conformal spin \(s_\alpha\) and their descendent states, obtained from the primary state by sequentially acting with the generators. For instance, examples of Kac--Moody descendants include \[|\phi^{\textsl{\tiny CFT}}_\beta\rangle = J^{\textsl{\tiny CFT}}_{-n} |\phi^{\textsl{\tiny CFT}}_\alpha\rangle, ~~ |\phi^{\textsl{\tiny CFT}}_\beta\rangle = \bar{J}^{\textsl{\tiny CFT}}_{-n} |\phi^{\textsl{\tiny CFT}}_\alpha\rangle, ~~ \text{for} \ n > 0.\] with \[\label{eq:ladder} \Delta_{\beta} = \Delta_{\alpha} + n, \quad s_{\beta} = s_{\alpha} \pm n, \quad \text{for} \ n > 0.\] It can be shown that all Kac--Moody primary states are also Virasoro primary states. However, there are Virasoro primary states that are not Kac--Moody primary states. For example, \[\begin{aligned}
|J^{\textsl{\tiny CFT}}\rangle &=& J^{\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle \\ |\bar{J}^{\textsl{\tiny CFT}}\rangle &=& \bar{J}^{\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle \\ |J\bar{J}^{\textsl{\tiny CFT}}\rangle &=& J^{\textsl{\tiny CFT}}_{-1}\bar{J}^{\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle, \end{aligned}\] that is, the states corresponding to the holomorphic current \(J^{\textsl{\tiny CFT}}\), the anti-holomophic current \(\bar{J}^{\textsl{\tiny CFT}}\) and their composite operator \(J\bar{J}^{\textsl{\tiny CFT}}\), which can be seen to be Virasoro primary states, all belong to the same Kac--Moody tower.
## Free compactified boson CFT
A simple example of CFT that has Kac--Moody symmetry is a free compactified boson CFT, with action \[S^{\textsl{\tiny CFT}}=\frac{1}{2}\int d\tau dx \, \left[ (\partial_\tau \phi)^2+(\partial_x \phi)^2 \right],\] where \(\phi\) is identified with \(\phi+2\pi R\), and \(R\) is the *compactification radius*. The holomorphic and anti-holomorphic conserved currents read \[J^{\textsl{\tiny CFT}}=i\partial\phi,~~ \bar{J}^{\textsl{\tiny CFT}}=i\bar{\partial}\phi,\] where \(\partial=(\partial_\tau+ i\partial_x)/2\) and \(\bar{\partial}=(\partial_\tau-i\partial_x)/2\) are holomorphic and anti-holomorphic derivatives. In this CFT, the conservation relations \(\bar{\partial}J^{\textsl{\tiny CFT}}=0\) and \(\partial \bar{J}^{\textsl{\tiny CFT}}=0\) follow from the equation of motion \(\partial\bar{\partial}\phi=0\). Now, using the definition of Kac--Moody generators in Eqs. [\[eq:U1JnCFT\]](#eq:U1JnCFT){reference-type="eqref" reference="eq:U1JnCFT"}-[\[eq:U1JnCFTb\]](#eq:U1JnCFTb){reference-type="eqref" reference="eq:U1JnCFTb"} and the canonical commutation \([\phi(x),\partial_\tau\phi(y)]=\delta(x-y)\), we can verify the Kac--Moody algebra, \[\label{eq:u1comm} \begin{aligned} &[J^{\textsl{\tiny CFT}}_m, J^{\textsl{\tiny CFT}}_n] = [\bar{J}^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = m \delta_{m+n,0},\\ &[J^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = 0, \end{aligned}\] which corresponds level constant \(k=1\). The free compactified boson CFT has infinitely many primary states with respect to the \(\mathfrak{u}(1)_1\) Kac--Moody algebra. They are vertex operators \(V^{\textsl{\tiny CFT}}_{Q,M}\) labelled by integers \(Q\) and \(M\), with scaling dimensions and conformal spins \[\Delta^{\textsl{\tiny CFT}}_{V_{Q,M}} = \frac{Q^2}{R^2} + \frac{R^2M^2}{4}, \quad S^{\textsl{\tiny CFT}}_{V_{Q,M}} = QM, \quad Q, M \in \mathbb{Z},\] where \(Q\) and \(M\) are the eigenvalues of \(Q^{\textsl{\tiny CFT}}\) and \(M^{\textsl{\tiny CFT}}\), respectively.
## Lattice Kac--Moody generators
Consider now a critical quantum spin chain on the circle with Hamiltonian \(H\) and global \(U(1)\) symmetry (see e.g. the XXZ model below). Our goal is to construct lattice operators \(J_n,\bar{J}_m\) that correspond to the Kac--Moody generators \(J^{\textsl{\tiny CFT}}_n,\bar{J}^{\textsl{\tiny CFT}}_m\). On the lattice, let the conserved \(U(1)\) charge be \[Q=\sum_{j} q_j,\] which commutes with the Hamiltonian \([H, Q] =0\). We identify \(Q\) with the CFT charge operator \(Q^{\textsl{\tiny CFT}}\) in Eq. [\[eq:U1QCFT\]](#eq:U1QCFT){reference-type="eqref" reference="eq:U1QCFT"}, and \(q\) with the CFT current density \(q^{\textsl{\tiny CFT}}\) in Eq. [\[eq:U1qCFT\]](#eq:U1qCFT){reference-type="eqref" reference="eq:U1qCFT"}. In order to find the Kac--Moody generators, one also needs to find the lattice operator \(m\) corresponding to \(m^{\textsl{\tiny CFT}}\) in Eq. [\[eq:U1mCFT\]](#eq:U1mCFT){reference-type="eqref" reference="eq:U1mCFT"}. Our proposal is that the corresponding lattice operators should satisfy an equation analogous to Eq. [\[eq:U1qmCFT\]](#eq:U1qmCFT){reference-type="eqref" reference="eq:U1qmCFT"}, after replacing the spatial derivative \(\partial_x\) with a finite difference, see Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"} below. Then by Eqs. [\[eq:U1qCFT\]](#eq:U1qCFT){reference-type="eqref" reference="eq:U1qCFT"}-[\[eq:U1mCFT\]](#eq:U1mCFT){reference-type="eqref" reference="eq:U1mCFT"}, the lattice operators \(J_j\) and \(\bar{J}_j\) that correspond to \(J^{\textsl{\tiny CFT}}\) and \(\bar{J}^{\textsl{\tiny CFT}}\) can be identified as in Eq. [\[eq:U1latticeJi\]](#eq:U1latticeJi){reference-type="eqref" reference="eq:U1latticeJi"} below. Finally the lattice operators \(J_n\) and \(\bar{J}_n\) corresponding to the Kac--Moody generators are constructed as Fourier modes in Eq. [\[eq:U1J\]](#eq:U1J){reference-type="eqref" reference="eq:U1J"}.
Eqs.[\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"}, [\[eq:U1latticeJi\]](#eq:U1latticeJi){reference-type="eqref" reference="eq:U1latticeJi"} and [\[eq:U1J\]](#eq:U1J){reference-type="eqref" reference="eq:U1J"} are the main proposal of this paper (together with their generalization below to the non-Abelian case). Let us elaborate on the form of \(m_j\) when \(H\) is a nearest-neighbor and a next-to-nearest-neighbor Hamiltonian. These specializations will be useful in the applications to specific lattice models.
### Nearest-neighbor Hamiltonian
Consider a nearest-neighbor Hamiltonian \(H = \sum_{j} h_{j,j+1}\) with a global symmetry that is realized on-site with charge \(Q = \sum_{j} q_j\). If follows from \([H, Q]=0\) that at for any pair \((j, j+1)\) of contiguous sites, \[= 0.\] The LHS of Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"} then becomes \[\begin{aligned} i [H, q_j] &= i ([h_{j-1,j}, q_j] + [h_{j, j+1}, q_j]) \\ &= i (-[h_{j-1,j}, q_{j-1}] + [h_{j, j+1}, q_j]), \end{aligned}\] and matching it with the RHS of Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"}, we obtain \[\label{eq:2local} m_j = i[h_{j,j+1}, q_j],\] which is our proposal as a lattice version of the locally conserved current \(m^{\textsl{\tiny CFT}}(x)\).
### Next-to-nearest-neighbor Hamiltonian
As a second example, consider a next-to-nearest-neighbor Hamiltonian \(H = \sum_j h_{j-1, j, j+1}\), again with a global symmetry that is realized on-site with charge \(Q = \sum_{j} q_j\). To simplify the notation, in the following \(h_{(j)} \coloneqq h_{j-1,j,j+1}\). At any set \((j-1,j,j+1)\) of three consecutive sites, \[= 0,\] we can check that \[\begin{aligned} i\left[H, \frac{q_{j}+q_{j+1}}{2}\right] = \frac{1}{2}(m_{j+1}-m_{j-1}) \end{aligned}\] where \[\label{eq:3local} m_{j} = i([h_{(j+1)}, q_j]-[h_{(j)}, q_{j+1}]).\] This is a discrete version of Eq. [\[eq:U1qmCFT\]](#eq:U1qmCFT){reference-type="eqref" reference="eq:U1qmCFT"} in the case of next-to-nearest-neighbor Hamiltonian. Therefore we identify \(m_j\) in Eq. [\[eq:3local\]](#eq:3local){reference-type="eqref" reference="eq:3local"} with \(m^{\textsl{\tiny CFT}}(x)\). To reiterate, given a local Hamiltonian \(H\) with a microscopic global symmetry realized on-site, that is with charge \(Q=\sum_j q_j\), our proposal gives a concrete way of constructing the lattice current density \(m_j\) via Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"} and subsequently lattice Kac--Moody generators \(J_n\) and \(\bar{J}_n\) via Eq. [\[eq:U1latticeJi\]](#eq:U1latticeJi){reference-type="eqref" reference="eq:U1latticeJi"}-[\[eq:U1J\]](#eq:U1J){reference-type="eqref" reference="eq:U1J"}.
## Example: XXZ model
As a test of our construction of lattice current generators \(J_n\) and \(\bar{J}_{n}\) for a critical quantum spin chain with a global \(U(1)\) symmetry, we consider the XXZ model \[\label {eq: XXZ} H =-\frac{2\gamma}{\pi \sin{\gamma}} \sum_{j=1}^{N}\left( S_{j}^{X}S_{j+1}^{X} + S_{j}^{Y}S_{j+1}^{Y}-\cos{\gamma}S_{j}^{Z}S_{j+1}^{Z} \right)\] with anisotropy \(\gamma \in [0, \pi )\) and \(S^\alpha = \sigma^\alpha/2\), where \(\sigma^{x}, \sigma^y\) and \(\sigma^z\) stand for the Pauli matrices. This model displays indeed a global \(U(1)\) symmetry realized on-site, generated by the charge operator \(Q = \sum_{j=1}^{N} S_{j}^{Z}\), which commutes with the Hamiltonian, \([Q, H] = 0\). In the continuum limit the XXZ model is equivalent via bosonization to the Gaussian model of a free massless boson compactified on circle with radius \(R = \sqrt{2\pi/(\pi-\gamma)}\) and the symmetry algebra in the corresponding CFT theory is the Kac--Moody algebra with \(\mathfrak{g} = \mathfrak{u}(1)\). As discussed in the last section, we use Eq. [\[eq:2local\]](#eq:2local){reference-type="eqref" reference="eq:2local"} to compute the lattice current \(m_j\), \[\label{eq:u1m} m_j = \frac{2 \gamma}{\pi \sin{\gamma}} \left( S_{j}^{X}S_{j+1}^{Y}-S_{j}^{Y}S_{j+1}^{X} \right).\] We note that the lattice currents \(q_j\) and \(m_j\) for the XXZ model can also be understood using bosonization, as discussed in Appendix [\[appendix:a\]](#appendix:a){reference-type="ref" reference="appendix:a"}. Following Eq. [\[eq:J\]](#eq:J){reference-type="eqref" reference="eq:J"} the lattice versions of Kac--Moody generators can be constructed as \[\label{eq:Jnxxz} \begin{aligned} J_n = \sum_{j=1}^{N} e^{+ijn \frac{2\pi}{N}} \frac{1}{2} \left[ S_{j}^{Z}+ \frac{2\gamma}{\pi \sin{\gamma}} \left( S_{j}^{X}S_{j+1}^{Y}-S_{j}^{Y}S_{j+1}^{X} \right) \right], \\ \bar{J}_n = \sum_{j=1}^{N} e^{-ijn \frac{2\pi}{N}} \frac{1}{2} \left[ {S_{j}^{Z}}-\frac{2\gamma}{\pi \sin{\gamma}} \left( S_{j}^{X}S_{j+1}^{Y}-S_{j}^{Y}S_{j+1}^{X} \right) \right]. \end{aligned}\] To numerically check our proposal, we use exact diagonalization to find a set of low energy eigenstates \(\ket{\psi_\alpha}\). We simultaneously diagonalize the Hamiltonian and the translation operator with periodic boundary conditions to get the scaling dimensions \(\Delta_{\alpha}\) and conformal spins \(s_{\alpha}\). We also sort the low-energy eigenstates into different sectors according to the eigenvalues of \(Q=\sum_j S^Z_j\). As an example, in Fig. [\[fig:spectrum\]](#fig:spectrum){reference-type="ref" reference="fig:spectrum"} the low-energy spectra of XXZ spin chain of 20 sites with \(\gamma = \pi/2\) in the \(Q = 0\) sector and \(Q=1\) sector are plotted. That the above lattice generators \(J_n\) and \(\bar{J}_n\) approximately satisfy the Kac--Moody algebra at low energies can be verified by directly evaluating Eqs. [\[eq:JmJn\]](#eq:JmJn){reference-type="eqref" reference="eq:JmJn"}-[\[eq:LmJn\]](#eq:LmJn){reference-type="eqref" reference="eq:LmJn"} on low energy states. Those constitute an infinite set of conditions. In the following, we will demonstrate it for a subset of these conditions that are physically important. Other conditions can be confirmed similarly. From Eq. [\[eq:LmJn\]](#eq:LmJn){reference-type="eqref" reference="eq:LmJn"} for \(m=n=0\), we see that the charge \(M = J_0-\bar{J}_0\) should approximately satisfy \([M,H]=0\) and \([M,Q]=0\) when acting on low-energy states. While for general anisotropy \(\gamma\) the operator \(M=\sum_j m_j\) does not necessarily commute with the Hamiltonian, in the specific case of the XX model (\(\gamma = \pi/2\)), \(M \sim \sum_j S_j^X S_{j+1}^Y-S_j^Y S_{j+1}^X\) does commute with \(H\). In fact, there exists a series of conserved charges in the XX model. However \(M\) does not exactly commute with \(Q\) on the lattice. To confirm that the microscopic global \(U(1)\) symmetry turns indeed into an emergent \(U(1)\times U(1)\) symmetry at low energies, we check both the commutation relation \(\mel{\psi}{[Q,M]}{\psi}\) and the expectation value \(\mel{\psi}{M}{\psi}\) in the eigenbasis of \(Q\). For low energy eigenstates, we find that \(Q\) and \(M\) commute, up to a finite size error. We further show that the expectation value of the charge \(M\) approaches an integer value, see Fig. [\[fig:xx_m\]](#fig:xx_m){reference-type="ref" reference="fig:xx_m"}. This is compatible with the emergence of a global \(U(1)\times U(1)\) symmetry and allows us to identify vertex operators \(V_{Q, M}\) according to their eigenvalues of \(Q\) and \(M\). We can also examine the commutation relations in Eq. [\[eq:JmJn\]](#eq:JmJn){reference-type="eqref" reference="eq:JmJn"} in the energy-momentum eigenbasis by calculating matrix elements \(\mel{\psi}{[J_n, J_m]}{\psi}\). The level constant \(k\) can be extracted from \(\mel{\psi }{[J_n, J_{-n}]}{\psi}\), according to Eq. [\[eq:u1comm\]](#eq:u1comm){reference-type="eqref" reference="eq:u1comm"}. For example, we can estimate \(k\) using the three expectation values \(k_1 = \mel{\mathbf{1}}{[J_1, J_{-1}]}{\mathbf{1}}\), \(k_2 = \frac{1}{2} \mel{\mathbf{1}}{[J_2, J_{-2}]}{\mathbf{1}}\) and \(k_3 = \frac{1}{3} \mel{\mathbf{1}}{[J_3, J_{-3}]}{\mathbf{1}}\), as shown in Fig. [\[fig:kfit\]](#fig:kfit){reference-type="ref" reference="fig:kfit"}. Linear extrapolation gives \(k_1 = 1.0000\), \(k_2 = 0.9999\) and \(k_3 = 0.9992\), which are in good agreement with theoretical value \(k=1\). We check the action of \(J_n\) on low energy eigenstates by examining matrix elements \(\mel{\psi_\alpha}{ J_n }{ \psi_\beta}\) (similarly for the anti-chiral part \(\bar{J}_n\)). By numerically checking whether a state is annihilated by \(J_{1}, \bar{J}_{1}\) and \(J_{2}, \bar{J}_{2}\), up to finite size error, we successfully identify all the Kac--Moody primaries. In Fig. [\[fig:spectrum\]](#fig:spectrum){reference-type="ref" reference="fig:spectrum"}, the candidate primary states in the \(Q=0\) and \(Q=1\) sectors are marked. Compared with the analytical result of free boson CFT in Fig. [\[fig:boson\]](#fig:boson){reference-type="ref" reference="fig:boson"}, our lattice generators identify the primary states correctly. By acting on a primary state with the lattice generators \(J_{n}\) and \(\bar{J}_{n}\), we can obtain all the descendent states of that primary state. Up to finite size error, \(J_n\) and \(\bar{J}_n\) act as ladder operators-they change the scaling dimension and conformal spin as illustrated in the example of Eq. [\[eq:ladder\]](#eq:ladder){reference-type="eqref" reference="eq:ladder"}. We plot a few examples of their actions in Fig. [\[fig:spectrum\]](#fig:spectrum){reference-type="ref" reference="fig:spectrum"}. Together we are able to organize all low-energy eigenstates into different Kac--Moody towers. In addition to the case of XX model with \(\gamma = \pi/2\), our proposal can be verified for general XXZ model. We provide numerical results for some other values of \(\gamma\) in Appendix [\[appendix:c\]](#appendix:c){reference-type="ref" reference="appendix:c"}. Up to an overall normalization constant of \(J\) and \(\bar{J}\), we find consistently \(k=1\) and expectation values of \(M\) to be integers, confirming the Kac--Moody algebra.
# Lattice realization of Kac--Moody algebra: the \(SU(2)\) case {#sec:su2}
In this section we generalize the construction of lattice Kac--Moody generators to general Lie groups. We will first introduce the Kac--Moody algebra with general semisimple Lie groups. Then we consider a specific example, the free compactified boson with compactification radius \(R=\sqrt{2}\), which possesses an \(\mathfrak{su}(2)_1\) Kac--Moody symmetry. We then consider critical quantum spin chains with semisimple Lie group symmetry and construct approximate lattice Kac--Moody generators. Finally we present the numerical verification with the Heisenberg model with next-to-nearest neighbor interactions.
## Kac--Moody algebra with a non-Abelian group
We consider a CFT that has an internal semisimple Lie group symmetry \(G\), where the Lie algebra is denoted as \(\mathfrak{g}\). The extended symmetry is generated by *Kac--Moody generators* that form the *Kac--Moody algebra*, denoted as \(\hat{\mathfrak{g}}_k\), where \(k\) is the *level constant* of the Kac--Moody algebra. In a CFT with Kac--Moody algebra, there exists a set of chiral and anti-chiral current operators \(J^{\alpha,\textsl{\tiny CFT}}\) and \(\bar{J}^{\alpha,\textsl{\tiny CFT}}\) with conformal dimensions \((1,0)\) and \((0,1)\), where \(\alpha=1,2,\cdots,\mathrm{dim}~\mathfrak{g}\) labels different currents. The Kac--Moody generators are Fourier modes of the currents, \[\begin{aligned}
\label{eq:JnCFT} J^{\alpha,\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{+inx\frac{2\pi}{L}} J^{\alpha,\textsl{\tiny CFT}}(x) \\ \bar{J}^{\alpha,\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{-inx\frac{2\pi}{L}} \bar{J}^{\alpha,\textsl{\tiny CFT}}(x). \end{aligned}\] They satisfy the Kac--Moody algebra \(\hat{\mathfrak{g}}_k\), \[\label{eq:comrel} \begin{aligned} &[\J{m}{\alpha},\J{n}{\beta}] = i \sum_\gamma f^{\alpha \beta \gamma} \J{m+n}{\gamma} + k m \delta^{\alpha \beta} \delta_{m+n, 0}\\ &[\J{m}{\alpha}, \Jbar{n}{\beta}] = 0\\ &[\Jbar{m}{\alpha}, \Jbar{n}{\beta}] = i \sum_\gamma f^{\alpha \beta \gamma} \Jbar{m+n}{\gamma} + k m \delta^{\alpha \beta} \delta_{m+n, 0} \end{aligned}\] where \(f^{\alpha \beta \gamma}\) are the structure constants of the Lie algebra \(\mathfrak{g}\) and \(k\) is the level constant. Setting \(n=m=0\) in Eq. [\[eq:comrel\]](#eq:comrel){reference-type="eqref" reference="eq:comrel"}, we see that the zero modes of the currents form exactly the ordinary Lie algebra \(\mathfrak{g}\) of the global symmetry \(G\), \[= i \sum_\gamma f^{\alpha \beta \gamma} \J{0}{\gamma}. \label{eq:j0comrel}\] Since \(J^{\alpha,\textsl{\tiny CFT}}_0\) and \(\bar{J}^{\alpha,\textsl{\tiny CFT}}_0\) are separately conserved, the global symmetry is \(G\times G\), as it also happened in the Abelian \(U(1)\) case of the previous section. The holomorphic and anti-holomorphic currents transform into each other under spatial parity. We may consider the linear combinations of currents that have definite parity, \[\begin{aligned}
\label{eq:qCFT} q^{\alpha, \textsl{\tiny CFT}}(x) &\equiv& \J{}{\alpha}(x) + \Jbar{}{\alpha}(x) \\ \label{eq:mCFT} m^{\alpha, \textsl{\tiny CFT}}(x) &\equiv& \J{}{\alpha}(x)-\Jbar{}{\alpha}(x). \end{aligned}\] They correspond to charges, \[\begin{aligned}
Q^{\alpha, \textsl{\tiny CFT}} &\equiv& \J{0}{\alpha} + \Jbar{0}{\alpha} \\ M^{\alpha, \textsl{\tiny CFT}} &\equiv& \J{0}{\alpha}-\Jbar{0}{\alpha}. \end{aligned}\] From Eq. [\[eq:j0comrel\]](#eq:j0comrel){reference-type="eqref" reference="eq:j0comrel"}, it follows that charges \(Q^{\alpha, \textsl{\tiny CFT}}\) satisfy the Lie algebra \(\mathfrak{g}\) \[= i \sum_\gamma f^{\alpha \beta \gamma} Q^{\gamma, \textsl{\tiny CFT}}.\]
## Free compactified boson at \(R=\sqrt{2}\)
At \(R=\sqrt{2}\), the free compactified boson has three Virasoro primary operators with conformal dimension \((1,0)\), \[\begin{aligned}
J^{x,\textsl{\tiny CFT}} &=& \frac{1}{\sqrt{2}}(V^{\textsl{\tiny CFT}}_{1,1}+V^{\textsl{\tiny CFT}}_{-1,-1}) \\ J^{y,\textsl{\tiny CFT}} &=& \frac{1}{\sqrt{2}i}(V^{\textsl{\tiny CFT}}_{1,1}-V^{\textsl{\tiny CFT}}_{-1,-1}) \\ J^{z,\textsl{\tiny CFT}} &=& i\partial\phi. \end{aligned}\] Together with three anti-holomorphic currents (with similar expressions), their Fourier modes constitute the \(\mathfrak{su}(2)_1\) algebra, \[\label{eq:struct} \begin{aligned} &[J^{\alpha,\textsl{\tiny CFT}}_m, J^{\beta,\textsl{\tiny CFT}}_n] = \sum_{\gamma} i\sqrt{2}\epsilon^{\alpha\beta\gamma}J^{\gamma,\textsl{\tiny CFT}}_{n+m}+m \delta_{m+n,0} \delta^{\alpha \beta},\\ &[J^{\alpha,\textsl{\tiny CFT}}_m, \bar{J}^{\beta,\textsl{\tiny CFT}}_n] = 0 \\ &[\bar{J}^{\alpha,\textsl{\tiny CFT}}_m, \bar{J}^{\beta,\textsl{\tiny CFT}}_n] =\sum_{\gamma} i\sqrt{2}\epsilon^{\alpha\beta\gamma}\bar{J}^{\gamma,\textsl{\tiny CFT}}_{n+m}+m \delta_{m+n,0} \delta^{\alpha \beta}, \end{aligned}\] where we can read off the structure constants \(f^{\alpha\beta\gamma}=\sqrt{2}\epsilon^{\alpha\beta\gamma}\) and the level constant \(k=1\). For later convenience, we also define \[\begin{aligned} &J^{\pm,\textsl{\tiny CFT}}_n = \frac{1}{\sqrt{2}}(J^{x,\textsl{\tiny CFT}}_n\pm i J^{y,\textsl{\tiny CFT}}_n),\\ &J^{3, \textsl{\tiny CFT}}_n = \frac{1}{\sqrt{2}}J^{z,\textsl{\tiny CFT}}_n \end{aligned}\] While there can be infinite (but countable) Virasoro primary states, in the presence of additional symmetry, the number of primary states with respect to the extended symmetry can be finite. In the case of \(\mathfrak{su}(2)_1\), a Kac--Moody primary state \(\ket{j^\textsl{\tiny CFT}}\) satisfies \[\label{eq:hws} \begin{aligned} J_{n}^{\alpha, \textsl{\tiny CFT}} \ket{j^\textsl{\tiny CFT}} &= 0 \quad \text{for}\quad n>0, \\ J_0^{+, \textsl{\tiny CFT}} \ket{j^\textsl{\tiny CFT}} &= 0,\\ J_0^{3, \textsl{\tiny CFT}} \ket{j^\textsl{\tiny CFT}} &= j\ket{j^\textsl{\tiny CFT}}. \end{aligned}\] where \(j = 0, \frac{1}{2}\). The only two \(SU(2)\) level 1 Kac--Moody primary operators are the identity operator \(\mathbf{1}^{\textsl{\tiny CFT}}\) with \(\Delta^{\textsl{\tiny CFT}}=0\) and \(s^{\textsl{\tiny CFT}}=0\) and \(V^{\textsl{\tiny CFT}}_{1,0}\) with \(\Delta^{\textsl{\tiny CFT}}=1/2\) and \(s^{\textsl{\tiny CFT}}=0\). Other Virasoro primary operators can be obtained by acting with Kac--Moody generators on these primaries, as in the examples shown in Fig. [\[fig:XXX_CFT01\]](#fig:XXX_CFT01){reference-type="ref" reference="fig:XXX_CFT01"}. For instance, in the Kac--Moody tower of the identity, \(|V^{\textsl{\tiny CFT}}_{1,1}\rangle\) and \(|V^{\textsl{\tiny CFT}}_{2,0}\rangle\) can be obtained in the following way, \[\begin{aligned}
\label{eq:kmtower1} J^{+,\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle=|V^{\textsl{\tiny CFT}}_{1,1}\rangle \\ \label{eq:kmtower2} \bar{J}^{+,\textsl{\tiny CFT}}_{-1}|V^{\textsl{\tiny CFT}}_{1,1}\rangle=|V^{\textsl{\tiny CFT}}_{2,0}\rangle. \end{aligned}\] Other Virasoro primary operators with scaling dimensions \(1\) or \(2\) can be obtained in an analogous way. In the Kac--Moody tower of \(V_{1,0}\), \[\begin{aligned}
\label{eq:kmtower3} J^{+,\textsl{\tiny CFT}}_{0}|V^{\textsl{\tiny CFT}}_{-1,0}\rangle=|V^{\textsl{\tiny CFT}}_{0,1}\rangle \\ \label{eq:kmtower4} J^{-,\textsl{\tiny CFT}}_{0}|V^{\textsl{\tiny CFT}}_{1,0}\rangle=|V^{\textsl{\tiny CFT}}_{0,-1}\rangle \\ \label{eq:kmtower5} J^{+,\textsl{\tiny CFT}}_{-2}|V^{\textsl{\tiny CFT}}_{1,0}\rangle=|V^{\textsl{\tiny CFT}}_{2,1}\rangle \\ \label{eq:kmtower6} J^{-,\textsl{\tiny CFT}}_{-2}|V^{\textsl{\tiny CFT}}_{-1,0}\rangle=|V^{\textsl{\tiny CFT}}_{-2,-1}\rangle. \end{aligned}\] Other Virasoro primary operators with scaling dimension \(1/2\) or \(5/2\) can be obtained in a similar way. These relations can be justified by the operator product expansion of vertex operators (see appendix [\[appendix:b\]](#appendix:b){reference-type="ref" reference="appendix:b"}).
## Lattice Kac--Moody generators for general \(G\)
Consider a lattice model with Hamiltonian \(H\) and an on-site symmetry \(G\) with Lie algebra \(\mathfrak{g}\). The symmetry is associated with a set of conserved charges \[Q^{\alpha} =\sum_{i}^{N} q^{\alpha}_i, ~~\alpha=1,2,\cdots,\mathrm{dim}~\mathfrak{g}.\] The charges commute with the Hamiltonian, \([H, Q^{\alpha}] = 0\), and satisfy \[= i f^{\alpha \beta \gamma} Q^{\gamma}.\] The one-site lattice operator \(q^{\alpha}_j\) corresponds to the CFT operator \(q^{\alpha,\textsl{\tiny CFT}}(x)\) in Eq. [\[eq:qCFT\]](#eq:qCFT){reference-type="eqref" reference="eq:qCFT"}, because they are both conserved and have even spatial parity. As in the \(U(1)\) case, we need to find the lattice operator \(m^{\alpha}_j\) corresponding to \(m^{\alpha,\textsl{\tiny CFT}}(x)\) in Eq. [\[eq:mCFT\]](#eq:mCFT){reference-type="eqref" reference="eq:mCFT"}. The derivation that leads to Eq. [\[eq:U1qmCFT\]](#eq:U1qmCFT){reference-type="eqref" reference="eq:U1qmCFT"} goes through for each conserved current \(J^{\alpha,\textsl{\tiny CFT}}(x)\), and we obtain \[\label{eq:qmCFT} i [H^{\textsl{\tiny CFT}},q^{\alpha,\textsl{\tiny CFT}}(x)]=\partial_x m^{\alpha,\textsl{\tiny CFT}}(x),\] which on the lattice transforms to Eq. [\[eq:qm\]](#eq:qm){reference-type="eqref" reference="eq:qm"} below. Then analogous to Eqs. [\[eq:qCFT\]](#eq:qCFT){reference-type="eqref" reference="eq:qCFT"}-[\[eq:mCFT\]](#eq:mCFT){reference-type="eqref" reference="eq:mCFT"}, we can write down the lattice operators that correspond to the chiral current density \(J^{\alpha}_j\) and anti-chiral current density \(\bar{J}^{\alpha}_j\) as in Eq. [\[eq:latticeJi\]](#eq:latticeJi){reference-type="eqref" reference="eq:latticeJi"} below. Finally the lattice Kac--Moody generators \(J^{\alpha}_n\) and \(\bar{J}^{\alpha}_n\) can be constructed as the Fourier modes of \(J^{\alpha}_j\) and \(\bar{J}^{\alpha}_j\).
This completes the construction of lattice Kac--Moody generators for general symmetry \(G\).
## Example: XXX model with next-to-nearest neighbor coupling
Consider an antiferromagnetic spin-1/2 chain with next-to-nearest neighbor coupling \(J_c=0.241167\) \[H = \mathcal{N} \left( \sum_{j=1}^{N} \vec{S}_j \cdot \vec{S}_{j+1} + J_c \sum_{j=1}^N \vec{S_j} \cdot \vec{S}_{j+2} \right)\] where \(S^{\alpha}_j = \frac{1}{2} \sigma_{j}^{\alpha}\) and \(\alpha = x, y, z.\) Overall normalization factor \(\mathcal{N} = 0.856\) is determined by fixing the energy spectrum to be Eq. [\[eq:ECFT\]](#eq:ECFT){reference-type="eqref" reference="eq:ECFT"} in the large-\(N\) limit. This model has three exact conserved charges \(Q^\alpha = \sum_{j=1}^{N} q^\alpha_j = \sum_{j=1}^{N} \sqrt{2} S_{j}^{\alpha}\) which commute with the Hamiltonian: \([Q^{\alpha}, H] = 0\). These charges satisfy \([Q^{\alpha}, Q^{\beta}] = i f^{\alpha \beta \gamma} Q^{\gamma}\) with structure constants \(f^{\alpha \beta \gamma} = \sqrt{2} \epsilon^{\alpha \beta \gamma}\). In the continuum limit, this model corresponds to the \(SU(2)\) WZW model at level \(k=1\). To find the lattice Kac--Moody generators, first we calculate the current \(m^\alpha_j\) using Eq. [\[eq:3local\]](#eq:3local){reference-type="eqref" reference="eq:3local"} \[m_j^\alpha =-f^{\alpha \beta \gamma} \mathcal{N} \left( S_j^{\beta} S_{j+1}^{\gamma} + J_c S_j^{\beta} S_{j+2}^{\gamma} \right).\] Then following Eq. [\[eq:J\]](#eq:J){reference-type="eqref" reference="eq:J"}, we construct \(J_n^{\alpha}\) and \(\bar{J}_n^{\alpha}\) as follows: \[\begin{aligned} J_n^{\alpha} = \sum_{j=1}^{N_0} e^{+ijn \frac{2\pi}{N}} \frac{\sqrt{2}}{2} \left[ S_j^\alpha-\epsilon^{\alpha \beta \gamma} \mathcal{N} \left( S_j^{\beta} S_{j+1}^{\gamma} + J_c S_j^{\beta} S_{j+2}^{\gamma} \right) \right], \\ \bar{J}_n^{\alpha} = \sum_{j=1}^{N_0} e^{-ijn \frac{2\pi}{N}} \frac{\sqrt{2}}{2} \left[ S_j^\alpha + \epsilon^{\alpha \beta \gamma} \mathcal{N} \left( S_j^{\beta} S_{j+1}^{\gamma} + J_c S_j^{\beta} S_{j+2}^{\gamma} \right) \right]. \end{aligned}\] In order to confirm that they correspond to Kac--Moody generators in the continuum, we perform simultaneous exact diagonalization of the above lattice Hamiltonian \(H\) and of the lattice translation operator to obtain the low-energy eigenstates. Scaling dimensions and conformal spins can be computed using Eqs. [\[eq:Elat\]](#eq:Elat){reference-type="eqref" reference="eq:Elat"}-[\[eq:Plat\]](#eq:Plat){reference-type="eqref" reference="eq:Plat"}.
Now we check various matrix elements of the lattice Kac--Moody generators. We start by examining the zero modes of the lattice generators \(J_0^{\alpha}\) and \(\bar{J}_0^{\alpha}\). While \(Q^{\alpha} = J_0^{\alpha} + \bar{J}_0^{\alpha}\) are exact symmetries, the charges \(M^{\alpha} = J_0^{\alpha}-\bar{J}_0^{\alpha}\) do not commute with the Hamiltonian on the lattice, i.e. \([M^{\alpha}, H] \neq 0\). However linear extrapolation of matrix elements \(\mel{\psi}{M^\alpha}{\psi}\) suggests that in the thermodynamic limit each \(M^\alpha\) becomes indeed a conserved charge, as illustrated in Fig. [\[fig:xxx_m\]](#fig:xxx_m){reference-type="ref" reference="fig:xxx_m"}. We also numerically confirmed that the matrix elements \(\mel{\psi}{[Q^\alpha, M^\beta]}{\psi}\) approximately vanish for low energy eigenstates, up to finite size error (not shown in Fig. [\[fig:xxx_m\]](#fig:xxx_m){reference-type="ref" reference="fig:xxx_m"}). Kac--Moody primary states can be identified numerically by checking conditions in Eq. [\[eq:hws\]](#eq:hws){reference-type="eqref" reference="eq:hws"}. Up to finite size error, the only eigenstates that satisfy the conditions in Eq. [\[eq:hws\]](#eq:hws){reference-type="eqref" reference="eq:hws"} are \(\ket{\mathbf{1}}=\ket{V_{0,0}}\) and \(\ket{V_{0, 1}}\). For other Virasoro primary states that are not Kac--Moody primary states, we confirm that they can be obtained as descendants of these two Kac--Moody primary states using the lattice generators. For example, among the four Virasoro primary states \(\ket{V_{0, \pm1}}\) and \(\ket{V_{\pm 1, 0}}\) at conformal spin \(S=0\) and dimension \(\Delta=1/2\), only \(\ket{V_{1, 0}}\) is a Kac--Moody primary state. We show that \(\ket{V_{0, \pm1}}\) are indeed Kac--Moody descendent states of \(\ket{V_{1, 0}}\) by checking matrix elements \(\mel{V_{0, \pm1}}{J^{\pm}_{0}}{V_{1, 0}}\), as plotted in Fig. [\[fig:xxx_matele\]](#fig:xxx_matele){reference-type="ref" reference="fig:xxx_matele"}. Acting on the Kac--Moody primary states with \(J_{-n}^\alpha\) and \(\bar{J}_{-n}^\alpha\), all other descendent states can be reached and classified into Kac--Moody towers. While \(J^3_n\) preserves the eigenvalue of \(J^z\), \(J^{\pm}_{-n}\) changes the eigenvalue of \(J^z\) by \(\pm n\). We illustrate some examples of the actions of \(J^{\alpha}_{-n}\) and \(\bar{J}^{\alpha}_{-n}\) in Fig. [\[fig:spectrum_xxx\]](#fig:spectrum_xxx){reference-type="ref" reference="fig:spectrum_xxx"} and present the finite size scaling results of some corresponding matrix elements in Fig. [\[fig:xxx_matele\]](#fig:xxx_matele){reference-type="ref" reference="fig:xxx_matele"}. Moreover, we are able to confirm not only the level constant \(k\) but also the structure constants \(f^{\alpha \beta \gamma}\). While the structure constants of the diagonal subalgebra are given, it is non-trivial to check the commutation relations of the chiral algebra Eq. [\[eq:struct\]](#eq:struct){reference-type="eqref" reference="eq:struct"}. For example, we confirm \([J^x_{m}, J^y_{n}]=i \sqrt{2}J^z_{m+n}\) in the low energy subspace by calculating the ratio between \(\mel{\psi}{[J^x_{m}, J^y_{n}]}{\psi}\) and \(\mel{\psi}{i \sqrt{2} J^z_{m+n}}{\psi}\). In Fig. [\[fig:xxx_kc\]](#fig:xxx_kc){reference-type="ref" reference="fig:xxx_kc"}, we show that in the thermodynamic limit \(\mel{\mathbf{1}}{J^y_{-1} J^x_{-1}}{\mathbf{1}} \approx 0.9787 \times \left(i \sqrt{2} \mel{\mathbf{1}}{J^z_{-2}}{\mathbf{1}} \right)\), which is in good agreement with Eq. [\[eq:struct\]](#eq:struct){reference-type="eqref" reference="eq:struct"}, up to finite-size effects. Other components of \(f^{\alpha \beta \gamma}\) can be checked numerically in similar fashion using the proposed lattice Kac--Moody generators.
# Discussion {#sec:discussion}
Since the groundbreaking work by Cardy and others in the 1980s, there have been a series of studies on numerically obtaining conformal data from critical lattice systems (see e.g. ). Our paper adds new capabilities and applications to this line of work, by addressing the very important case of quantum critical spin chains with a global symmetry. As we have reviewed, a critical quantum spin chains with a microscopic Lie group symmetry \(G\) corresponds to a CFT with Kac--Moody symmetry. In this paper, we have proposed a concrete construction of Kac--Moody currents \(J_n^{\alpha}\) and \(\bar{J}_n^{\alpha}\) on lattice with the Hamiltonian and its microscopic symmetry as the only input. Our construction allows us to observe the emergence of Kac--Moody symmetries numerically. We illustrated our methods using two spin models, each having a global symmetry with Lie group \(G\), namely the XXZ model as an Abelian example and the XXX (or Heisenberg) model with next-to-nearest neighbor coupling as a non-Abelian example. First, we obtained the low-energy spectrum of the Hamiltonian using exact diagonalization. We then verified our proposal by studying the action of lattice Kac--Moody generators on low-energy energy-momentum eigenstates. As we consider larger system sizes, a second copy of the symmetry is observed to emerge, making the global symmetry \(G\times G\). We also successfully demonstrated the identification of the Kac--Moody primary states and their Kac--Moody towers. We show that the proposal works for both the Abelian and non-Abelian cases, and with independence of whether the model is integrable. To apply our method to larger system sizes, and thus further reduce finite-size corrections in the extracted conformal data, tensor network techniques such as puMPS can be employed. It would also be interesting to study the realizations of other extended symmetries on the lattice, such as the \(\mathcal{W}\)-algebra appearing in the three-state Potts model. *Note:* In subsequent work (which was posted recently ), one of the authors generalized some of the aspects of our work to a more general setting that includes critical systems for which an explicit local Hamiltonian description is not available.
# [\[sec:level1\]]{#sec:level1 label="sec:level1"}Introduction
Near a continuous phase transition, microscopically different systems may display low-energy, long-distance behaviour that is remarkably similar. Such universality is explained by saying that those systems flow to the same fixed point of the renormalization group (RG). In 1+1 dimensions, the universality class of a (massless) RG fixed point is often described by a conformal field theory (CFT) which, in turn, is completely characterized by a set of parameters known as the conformal data. In this work we are concerned with critical quantum spin chains, as specified in terms of a local lattice Hamiltonian in one spatial dimension. Given one such Hamiltonian, a natural goal is to numerically compute the conformal data of the emergent 1+1 CFT that describes the behaviour of the spin chain at long distances and low energies. In other words, to numerically characterize the emergent universal behaviour of the phase transition realized by the quantum spin chain. For this purpose, one may follow an approach, initiated in the 1980's, that is based on the *operator-state correspondence*. Given a 1+1 CFT on the circle, this correspondence establishes that each simultaneous eigenstate \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) of the Hamiltonian \(H^{\textsl{\tiny CFT}}\) and momentum \(P^{\textsl{\tiny CFT}}\) of the CFT on the circle corresponds to a scaling operator \(\psi^{\textsl{\tiny CFT}}_\alpha\) (an operator that transforms covariantly under scale transformations). Moreover, the energy and momentum of the state \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) relate to the universal scaling properties of the operator \(\psi^{\textsl{\tiny CFT}}_\alpha\). Importantly, the low energy states of a critical quantum spin chain on the circle are organized as in the emergent CFT. Accordingly, Cardy and others proposed that one could extract the conformal data by simply studying the low energy states of the critical quantum spin chain. Over the years, several other authors have contributed additional insights into this strategy. A crucial step was to identify operators on the lattice with operators of the CFT. In particular, certain lattice operators can be identified with Virasoro generators of the conformal symmetry in the CFT. This means that such lattice operators act on the low-energy states of the critical quantum spin chain approximately (up to finite size corrections) in the same way as the Virasoro generators act on the corresponding states in the CFT. The lattice Virasoro generators allow us to see the emergence of conformal symmetry directly on the lattice, offering a way to systematically identify primary operators and conformal towers in the low-energy spectrum of a quantum critical spin chain. In some CFTs, conformal symmetry is enhanced to a larger symmetry, which further relates different Virasoro primary operators and conformal towers. Extended symmetries exist in many CFTs that describe lattice models, such as quantum spin chains, classical statistical-mechanics models, and edge modes of topological orders. One important implication of the extended symmetry is that it reduces the number of independent primary operators in the conformal data. In the cases known as rational CFTs, the number of primary operators with respect to the extended symmetry is finite, while the number of Virasoro primary fields can be infinite. This makes the extraction of complete conformal data possible. A prominent example is when conformal symmetry is enhanced by a global symmetry with Lie group \(G\). In this case, scaling operators are organized by an extension of the Virasoro algebra, the Kac--Moody algebra, denoted by \(\mathfrak{g}_k\), where \(\mathfrak{g}\) is the Lie algebra of the the group \(G\) and \(k\in \mathbb{Z}\) is the *level constant*. A remarkable consequence of the extension is that the global symmetry of the CFT becomes \(G\times G\), which acts independently on left and right moving fields. In this paper we investigate how the Kac--Moody symmetry emerges from critical quantum spin chains with a Lie group symmetry \(G\). We construct lattice operators that correspond to the generators of the Kac--Moody algebra. We numerically confirm that the global symmetry is enhanced to \(G\times G\) in the low-energy subspace and that the eigenstates can be organized into Kac--Moody towers. We test our proposal with the XXZ model for the \(G=U(1)\) case and the Heisenberg model with next-to-nearest-neighbor interactions for the \(G=SU(2)\) case. In both cases we find that the proposed lattice Kac--Moody generators act on the low-energy states as their CFT counterparts in the thermodynamic limit. In particular, we can extract the level constant \(k\) by computing the commutators of lattice Kac--Moody generators in the low-energy subspace. In previous work by some of the authors, lattice operators that correspond to generators of another form of extended symmetry, namely superconformal symmetry, were found numerically. Here, in contrast, we provide an analytical ansatz for the lattice Kac--Moody generators. For the specific case of the XXZ model, which is an integrable model, the lattice Kac--Moody generators can also be found via bosonization (as reviewed in Appendix [\[appendix:a\]](#appendix:a){reference-type="ref" reference="appendix:a"}), and we can use this previous result to further validate our proposal. In contrast, no previous results appear to be known for the Heisenberg model with next-to-near-neighbor interactions. The performance of our method, which treats all models on the same footing, is seen to not rely on integrablity. The rest of the paper is organized as follows. In Sec. [8](#sec:review){reference-type="ref" reference="sec:review"}, we review the conformal symmetry and the Virasoro algebra in a CFT, as well as the approximate lattice version of Virasoro generators. In Sec. [9](#sec:u1){reference-type="ref" reference="sec:u1"}, we consider the Kac--Moody algebra for the Abelian group \(G=U(1)\). We first review the free boson CFT and the action of the Kac--Moody algebra, and then propose our method to build an approximate lattice version of the Kac--Moody generators. We also present numerical tests of the lattice Kac--Moody generators in the XXZ quantum spin model. In Sec. [10](#sec:su2){reference-type="ref" reference="sec:su2"} we generalize the formalism to general non-Abelian group \(G\). We will focus on \(G=SU(2)\) and construct an approximate lattice version of the Kac--Moody generators for the Heisenberg model with next-to-nearest-neighbor interactions. In Sec. [11](#sec:discussion){reference-type="ref" reference="sec:discussion"} we conclude with discussions and future directions.
# Conformal symmetry and lattice Virasoro generators {#sec:review}
In this section we review conformal symmetry, the Virasoro algebra and its approximate realization in the low-energy states of a quantum critical spin chain, see Ref. and Ref. for an introduction. Throughout this paper, we use superscript \(^\textsl{\tiny CFT}\) to denote objects in a CFT, e.g. \(T^\textsl{\tiny CFT}\), \(L_n^\textsl{\tiny CFT}\), \(\J{n}{\alpha}\), and in this way distinguish them from the corresponding objects in a lattice, e.g. \(T\), \(L_n\), \(J_n^{\alpha}\).
## Virasoro algebra
Consider a conformal field theory in 1+1 dimensions. Conformal transformations on the plane are generated by Virasoro generators \(L^{\textsl{\tiny CFT}}_n,\bar{L}^{\textsl{\tiny CFT}}_n~(n\in\mathbb{Z})\) that satisfy the Virasoro algebra, \[\begin{aligned}
\label{VirasoroCFT} &[L^{\textsl{\tiny CFT}}_n,L^{\textsl{\tiny CFT}}_m]=(n-m) L^{\textsl{\tiny CFT}}_{n+m} + \frac{c^{\textsl{\tiny CFT}}}{12}n (n^2-1) \delta_{n+m,0} \\ &[\bar{L}^{\textsl{\tiny CFT}}_n,\bar{L}^{\textsl{\tiny CFT}}_m]=(n-m) \bar{L}^{\textsl{\tiny CFT}}_{n+m} + \frac{c^{\textsl{\tiny CFT}}}{12}n (n^2-1) \delta_{n+m,0} \\ &[L^{\textsl{\tiny CFT}}_n,\bar{L}^{\textsl{\tiny CFT}}_m]=0. \end{aligned}\] In particular, dilations and rotations of the plane are generated by \(D^{\textsl{\tiny CFT}}=L^{\textsl{\tiny CFT}}_0+\bar{L}^{\textsl{\tiny CFT}}_0\) and \(R^{\textsl{\tiny CFT}}=L^{\textsl{\tiny CFT}}_0-\bar{L}^{\textsl{\tiny CFT}}_0\), respectively. The Hilbert space of the CFT is supported on circles around the origin. Any state in the Hilbert space can be spanned by simultaneous eigenstates \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) of \(L^{\textsl{\tiny CFT}}_0\) and \(\bar{L}^{\textsl{\tiny CFT}}_0\), \[L^{\textsl{\tiny CFT}}_0|\psi^{\textsl{\tiny CFT}}_\alpha\rangle=h^{\textsl{\tiny CFT}}_\alpha|\psi^{\textsl{\tiny CFT}}_\alpha\rangle, ~~ \bar{L}^{\textsl{\tiny CFT}}_0|\psi^{\textsl{\tiny CFT}}_\alpha\rangle=\bar{h}^{\textsl{\tiny CFT}}_\alpha|\psi^{\textsl{\tiny CFT}}_\alpha\rangle,\] where \((h^{\textsl{\tiny CFT}}_\alpha,\bar{h}^{\textsl{\tiny CFT}}_\alpha)\) are the holomorphic and anti-holomorphic conformal dimensions of the state. The eigenvalues of \(D^{\textsl{\tiny CFT}}\) and \(R^{\textsl{\tiny CFT}}\) are \(\Delta^{\textsl{\tiny CFT}}_\alpha=h^{\textsl{\tiny CFT}}_\alpha+\bar{h}^{\textsl{\tiny CFT}}_\alpha\) and \(s^{\textsl{\tiny CFT}}_\alpha=h^{\textsl{\tiny CFT}}_\alpha-\bar{h}^{\textsl{\tiny CFT}}_\alpha\), known as scaling dimensions and conformal spins, respectively. Setting \(n=0\) in Eq. [\[VirasoroCFT\]](#VirasoroCFT){reference-type="eqref" reference="VirasoroCFT"}, we see that acting with \(L^{\textsl{\tiny CFT}}_m\) on a state changes the holomorphic dimension by \(-m\). Therefore, \(L^{\textsl{\tiny CFT}}_m\) lowers the scaling dimensions and the conformal spins by \(m\). Similarly, \(\bar{L}^{\textsl{\tiny CFT}}_m\) lowers the scaling dimensions by \(m\) and increases the conformal spins by \(m\). A Virasoro primary state \(|\phi^{\textsl{\tiny CFT}}_\alpha\rangle\) is defined such that its scaling dimension cannot be lowered, \[\label{eq:primarydef} L^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \bar{L}^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \forall n>0.\] Each primary state is associated with a Virasoro conformal tower, which contains that primary state as well as all its Virasoro descendant states. The descendant states are obtained by acting successively with raising operators \(L^{\textsl{\tiny CFT}}_m,\bar{L}^{\textsl{\tiny CFT}}_m~(m<0)\) on the primary state.
## CFT on the cylinder
In 1+1 dimensions one can use a conformal transformation to map the CFT from the plane to a cylinder \(S^{1}\times\mathbb R\), where the axial direction is the imaginary time direction with coordinate \(\tau\in(-\infty,\infty)\), and the angular direction is the spatial direction with coordinate \(x\in [0,L)\). The Hilbert space is supported on the equal-time slices. We will focus on the \(\tau=0\) slice, and denote the fields with its spatial coordinate \(x\). In any CFT, there exist the holomorphic and anti-holomorphic energy-momentum tensors \(T^\textsl{\tiny CFT}\) and \(\bar{T}^\textsl{\tiny CFT}\), with conformal dimensions \((2,0)\) and \((0,2)\), respectively. The Hamiltonian and momentum can be expressed as an integral of the stress tensor, \[\begin{aligned}
H^{\textsl{\tiny CFT}} &=& \int_0^{L} dx\, h^\textsl{\tiny CFT}(x) \\ P^{\textsl{\tiny CFT}} &=& \int_0^{L} dx\, p^\textsl{\tiny CFT}(x), \end{aligned}\] where \[\begin{aligned}
h^\textsl{\tiny CFT}(x) &=& \frac{1}{2\pi}(T^\textsl{\tiny CFT}(x)+\bar{T}^\textsl{\tiny CFT}(x)) \\ p^\textsl{\tiny CFT}(x) &=& \frac{1}{2\pi}(T^\textsl{\tiny CFT}(x)-\bar{T}^\textsl{\tiny CFT}(x)). \end{aligned}\] The Virasoro generators are the Fourier modes of \(T^\textsl{\tiny CFT}\) and \(\bar{T}^\textsl{\tiny CFT}\), \[\begin{aligned} \label{eq:LnCFT} L^{\textsl{\tiny CFT}}_n = \frac{L}{(2\pi)^2} \int_{0}^{L} dx e^{+ i n x \frac{2\pi}{L}} T^\textsl{\tiny CFT} (x) + \frac{c^{\textsl{\tiny CFT}}}{24} \delta_{n,0}, \\ \bar{L}^{\textsl{\tiny CFT}}_n = \frac{L}{(2\pi)^2} \int_{0}^{L} dx e^{-i n x \frac{2\pi}{L}} \bar{T}^\textsl{\tiny CFT} (x) + \frac{c^{\textsl{\tiny CFT}}}{24} \delta_{n,0}. \end{aligned}\] We may express the Hamiltonian and momentum as \[\begin{aligned}
\label{HCFT} H^{\textsl{\tiny CFT}}&=&\frac{2\pi}{L}\left(L^{\textsl{\tiny CFT}}_0+\bar{L}^{\textsl{\tiny CFT}}_0-\frac{c^{\textsl{\tiny CFT}}}{12}\right) \\ \label{PCFT} P^{\textsl{\tiny CFT}}&=& \frac{2\pi}{L}\left(L^{\textsl{\tiny CFT}}_0-\bar{L}^{\textsl{\tiny CFT}}_0\right). \end{aligned}\] We see that each simultaneous eigenstate \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\) of \(H^{\textsl{\tiny CFT}}\) and \(P^{\textsl{\tiny CFT}}\) has energy and momentum \[\begin{aligned}
\label{eq:ECFT} E^{\textsl{\tiny CFT}}_\alpha&=&\frac{2\pi}{L}\left(\Delta^{\textsl{\tiny CFT}}_\alpha-\frac{c^{\textsl{\tiny CFT}}}{12}\right) \\ \label{eq:PCFT} P^{\textsl{\tiny CFT}}_\alpha&=& \frac{2\pi}{L} s^{\textsl{\tiny CFT}}_\alpha. \end{aligned}\] It is also useful to express the Fourier mode of the Hamiltonian density as a linear combination of Virasoro generators \(L_n^{\textsl{\tiny CFT}}\) and \(\bar{L}_{-n}^{\textsl{\tiny CFT}}\), \[L^{\textsl{\tiny CFT}}_n+\bar{L}^{\textsl{\tiny CFT}}_{-n}= H^{\textsl{\tiny CFT}}_n \equiv \frac{L}{2\pi} \int_{0}^{L} dx e^{+ i n x \frac{2\pi}{L}} h^\textsl{\tiny CFT} (x).\] These operators transform states within the same conformal tower of the CFT.
## Lattice Virasoro generators
Given a critical quantum spin chain on a circle with Hamiltonian \[H=\sum_{j} h_j,\] its low-energy eigenstates \(|\psi_\alpha\rangle\) are in one-to-one correspondence with CFT states \(|\psi^{\textsl{\tiny CFT}}_\alpha\rangle\). The energies and momenta of \(|\psi_\alpha\rangle\) are related to the scaling dimensions and conformal spins by \[\begin{aligned}
\label{eq:Elat} E_\alpha &=& A+B\frac{2\pi}{N}\left(\Delta^{\textsl{\tiny CFT}}_\alpha-\frac{c^{\textsl{\tiny CFT}}}{12}\right)+O(N^{-x}) \\ \label{eq:Plat} P_\alpha &=& \frac{2\pi}{N}s^{\textsl{\tiny CFT}}_\alpha, \end{aligned}\] where the \(O(N^{-x})\) with \(x>1\) is the non-universal finite-size correction. These relations are direct lattice versions of expressions Eqs. [\[eq:ECFT\]](#eq:ECFT){reference-type="eqref" reference="eq:ECFT"}-[\[eq:PCFT\]](#eq:PCFT){reference-type="eqref" reference="eq:PCFT"} for the CFT, in that upon identifying the circle size \(L\) with the size \(N\) of the spin chain, we could rewrite them as \(E_{\alpha} = A + BE_{\alpha} + O(N^{-x})\) and \(P_{\alpha} = P_{\alpha}^{\textsl{\tiny CFT}}\). From Eqs. [\[eq:Elat\]](#eq:Elat){reference-type="eqref" reference="eq:Elat"}-[\[eq:Plat\]](#eq:Plat){reference-type="eqref" reference="eq:Plat"}, one may extract approximate scaling dimensions \(\Delta_\alpha\) and exact conformal spins \(s_\alpha\) from the low-energy spectrum of the spin chain. In particular, the constants \(A,B\) (which depend on how the lattice Hamiltonian is normalized) may be determined by demanding that the scaling dimensions of the identity operator and the stress tensor be \(\Delta_{\mathbf{1}}=\Delta^{\textsl{\tiny CFT}}_{\mathbf{1}}=0\) and \(\Delta_{T}=\Delta^{\textsl{\tiny CFT}}_T=2\). Notice that the states \(\ket{\mathbf{1}^{\textsl{\tiny CFT}}}\) and \(\ket{T^{\textsl{\tiny CFT}}}\) are present in any CFT, so the above procedure to determine constants \(A\) and \(B\) can always be applied. The lattice operator \(H_n\) that corresponds to \(H^{\textsl{\tiny CFT}}_n\) is the Fourier mode of the lattice Hamiltonian density \(h_j\), \[H_n=\frac{N}{B}\sum_{j} e^{inj \frac{2\pi}{N}} h_j \sim H^{\textsl{\tiny CFT}}_n.\] There is both analytical and numerical evidence that \(H_n\) acts on the low-energy eigenstates of the lattice Hamiltonian as \(H^{\textsl{\tiny CFT}}_n\) acts on the corresponding CFT states. We stress that the lattice Virasoro generators \(H_n\) only (approximately) satisfy the algebra obeyed by the CFT operators \(H^{\textsl{\tiny CFT}}_n\) when projected onto the low-energy eigenstates of the quantum spin chains. At the level of lattice operators, it is easily checked (e.g. numerically) that the \(H_n\) operators do not satisfy the Virasoro algebra. Similar to the \(O(N^{-x})\) term in Eq. [\[eq:Elat\]](#eq:Elat){reference-type="eqref" reference="eq:Elat"}, there are also non-universal finite-size corrections in the matrix elements of \(H_n\), which can be reduced by an extrapolation to the thermodynamic limit (that is, the limit of an inifnitely large spin chain, \(N \rightarrow \infty\)). Since the lattice Virasoro generators \(H_n\) connect low-energy eigenstates within the same conformal tower, they can be used to identify Virasoro primary states and their conformal towers on the lattice.
# Lattice realization of Kac--Moody algebra: the \(U(1)\) case {#sec:u1}
In this section we first review the \(U(1)\) Kac--Moody algebra and its manifestation in the free compactified boson CFT. We then consider a critical quantum spin chain with \(U(1)\) symmetry and construct a lattice version of the Kac--Moody generators. Finally we numerically verify the actions of the proposed lattice Kac--Moody generators in the XXZ model.
## \(U(1)\) Kac--Moody algebra
We consider a CFT with a global \(U(1)\) symmetry, with conserved \(U(1)\) charge \(Q^{\textsl{\tiny CFT}}\), which is an integral over space of a conserved local current \(q^{\textsl{\tiny CFT}}(x)\), \[\label{eq:U1QCFT} Q^{\textsl{\tiny CFT}}=\int dx\, q^{\textsl{\tiny CFT}}(x).\] In a CFT, it turns out that the conserved current can be further divided into holomorphic and anti-holomorphic parts which are separately conserved, \[\label{eq:U1qCFT} q^{\textsl{\tiny CFT}}(x)=J^{\textsl{\tiny CFT}}(x)+\bar{J}^{\textsl{\tiny CFT}}(x),\] where \(J^{\textsl{\tiny CFT}}\) and \(\bar{J}^{\textsl{\tiny CFT}}\) are *current operators* with conformal dimensions \((1,0)\) and \((0,1)\), respectively. Since both current operators satisfy conservation laws, namely \(\bar{\partial} J^{\textsl{\tiny CFT}}=0\) and \(\partial \bar{J}^{\textsl{\tiny CFT}}=0\), there is an additional conserved charge \[\label{eq:U1MCFT} M^{\textsl{\tiny CFT}}=\int dx\, m^{\textsl{\tiny CFT}}(x),\] where \[\label{eq:U1mCFT} m^{\textsl{\tiny CFT}}(x)=J^{\textsl{\tiny CFT}}(x)-\bar{J}^{\textsl{\tiny CFT}}(x).\] The global symmetry is therefore \(U(1)\times U(1)\). For later reference, let us combine Eqs. [\[eq:U1qCFT\]](#eq:U1qCFT){reference-type="eqref" reference="eq:U1qCFT"}-[\[eq:U1mCFT\]](#eq:U1mCFT){reference-type="eqref" reference="eq:U1mCFT"} and the conservation relations above, to obtain \[-i\partial_\tau q^{\textsl{\tiny CFT}}(x)=\partial_x m^{\textsl{\tiny CFT}}(x).\] Then, upon substituting the time derivative with the commutator by the Hamiltonian we further obtain \[\label{eq:U1qmCFT} i [H^{\textsl{\tiny CFT}},q^{\textsl{\tiny CFT}}(x)]=\partial_x m^{\textsl{\tiny CFT}}(x).\] Similar to Eq. [\[eq:LnCFT\]](#eq:LnCFT){reference-type="eqref" reference="eq:LnCFT"}, which defined the Virasoro generators, we may define the generators of the Kac--Moody algebra by \[\begin{aligned}
\label{eq:U1JnCFT} J^{\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{+inx\frac{2\pi}{L}} J^{\textsl{\tiny CFT}}(x) \\ \bar{J}^{\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{-inx\frac{2\pi}{L}} \bar{J}^{\textsl{\tiny CFT}}(x).\label{eq:U1JnCFTb} \end{aligned}\] They satisfy the \(\mathfrak{u}(1)_k\) Kac--Moody algebra for some value \(k\in \mathbb{Z}\), known as the level constant, \[\label{eq:JmJn} \begin{aligned} &[J^{\textsl{\tiny CFT}}_m, J^{\textsl{\tiny CFT}}_n] = km \delta_{m+n,0},\\ &[\bar{J}^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = km \delta_{m+n,0},\\ &[J^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = 0. \end{aligned}\] The Virasoro generators and Kac--Moody generators satisfy the commutation relations \[\label{eq:LmJn} \begin{aligned} [L^{\textsl{\tiny CFT}}_m, J^{\textsl{\tiny CFT}}_n] =-n J^{\textsl{\tiny CFT}}_{m+n}, \\ [\bar{L}^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] =-n \bar{J}^{\textsl{\tiny CFT}}_{m+n}, \end{aligned}\] which is compatible with the fact that \(J^{\textsl{\tiny CFT}}\) is a Virasoro primary operator with conformal dimensions \((1,0)\). Setting \(m=0\) in Eq. [\[eq:LmJn\]](#eq:LmJn){reference-type="eqref" reference="eq:LmJn"}, we see that \(J^{\textsl{\tiny CFT}}_m\) changes the holomorphic dimension by \(-m\), as what \(L^{\textsl{\tiny CFT}}_m\) does. Therefore, \(J^{\textsl{\tiny CFT}}_{m}\) is a raising operator with negative \(m\), and a lowering operator with positive \(m\). When \(m=0\), \(J^{\textsl{\tiny CFT}}_{0}\) commutes with \(L^{\textsl{\tiny CFT}}_0\) and they have the same eigenstates. This reflects the fact that \(J^{\textsl{\tiny CFT}}(x)\) is a conserved current. In analogy with the definition of Virasoro primary states in Eq. [\[eq:primarydef\]](#eq:primarydef){reference-type="eqref" reference="eq:primarydef"}, we define Kac--Moody primary states, \[J^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \bar{J}^{\textsl{\tiny CFT}}_n|\phi^{\textsl{\tiny CFT}}_\alpha\rangle=0, ~~ \forall n>0.\] A Kac--Moody tower consists of a Kac--Moody primary state \(|\phi^{\textsl{\tiny CFT}}_\alpha\rangle\) with scaling dimension \(\Delta_\alpha\) and conformal spin \(s_\alpha\) and their descendent states, obtained from the primary state by sequentially acting with the generators. For instance, examples of Kac--Moody descendants include \[|\phi^{\textsl{\tiny CFT}}_\beta\rangle = J^{\textsl{\tiny CFT}}_{-n} |\phi^{\textsl{\tiny CFT}}_\alpha\rangle, ~~ |\phi^{\textsl{\tiny CFT}}_\beta\rangle = \bar{J}^{\textsl{\tiny CFT}}_{-n} |\phi^{\textsl{\tiny CFT}}_\alpha\rangle, ~~ \text{for} \ n > 0.\] with \[\label{eq:ladder} \Delta_{\beta} = \Delta_{\alpha} + n, \quad s_{\beta} = s_{\alpha} \pm n, \quad \text{for} \ n > 0.\] It can be shown that all Kac--Moody primary states are also Virasoro primary states. However, there are Virasoro primary states that are not Kac--Moody primary states. For example, \[\begin{aligned}
|J^{\textsl{\tiny CFT}}\rangle &=& J^{\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle \\ |\bar{J}^{\textsl{\tiny CFT}}\rangle &=& \bar{J}^{\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle \\ |J\bar{J}^{\textsl{\tiny CFT}}\rangle &=& J^{\textsl{\tiny CFT}}_{-1}\bar{J}^{\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle, \end{aligned}\] that is, the states corresponding to the holomorphic current \(J^{\textsl{\tiny CFT}}\), the anti-holomophic current \(\bar{J}^{\textsl{\tiny CFT}}\) and their composite operator \(J\bar{J}^{\textsl{\tiny CFT}}\), which can be seen to be Virasoro primary states, all belong to the same Kac--Moody tower.
## Free compactified boson CFT
A simple example of CFT that has Kac--Moody symmetry is a free compactified boson CFT, with action \[S^{\textsl{\tiny CFT}}=\frac{1}{2}\int d\tau dx \, \left[ (\partial_\tau \phi)^2+(\partial_x \phi)^2 \right],\] where \(\phi\) is identified with \(\phi+2\pi R\), and \(R\) is the *compactification radius*. The holomorphic and anti-holomorphic conserved currents read \[J^{\textsl{\tiny CFT}}=i\partial\phi,~~ \bar{J}^{\textsl{\tiny CFT}}=i\bar{\partial}\phi,\] where \(\partial=(\partial_\tau+ i\partial_x)/2\) and \(\bar{\partial}=(\partial_\tau-i\partial_x)/2\) are holomorphic and anti-holomorphic derivatives. In this CFT, the conservation relations \(\bar{\partial}J^{\textsl{\tiny CFT}}=0\) and \(\partial \bar{J}^{\textsl{\tiny CFT}}=0\) follow from the equation of motion \(\partial\bar{\partial}\phi=0\). Now, using the definition of Kac--Moody generators in Eqs. [\[eq:U1JnCFT\]](#eq:U1JnCFT){reference-type="eqref" reference="eq:U1JnCFT"}-[\[eq:U1JnCFTb\]](#eq:U1JnCFTb){reference-type="eqref" reference="eq:U1JnCFTb"} and the canonical commutation \([\phi(x),\partial_\tau\phi(y)]=\delta(x-y)\), we can verify the Kac--Moody algebra, \[\label{eq:u1comm} \begin{aligned} &[J^{\textsl{\tiny CFT}}_m, J^{\textsl{\tiny CFT}}_n] = [\bar{J}^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = m \delta_{m+n,0},\\ &[J^{\textsl{\tiny CFT}}_m, \bar{J}^{\textsl{\tiny CFT}}_n] = 0, \end{aligned}\] which corresponds level constant \(k=1\). The free compactified boson CFT has infinitely many primary states with respect to the \(\mathfrak{u}(1)_1\) Kac--Moody algebra. They are vertex operators \(V^{\textsl{\tiny CFT}}_{Q,M}\) labelled by integers \(Q\) and \(M\), with scaling dimensions and conformal spins \[\Delta^{\textsl{\tiny CFT}}_{V_{Q,M}} = \frac{Q^2}{R^2} + \frac{R^2M^2}{4}, \quad S^{\textsl{\tiny CFT}}_{V_{Q,M}} = QM, \quad Q, M \in \mathbb{Z},\] where \(Q\) and \(M\) are the eigenvalues of \(Q^{\textsl{\tiny CFT}}\) and \(M^{\textsl{\tiny CFT}}\), respectively.
## Lattice Kac--Moody generators
Consider now a critical quantum spin chain on the circle with Hamiltonian \(H\) and global \(U(1)\) symmetry (see e.g. the XXZ model below). Our goal is to construct lattice operators \(J_n,\bar{J}_m\) that correspond to the Kac--Moody generators \(J^{\textsl{\tiny CFT}}_n,\bar{J}^{\textsl{\tiny CFT}}_m\). On the lattice, let the conserved \(U(1)\) charge be \[Q=\sum_{j} q_j,\] which commutes with the Hamiltonian \([H, Q] =0\). We identify \(Q\) with the CFT charge operator \(Q^{\textsl{\tiny CFT}}\) in Eq. [\[eq:U1QCFT\]](#eq:U1QCFT){reference-type="eqref" reference="eq:U1QCFT"}, and \(q\) with the CFT current density \(q^{\textsl{\tiny CFT}}\) in Eq. [\[eq:U1qCFT\]](#eq:U1qCFT){reference-type="eqref" reference="eq:U1qCFT"}. In order to find the Kac--Moody generators, one also needs to find the lattice operator \(m\) corresponding to \(m^{\textsl{\tiny CFT}}\) in Eq. [\[eq:U1mCFT\]](#eq:U1mCFT){reference-type="eqref" reference="eq:U1mCFT"}. Our proposal is that the corresponding lattice operators should satisfy an equation analogous to Eq. [\[eq:U1qmCFT\]](#eq:U1qmCFT){reference-type="eqref" reference="eq:U1qmCFT"}, after replacing the spatial derivative \(\partial_x\) with a finite difference, see Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"} below. Then by Eqs. [\[eq:U1qCFT\]](#eq:U1qCFT){reference-type="eqref" reference="eq:U1qCFT"}-[\[eq:U1mCFT\]](#eq:U1mCFT){reference-type="eqref" reference="eq:U1mCFT"}, the lattice operators \(J_j\) and \(\bar{J}_j\) that correspond to \(J^{\textsl{\tiny CFT}}\) and \(\bar{J}^{\textsl{\tiny CFT}}\) can be identified as in Eq. [\[eq:U1latticeJi\]](#eq:U1latticeJi){reference-type="eqref" reference="eq:U1latticeJi"} below. Finally the lattice operators \(J_n\) and \(\bar{J}_n\) corresponding to the Kac--Moody generators are constructed as Fourier modes in Eq. [\[eq:U1J\]](#eq:U1J){reference-type="eqref" reference="eq:U1J"}.
Eqs.[\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"}, [\[eq:U1latticeJi\]](#eq:U1latticeJi){reference-type="eqref" reference="eq:U1latticeJi"} and [\[eq:U1J\]](#eq:U1J){reference-type="eqref" reference="eq:U1J"} are the main proposal of this paper (together with their generalization below to the non-Abelian case). Let us elaborate on the form of \(m_j\) when \(H\) is a nearest-neighbor and a next-to-nearest-neighbor Hamiltonian. These specializations will be useful in the applications to specific lattice models.
### Nearest-neighbor Hamiltonian
Consider a nearest-neighbor Hamiltonian \(H = \sum_{j} h_{j,j+1}\) with a global symmetry that is realized on-site with charge \(Q = \sum_{j} q_j\). If follows from \([H, Q]=0\) that at for any pair \((j, j+1)\) of contiguous sites, \[= 0.\] The LHS of Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"} then becomes \[\begin{aligned} i [H, q_j] &= i ([h_{j-1,j}, q_j] + [h_{j, j+1}, q_j]) \\ &= i (-[h_{j-1,j}, q_{j-1}] + [h_{j, j+1}, q_j]), \end{aligned}\] and matching it with the RHS of Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"}, we obtain \[\label{eq:2local} m_j = i[h_{j,j+1}, q_j],\] which is our proposal as a lattice version of the locally conserved current \(m^{\textsl{\tiny CFT}}(x)\).
### Next-to-nearest-neighbor Hamiltonian
As a second example, consider a next-to-nearest-neighbor Hamiltonian \(H = \sum_j h_{j-1, j, j+1}\), again with a global symmetry that is realized on-site with charge \(Q = \sum_{j} q_j\). To simplify the notation, in the following \(h_{(j)} \coloneqq h_{j-1,j,j+1}\). At any set \((j-1,j,j+1)\) of three consecutive sites, \[= 0,\] we can check that \[\begin{aligned} i\left[H, \frac{q_{j}+q_{j+1}}{2}\right] = \frac{1}{2}(m_{j+1}-m_{j-1}) \end{aligned}\] where \[\label{eq:3local} m_{j} = i([h_{(j+1)}, q_j]-[h_{(j)}, q_{j+1}]).\] This is a discrete version of Eq. [\[eq:U1qmCFT\]](#eq:U1qmCFT){reference-type="eqref" reference="eq:U1qmCFT"} in the case of next-to-nearest-neighbor Hamiltonian. Therefore we identify \(m_j\) in Eq. [\[eq:3local\]](#eq:3local){reference-type="eqref" reference="eq:3local"} with \(m^{\textsl{\tiny CFT}}(x)\). To reiterate, given a local Hamiltonian \(H\) with a microscopic global symmetry realized on-site, that is with charge \(Q=\sum_j q_j\), our proposal gives a concrete way of constructing the lattice current density \(m_j\) via Eq. [\[eq:U1qm\]](#eq:U1qm){reference-type="eqref" reference="eq:U1qm"} and subsequently lattice Kac--Moody generators \(J_n\) and \(\bar{J}_n\) via Eq. [\[eq:U1latticeJi\]](#eq:U1latticeJi){reference-type="eqref" reference="eq:U1latticeJi"}-[\[eq:U1J\]](#eq:U1J){reference-type="eqref" reference="eq:U1J"}.
## Example: XXZ model
As a test of our construction of lattice current generators \(J_n\) and \(\bar{J}_{n}\) for a critical quantum spin chain with a global \(U(1)\) symmetry, we consider the XXZ model \[\label {eq: XXZ} H =-\frac{2\gamma}{\pi \sin{\gamma}} \sum_{j=1}^{N}\left( S_{j}^{X}S_{j+1}^{X} + S_{j}^{Y}S_{j+1}^{Y}-\cos{\gamma}S_{j}^{Z}S_{j+1}^{Z} \right)\] with anisotropy \(\gamma \in [0, \pi )\) and \(S^\alpha = \sigma^\alpha/2\), where \(\sigma^{x}, \sigma^y\) and \(\sigma^z\) stand for the Pauli matrices. This model displays indeed a global \(U(1)\) symmetry realized on-site, generated by the charge operator \(Q = \sum_{j=1}^{N} S_{j}^{Z}\), which commutes with the Hamiltonian, \([Q, H] = 0\). In the continuum limit the XXZ model is equivalent via bosonization to the Gaussian model of a free massless boson compactified on circle with radius \(R = \sqrt{2\pi/(\pi-\gamma)}\) and the symmetry algebra in the corresponding CFT theory is the Kac--Moody algebra with \(\mathfrak{g} = \mathfrak{u}(1)\). As discussed in the last section, we use Eq. [\[eq:2local\]](#eq:2local){reference-type="eqref" reference="eq:2local"} to compute the lattice current \(m_j\), \[\label{eq:u1m} m_j = \frac{2 \gamma}{\pi \sin{\gamma}} \left( S_{j}^{X}S_{j+1}^{Y}-S_{j}^{Y}S_{j+1}^{X} \right).\] We note that the lattice currents \(q_j\) and \(m_j\) for the XXZ model can also be understood using bosonization, as discussed in Appendix [\[appendix:a\]](#appendix:a){reference-type="ref" reference="appendix:a"}. Following Eq. [\[eq:J\]](#eq:J){reference-type="eqref" reference="eq:J"} the lattice versions of Kac--Moody generators can be constructed as \[\label{eq:Jnxxz} \begin{aligned} J_n = \sum_{j=1}^{N} e^{+ijn \frac{2\pi}{N}} \frac{1}{2} \left[ S_{j}^{Z}+ \frac{2\gamma}{\pi \sin{\gamma}} \left( S_{j}^{X}S_{j+1}^{Y}-S_{j}^{Y}S_{j+1}^{X} \right) \right], \\ \bar{J}_n = \sum_{j=1}^{N} e^{-ijn \frac{2\pi}{N}} \frac{1}{2} \left[ {S_{j}^{Z}}-\frac{2\gamma}{\pi \sin{\gamma}} \left( S_{j}^{X}S_{j+1}^{Y}-S_{j}^{Y}S_{j+1}^{X} \right) \right]. \end{aligned}\] To numerically check our proposal, we use exact diagonalization to find a set of low energy eigenstates \(\ket{\psi_\alpha}\). We simultaneously diagonalize the Hamiltonian and the translation operator with periodic boundary conditions to get the scaling dimensions \(\Delta_{\alpha}\) and conformal spins \(s_{\alpha}\). We also sort the low-energy eigenstates into different sectors according to the eigenvalues of \(Q=\sum_j S^Z_j\). As an example, in Fig. [\[fig:spectrum\]](#fig:spectrum){reference-type="ref" reference="fig:spectrum"} the low-energy spectra of XXZ spin chain of 20 sites with \(\gamma = \pi/2\) in the \(Q = 0\) sector and \(Q=1\) sector are plotted. That the above lattice generators \(J_n\) and \(\bar{J}_n\) approximately satisfy the Kac--Moody algebra at low energies can be verified by directly evaluating Eqs. [\[eq:JmJn\]](#eq:JmJn){reference-type="eqref" reference="eq:JmJn"}-[\[eq:LmJn\]](#eq:LmJn){reference-type="eqref" reference="eq:LmJn"} on low energy states. Those constitute an infinite set of conditions. In the following, we will demonstrate it for a subset of these conditions that are physically important. Other conditions can be confirmed similarly. From Eq. [\[eq:LmJn\]](#eq:LmJn){reference-type="eqref" reference="eq:LmJn"} for \(m=n=0\), we see that the charge \(M = J_0-\bar{J}_0\) should approximately satisfy \([M,H]=0\) and \([M,Q]=0\) when acting on low-energy states. While for general anisotropy \(\gamma\) the operator \(M=\sum_j m_j\) does not necessarily commute with the Hamiltonian, in the specific case of the XX model (\(\gamma = \pi/2\)), \(M \sim \sum_j S_j^X S_{j+1}^Y-S_j^Y S_{j+1}^X\) does commute with \(H\). In fact, there exists a series of conserved charges in the XX model. However \(M\) does not exactly commute with \(Q\) on the lattice. To confirm that the microscopic global \(U(1)\) symmetry turns indeed into an emergent \(U(1)\times U(1)\) symmetry at low energies, we check both the commutation relation \(\mel{\psi}{[Q,M]}{\psi}\) and the expectation value \(\mel{\psi}{M}{\psi}\) in the eigenbasis of \(Q\). For low energy eigenstates, we find that \(Q\) and \(M\) commute, up to a finite size error. We further show that the expectation value of the charge \(M\) approaches an integer value, see Fig. [\[fig:xx_m\]](#fig:xx_m){reference-type="ref" reference="fig:xx_m"}. This is compatible with the emergence of a global \(U(1)\times U(1)\) symmetry and allows us to identify vertex operators \(V_{Q, M}\) according to their eigenvalues of \(Q\) and \(M\). We can also examine the commutation relations in Eq. [\[eq:JmJn\]](#eq:JmJn){reference-type="eqref" reference="eq:JmJn"} in the energy-momentum eigenbasis by calculating matrix elements \(\mel{\psi}{[J_n, J_m]}{\psi}\). The level constant \(k\) can be extracted from \(\mel{\psi }{[J_n, J_{-n}]}{\psi}\), according to Eq. [\[eq:u1comm\]](#eq:u1comm){reference-type="eqref" reference="eq:u1comm"}. For example, we can estimate \(k\) using the three expectation values \(k_1 = \mel{\mathbf{1}}{[J_1, J_{-1}]}{\mathbf{1}}\), \(k_2 = \frac{1}{2} \mel{\mathbf{1}}{[J_2, J_{-2}]}{\mathbf{1}}\) and \(k_3 = \frac{1}{3} \mel{\mathbf{1}}{[J_3, J_{-3}]}{\mathbf{1}}\), as shown in Fig. [\[fig:kfit\]](#fig:kfit){reference-type="ref" reference="fig:kfit"}. Linear extrapolation gives \(k_1 = 1.0000\), \(k_2 = 0.9999\) and \(k_3 = 0.9992\), which are in good agreement with theoretical value \(k=1\). We check the action of \(J_n\) on low energy eigenstates by examining matrix elements \(\mel{\psi_\alpha}{ J_n }{ \psi_\beta}\) (similarly for the anti-chiral part \(\bar{J}_n\)). By numerically checking whether a state is annihilated by \(J_{1}, \bar{J}_{1}\) and \(J_{2}, \bar{J}_{2}\), up to finite size error, we successfully identify all the Kac--Moody primaries. In Fig. [\[fig:spectrum\]](#fig:spectrum){reference-type="ref" reference="fig:spectrum"}, the candidate primary states in the \(Q=0\) and \(Q=1\) sectors are marked. Compared with the analytical result of free boson CFT in Fig. [\[fig:boson\]](#fig:boson){reference-type="ref" reference="fig:boson"}, our lattice generators identify the primary states correctly. By acting on a primary state with the lattice generators \(J_{n}\) and \(\bar{J}_{n}\), we can obtain all the descendent states of that primary state. Up to finite size error, \(J_n\) and \(\bar{J}_n\) act as ladder operators-they change the scaling dimension and conformal spin as illustrated in the example of Eq. [\[eq:ladder\]](#eq:ladder){reference-type="eqref" reference="eq:ladder"}. We plot a few examples of their actions in Fig. [\[fig:spectrum\]](#fig:spectrum){reference-type="ref" reference="fig:spectrum"}. Together we are able to organize all low-energy eigenstates into different Kac--Moody towers. In addition to the case of XX model with \(\gamma = \pi/2\), our proposal can be verified for general XXZ model. We provide numerical results for some other values of \(\gamma\) in Appendix [\[appendix:c\]](#appendix:c){reference-type="ref" reference="appendix:c"}. Up to an overall normalization constant of \(J\) and \(\bar{J}\), we find consistently \(k=1\) and expectation values of \(M\) to be integers, confirming the Kac--Moody algebra.
# Lattice realization of Kac--Moody algebra: the \(SU(2)\) case {#sec:su2}
In this section we generalize the construction of lattice Kac--Moody generators to general Lie groups. We will first introduce the Kac--Moody algebra with general semisimple Lie groups. Then we consider a specific example, the free compactified boson with compactification radius \(R=\sqrt{2}\), which possesses an \(\mathfrak{su}(2)_1\) Kac--Moody symmetry. We then consider critical quantum spin chains with semisimple Lie group symmetry and construct approximate lattice Kac--Moody generators. Finally we present the numerical verification with the Heisenberg model with next-to-nearest neighbor interactions.
## Kac--Moody algebra with a non-Abelian group
We consider a CFT that has an internal semisimple Lie group symmetry \(G\), where the Lie algebra is denoted as \(\mathfrak{g}\). The extended symmetry is generated by *Kac--Moody generators* that form the *Kac--Moody algebra*, denoted as \(\hat{\mathfrak{g}}_k\), where \(k\) is the *level constant* of the Kac--Moody algebra. In a CFT with Kac--Moody algebra, there exists a set of chiral and anti-chiral current operators \(J^{\alpha,\textsl{\tiny CFT}}\) and \(\bar{J}^{\alpha,\textsl{\tiny CFT}}\) with conformal dimensions \((1,0)\) and \((0,1)\), where \(\alpha=1,2,\cdots,\mathrm{dim}~\mathfrak{g}\) labels different currents. The Kac--Moody generators are Fourier modes of the currents, \[\begin{aligned}
\label{eq:JnCFT} J^{\alpha,\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{+inx\frac{2\pi}{L}} J^{\alpha,\textsl{\tiny CFT}}(x) \\ \bar{J}^{\alpha,\textsl{\tiny CFT}}_n &=& \frac{1}{2\pi}\int_0^L dx\, e^{-inx\frac{2\pi}{L}} \bar{J}^{\alpha,\textsl{\tiny CFT}}(x). \end{aligned}\] They satisfy the Kac--Moody algebra \(\hat{\mathfrak{g}}_k\), \[\label{eq:comrel} \begin{aligned} &[\J{m}{\alpha},\J{n}{\beta}] = i \sum_\gamma f^{\alpha \beta \gamma} \J{m+n}{\gamma} + k m \delta^{\alpha \beta} \delta_{m+n, 0}\\ &[\J{m}{\alpha}, \Jbar{n}{\beta}] = 0\\ &[\Jbar{m}{\alpha}, \Jbar{n}{\beta}] = i \sum_\gamma f^{\alpha \beta \gamma} \Jbar{m+n}{\gamma} + k m \delta^{\alpha \beta} \delta_{m+n, 0} \end{aligned}\] where \(f^{\alpha \beta \gamma}\) are the structure constants of the Lie algebra \(\mathfrak{g}\) and \(k\) is the level constant. Setting \(n=m=0\) in Eq. [\[eq:comrel\]](#eq:comrel){reference-type="eqref" reference="eq:comrel"}, we see that the zero modes of the currents form exactly the ordinary Lie algebra \(\mathfrak{g}\) of the global symmetry \(G\), \[= i \sum_\gamma f^{\alpha \beta \gamma} \J{0}{\gamma}. \label{eq:j0comrel}\] Since \(J^{\alpha,\textsl{\tiny CFT}}_0\) and \(\bar{J}^{\alpha,\textsl{\tiny CFT}}_0\) are separately conserved, the global symmetry is \(G\times G\), as it also happened in the Abelian \(U(1)\) case of the previous section. The holomorphic and anti-holomorphic currents transform into each other under spatial parity. We may consider the linear combinations of currents that have definite parity, \[\begin{aligned}
\label{eq:qCFT} q^{\alpha, \textsl{\tiny CFT}}(x) &\equiv& \J{}{\alpha}(x) + \Jbar{}{\alpha}(x) \\ \label{eq:mCFT} m^{\alpha, \textsl{\tiny CFT}}(x) &\equiv& \J{}{\alpha}(x)-\Jbar{}{\alpha}(x). \end{aligned}\] They correspond to charges, \[\begin{aligned}
Q^{\alpha, \textsl{\tiny CFT}} &\equiv& \J{0}{\alpha} + \Jbar{0}{\alpha} \\ M^{\alpha, \textsl{\tiny CFT}} &\equiv& \J{0}{\alpha}-\Jbar{0}{\alpha}. \end{aligned}\] From Eq. [\[eq:j0comrel\]](#eq:j0comrel){reference-type="eqref" reference="eq:j0comrel"}, it follows that charges \(Q^{\alpha, \textsl{\tiny CFT}}\) satisfy the Lie algebra \(\mathfrak{g}\) \[= i \sum_\gamma f^{\alpha \beta \gamma} Q^{\gamma, \textsl{\tiny CFT}}.\]
## Free compactified boson at \(R=\sqrt{2}\)
At \(R=\sqrt{2}\), the free compactified boson has three Virasoro primary operators with conformal dimension \((1,0)\), \[\begin{aligned}
J^{x,\textsl{\tiny CFT}} &=& \frac{1}{\sqrt{2}}(V^{\textsl{\tiny CFT}}_{1,1}+V^{\textsl{\tiny CFT}}_{-1,-1}) \\ J^{y,\textsl{\tiny CFT}} &=& \frac{1}{\sqrt{2}i}(V^{\textsl{\tiny CFT}}_{1,1}-V^{\textsl{\tiny CFT}}_{-1,-1}) \\ J^{z,\textsl{\tiny CFT}} &=& i\partial\phi. \end{aligned}\] Together with three anti-holomorphic currents (with similar expressions), their Fourier modes constitute the \(\mathfrak{su}(2)_1\) algebra, \[\label{eq:struct} \begin{aligned} &[J^{\alpha,\textsl{\tiny CFT}}_m, J^{\beta,\textsl{\tiny CFT}}_n] = \sum_{\gamma} i\sqrt{2}\epsilon^{\alpha\beta\gamma}J^{\gamma,\textsl{\tiny CFT}}_{n+m}+m \delta_{m+n,0} \delta^{\alpha \beta},\\ &[J^{\alpha,\textsl{\tiny CFT}}_m, \bar{J}^{\beta,\textsl{\tiny CFT}}_n] = 0 \\ &[\bar{J}^{\alpha,\textsl{\tiny CFT}}_m, \bar{J}^{\beta,\textsl{\tiny CFT}}_n] =\sum_{\gamma} i\sqrt{2}\epsilon^{\alpha\beta\gamma}\bar{J}^{\gamma,\textsl{\tiny CFT}}_{n+m}+m \delta_{m+n,0} \delta^{\alpha \beta}, \end{aligned}\] where we can read off the structure constants \(f^{\alpha\beta\gamma}=\sqrt{2}\epsilon^{\alpha\beta\gamma}\) and the level constant \(k=1\). For later convenience, we also define \[\begin{aligned} &J^{\pm,\textsl{\tiny CFT}}_n = \frac{1}{\sqrt{2}}(J^{x,\textsl{\tiny CFT}}_n\pm i J^{y,\textsl{\tiny CFT}}_n),\\ &J^{3, \textsl{\tiny CFT}}_n = \frac{1}{\sqrt{2}}J^{z,\textsl{\tiny CFT}}_n \end{aligned}\] While there can be infinite (but countable) Virasoro primary states, in the presence of additional symmetry, the number of primary states with respect to the extended symmetry can be finite. In the case of \(\mathfrak{su}(2)_1\), a Kac--Moody primary state \(\ket{j^\textsl{\tiny CFT}}\) satisfies \[\label{eq:hws} \begin{aligned} J_{n}^{\alpha, \textsl{\tiny CFT}} \ket{j^\textsl{\tiny CFT}} &= 0 \quad \text{for}\quad n>0, \\ J_0^{+, \textsl{\tiny CFT}} \ket{j^\textsl{\tiny CFT}} &= 0,\\ J_0^{3, \textsl{\tiny CFT}} \ket{j^\textsl{\tiny CFT}} &= j\ket{j^\textsl{\tiny CFT}}. \end{aligned}\] where \(j = 0, \frac{1}{2}\). The only two \(SU(2)\) level 1 Kac--Moody primary operators are the identity operator \(\mathbf{1}^{\textsl{\tiny CFT}}\) with \(\Delta^{\textsl{\tiny CFT}}=0\) and \(s^{\textsl{\tiny CFT}}=0\) and \(V^{\textsl{\tiny CFT}}_{1,0}\) with \(\Delta^{\textsl{\tiny CFT}}=1/2\) and \(s^{\textsl{\tiny CFT}}=0\). Other Virasoro primary operators can be obtained by acting with Kac--Moody generators on these primaries, as in the examples shown in Fig. [\[fig:XXX_CFT01\]](#fig:XXX_CFT01){reference-type="ref" reference="fig:XXX_CFT01"}. For instance, in the Kac--Moody tower of the identity, \(|V^{\textsl{\tiny CFT}}_{1,1}\rangle\) and \(|V^{\textsl{\tiny CFT}}_{2,0}\rangle\) can be obtained in the following way, \[\begin{aligned}
\label{eq:kmtower1} J^{+,\textsl{\tiny CFT}}_{-1}|\mathbf{1}^{\textsl{\tiny CFT}}\rangle=|V^{\textsl{\tiny CFT}}_{1,1}\rangle \\ \label{eq:kmtower2} \bar{J}^{+,\textsl{\tiny CFT}}_{-1}|V^{\textsl{\tiny CFT}}_{1,1}\rangle=|V^{\textsl{\tiny CFT}}_{2,0}\rangle. \end{aligned}\] Other Virasoro primary operators with scaling dimensions \(1\) or \(2\) can be obtained in an analogous way. In the Kac--Moody tower of \(V_{1,0}\), \[\begin{aligned}
\label{eq:kmtower3} J^{+,\textsl{\tiny CFT}}_{0}|V^{\textsl{\tiny CFT}}_{-1,0}\rangle=|V^{\textsl{\tiny CFT}}_{0,1}\rangle \\ \label{eq:kmtower4} J^{-,\textsl{\tiny CFT}}_{0}|V^{\textsl{\tiny CFT}}_{1,0}\rangle=|V^{\textsl{\tiny CFT}}_{0,-1}\rangle \\ \label{eq:kmtower5} J^{+,\textsl{\tiny CFT}}_{-2}|V^{\textsl{\tiny CFT}}_{1,0}\rangle=|V^{\textsl{\tiny CFT}}_{2,1}\rangle \\ \label{eq:kmtower6} J^{-,\textsl{\tiny CFT}}_{-2}|V^{\textsl{\tiny CFT}}_{-1,0}\rangle=|V^{\textsl{\tiny CFT}}_{-2,-1}\rangle. \end{aligned}\] Other Virasoro primary operators with scaling dimension \(1/2\) or \(5/2\) can be obtained in a similar way. These relations can be justified by the operator product expansion of vertex operators (see appendix [\[appendix:b\]](#appendix:b){reference-type="ref" reference="appendix:b"}).
## Lattice Kac--Moody generators for general \(G\)
Consider a lattice model with Hamiltonian \(H\) and an on-site symmetry \(G\) with Lie algebra \(\mathfrak{g}\). The symmetry is associated with a set of conserved charges \[Q^{\alpha} =\sum_{i}^{N} q^{\alpha}_i, ~~\alpha=1,2,\cdots,\mathrm{dim}~\mathfrak{g}.\] The charges commute with the Hamiltonian, \([H, Q^{\alpha}] = 0\), and satisfy \[= i f^{\alpha \beta \gamma} Q^{\gamma}.\] The one-site lattice operator \(q^{\alpha}_j\) corresponds to the CFT operator \(q^{\alpha,\textsl{\tiny CFT}}(x)\) in Eq. [\[eq:qCFT\]](#eq:qCFT){reference-type="eqref" reference="eq:qCFT"}, because they are both conserved and have even spatial parity. As in the \(U(1)\) case, we need to find the lattice operator \(m^{\alpha}_j\) corresponding to \(m^{\alpha,\textsl{\tiny CFT}}(x)\) in Eq. [\[eq:mCFT\]](#eq:mCFT){reference-type="eqref" reference="eq:mCFT"}. The derivation that leads to Eq. [\[eq:U1qmCFT\]](#eq:U1qmCFT){reference-type="eqref" reference="eq:U1qmCFT"} goes through for each conserved current \(J^{\alpha,\textsl{\tiny CFT}}(x)\), and we obtain \[\label{eq:qmCFT} i [H^{\textsl{\tiny CFT}},q^{\alpha,\textsl{\tiny CFT}}(x)]=\partial_x m^{\alpha,\textsl{\tiny CFT}}(x),\] which on the lattice transforms to Eq. [\[eq:qm\]](#eq:qm){reference-type="eqref" reference="eq:qm"} below. Then analogous to Eqs. [\[eq:qCFT\]](#eq:qCFT){reference-type="eqref" reference="eq:qCFT"}-[\[eq:mCFT\]](#eq:mCFT){reference-type="eqref" reference="eq:mCFT"}, we can write down the lattice operators that correspond to the chiral current density \(J^{\alpha}_j\) and anti-chiral current density \(\bar{J}^{\alpha}_j\) as in Eq. [\[eq:latticeJi\]](#eq:latticeJi){reference-type="eqref" reference="eq:latticeJi"} below. Finally the lattice Kac--Moody generators \(J^{\alpha}_n\) and \(\bar{J}^{\alpha}_n\) can be constructed as the Fourier modes of \(J^{\alpha}_j\) and \(\bar{J}^{\alpha}_j\).
This completes the construction of lattice Kac--Moody generators for general symmetry \(G\).
## Example: XXX model with next-to-nearest neighbor coupling
Consider an antiferromagnetic spin-1/2 chain with next-to-nearest neighbor coupling \(J_c=0.241167\) \[H = \mathcal{N} \left( \sum_{j=1}^{N} \vec{S}_j \cdot \vec{S}_{j+1} + J_c \sum_{j=1}^N \vec{S_j} \cdot \vec{S}_{j+2} \right)\] where \(S^{\alpha}_j = \frac{1}{2} \sigma_{j}^{\alpha}\) and \(\alpha = x, y, z.\) Overall normalization factor \(\mathcal{N} = 0.856\) is determined by fixing the energy spectrum to be Eq. [\[eq:ECFT\]](#eq:ECFT){reference-type="eqref" reference="eq:ECFT"} in the large-\(N\) limit. This model has three exact conserved charges \(Q^\alpha = \sum_{j=1}^{N} q^\alpha_j = \sum_{j=1}^{N} \sqrt{2} S_{j}^{\alpha}\) which commute with the Hamiltonian: \([Q^{\alpha}, H] = 0\). These charges satisfy \([Q^{\alpha}, Q^{\beta}] = i f^{\alpha \beta \gamma} Q^{\gamma}\) with structure constants \(f^{\alpha \beta \gamma} = \sqrt{2} \epsilon^{\alpha \beta \gamma}\). In the continuum limit, this model corresponds to the \(SU(2)\) WZW model at level \(k=1\). To find the lattice Kac--Moody generators, first we calculate the current \(m^\alpha_j\) using Eq. [\[eq:3local\]](#eq:3local){reference-type="eqref" reference="eq:3local"} \[m_j^\alpha =-f^{\alpha \beta \gamma} \mathcal{N} \left( S_j^{\beta} S_{j+1}^{\gamma} + J_c S_j^{\beta} S_{j+2}^{\gamma} \right).\] Then following Eq. [\[eq:J\]](#eq:J){reference-type="eqref" reference="eq:J"}, we construct \(J_n^{\alpha}\) and \(\bar{J}_n^{\alpha}\) as follows: \[\begin{aligned} J_n^{\alpha} = \sum_{j=1}^{N_0} e^{+ijn \frac{2\pi}{N}} \frac{\sqrt{2}}{2} \left[ S_j^\alpha-\epsilon^{\alpha \beta \gamma} \mathcal{N} \left( S_j^{\beta} S_{j+1}^{\gamma} + J_c S_j^{\beta} S_{j+2}^{\gamma} \right) \right], \\ \bar{J}_n^{\alpha} = \sum_{j=1}^{N_0} e^{-ijn \frac{2\pi}{N}} \frac{\sqrt{2}}{2} \left[ S_j^\alpha + \epsilon^{\alpha \beta \gamma} \mathcal{N} \left( S_j^{\beta} S_{j+1}^{\gamma} + J_c S_j^{\beta} S_{j+2}^{\gamma} \right) \right]. \end{aligned}\] In order to confirm that they correspond to Kac--Moody generators in the continuum, we perform simultaneous exact diagonalization of the above lattice Hamiltonian \(H\) and of the lattice translation operator to obtain the low-energy eigenstates. Scaling dimensions and conformal spins can be computed using Eqs. [\[eq:Elat\]](#eq:Elat){reference-type="eqref" reference="eq:Elat"}-[\[eq:Plat\]](#eq:Plat){reference-type="eqref" reference="eq:Plat"}.
Now we check various matrix elements of the lattice Kac--Moody generators. We start by examining the zero modes of the lattice generators \(J_0^{\alpha}\) and \(\bar{J}_0^{\alpha}\). While \(Q^{\alpha} = J_0^{\alpha} + \bar{J}_0^{\alpha}\) are exact symmetries, the charges \(M^{\alpha} = J_0^{\alpha}-\bar{J}_0^{\alpha}\) do not commute with the Hamiltonian on the lattice, i.e. \([M^{\alpha}, H] \neq 0\). However linear extrapolation of matrix elements \(\mel{\psi}{M^\alpha}{\psi}\) suggests that in the thermodynamic limit each \(M^\alpha\) becomes indeed a conserved charge, as illustrated in Fig. [\[fig:xxx_m\]](#fig:xxx_m){reference-type="ref" reference="fig:xxx_m"}. We also numerically confirmed that the matrix elements \(\mel{\psi}{[Q^\alpha, M^\beta]}{\psi}\) approximately vanish for low energy eigenstates, up to finite size error (not shown in Fig. [\[fig:xxx_m\]](#fig:xxx_m){reference-type="ref" reference="fig:xxx_m"}). Kac--Moody primary states can be identified numerically by checking conditions in Eq. [\[eq:hws\]](#eq:hws){reference-type="eqref" reference="eq:hws"}. Up to finite size error, the only eigenstates that satisfy the conditions in Eq. [\[eq:hws\]](#eq:hws){reference-type="eqref" reference="eq:hws"} are \(\ket{\mathbf{1}}=\ket{V_{0,0}}\) and \(\ket{V_{0, 1}}\). For other Virasoro primary states that are not Kac--Moody primary states, we confirm that they can be obtained as descendants of these two Kac--Moody primary states using the lattice generators. For example, among the four Virasoro primary states \(\ket{V_{0, \pm1}}\) and \(\ket{V_{\pm 1, 0}}\) at conformal spin \(S=0\) and dimension \(\Delta=1/2\), only \(\ket{V_{1, 0}}\) is a Kac--Moody primary state. We show that \(\ket{V_{0, \pm1}}\) are indeed Kac--Moody descendent states of \(\ket{V_{1, 0}}\) by checking matrix elements \(\mel{V_{0, \pm1}}{J^{\pm}_{0}}{V_{1, 0}}\), as plotted in Fig. [\[fig:xxx_matele\]](#fig:xxx_matele){reference-type="ref" reference="fig:xxx_matele"}. Acting on the Kac--Moody primary states with \(J_{-n}^\alpha\) and \(\bar{J}_{-n}^\alpha\), all other descendent states can be reached and classified into Kac--Moody towers. While \(J^3_n\) preserves the eigenvalue of \(J^z\), \(J^{\pm}_{-n}\) changes the eigenvalue of \(J^z\) by \(\pm n\). We illustrate some examples of the actions of \(J^{\alpha}_{-n}\) and \(\bar{J}^{\alpha}_{-n}\) in Fig. [\[fig:spectrum_xxx\]](#fig:spectrum_xxx){reference-type="ref" reference="fig:spectrum_xxx"} and present the finite size scaling results of some corresponding matrix elements in Fig. [\[fig:xxx_matele\]](#fig:xxx_matele){reference-type="ref" reference="fig:xxx_matele"}. Moreover, we are able to confirm not only the level constant \(k\) but also the structure constants \(f^{\alpha \beta \gamma}\). While the structure constants of the diagonal subalgebra are given, it is non-trivial to check the commutation relations of the chiral algebra Eq. [\[eq:struct\]](#eq:struct){reference-type="eqref" reference="eq:struct"}. For example, we confirm \([J^x_{m}, J^y_{n}]=i \sqrt{2}J^z_{m+n}\) in the low energy subspace by calculating the ratio between \(\mel{\psi}{[J^x_{m}, J^y_{n}]}{\psi}\) and \(\mel{\psi}{i \sqrt{2} J^z_{m+n}}{\psi}\). In Fig. [\[fig:xxx_kc\]](#fig:xxx_kc){reference-type="ref" reference="fig:xxx_kc"}, we show that in the thermodynamic limit \(\mel{\mathbf{1}}{J^y_{-1} J^x_{-1}}{\mathbf{1}} \approx 0.9787 \times \left(i \sqrt{2} \mel{\mathbf{1}}{J^z_{-2}}{\mathbf{1}} \right)\), which is in good agreement with Eq. [\[eq:struct\]](#eq:struct){reference-type="eqref" reference="eq:struct"}, up to finite-size effects. Other components of \(f^{\alpha \beta \gamma}\) can be checked numerically in similar fashion using the proposed lattice Kac--Moody generators.
# Discussion {#sec:discussion}
Since the groundbreaking work by Cardy and others in the 1980s, there have been a series of studies on numerically obtaining conformal data from critical lattice systems (see e.g. ). Our paper adds new capabilities and applications to this line of work, by addressing the very important case of quantum critical spin chains with a global symmetry. As we have reviewed, a critical quantum spin chains with a microscopic Lie group symmetry \(G\) corresponds to a CFT with Kac--Moody symmetry. In this paper, we have proposed a concrete construction of Kac--Moody currents \(J_n^{\alpha}\) and \(\bar{J}_n^{\alpha}\) on lattice with the Hamiltonian and its microscopic symmetry as the only input. Our construction allows us to observe the emergence of Kac--Moody symmetries numerically. We illustrated our methods using two spin models, each having a global symmetry with Lie group \(G\), namely the XXZ model as an Abelian example and the XXX (or Heisenberg) model with next-to-nearest neighbor coupling as a non-Abelian example. First, we obtained the low-energy spectrum of the Hamiltonian using exact diagonalization. We then verified our proposal by studying the action of lattice Kac--Moody generators on low-energy energy-momentum eigenstates. As we consider larger system sizes, a second copy of the symmetry is observed to emerge, making the global symmetry \(G\times G\). We also successfully demonstrated the identification of the Kac--Moody primary states and their Kac--Moody towers. We show that the proposal works for both the Abelian and non-Abelian cases, and with independence of whether the model is integrable. To apply our method to larger system sizes, and thus further reduce finite-size corrections in the extracted conformal data, tensor network techniques such as puMPS can be employed. It would also be interesting to study the realizations of other extended symmetries on the lattice, such as the \(\mathcal{W}\)-algebra appearing in the three-state Potts model. *Note:* In subsequent work (which was posted recently ), one of the authors generalized some of the aspects of our work to a more general setting that includes critical systems for which an explicit local Hamiltonian description is not available. | {'timestamp': '2022-06-06T02:17:02', 'yymm': '2206', 'arxiv_id': '2206.01656', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01656'} |
# Introduction
How would this patient's scans look if they had a different [ef]{acronym-label="ef" acronym-form="singular+short"}? How would this [us]{acronym-label="us" acronym-form="singular+short"} view appear if I turned the probe by 5 degrees? These are important causality related questions that physicians and operators ask explicitly or implicitly during the course of an examination in order to reason about the possible pathologies of the patient. In the second case, the interventional query of turning the probe is easy to resolve, by performing the action. However, in the first case, we ask a counterfactual question which cannot be answered directly. Indeed, it falls under the third and highest rung of Pearl's hierarchy of causation. Counterfactual queries probe into alternative scenarios that might have occurred had our actions been different. For the first question of this paper, we ask ourselves how the patient's scans would look if they had a different [ef]{acronym-label="ef" acronym-form="singular+short"}. Here, the treatment would be the different ejection fraction, and the outcome, the different set of scans. Note that this is a query that is *counter-to* our observed knowledge that the patients scans exhibited a specific [ef]{acronym-label="ef" acronym-form="singular+short"}. As such, standard Bayesian Inference, which conditions on the observed data without any further considerations, is not able to answer this type of question. **Related works**: Generating synthetic [us]{acronym-label="us" acronym-form="singular+short"} images can be performed with physics-based simulators and other techniques, such as registration-based methods. However, these methods are usually very computationally expensive and do not generate fully realistic images. With the shift toward deep learning, [gan]{acronym-label="gan" acronym-form="singular+short"}-based techniques have emerged. They can be based on simulated [us]{acronym-label="us" acronym-form="singular+short"} priors or other imaging modalities (MRI, CT) to condition the anatomy of the generated [us]{acronym-label="us" acronym-form="singular+short"} images. Recently, many works explore machine learning as a tool to estimate interventional conditional distributions . However, fewer works focus on the counterfactual query estimation. explore the Abduction-Action-Prediction paradigm and use deep neural networks for the abduction step, which is computationally very expensive. derive a parametric mathematical model for the estimation of one of the probabilities of causation, while use to develop deep twin networks. The computer vision field also has a lot of interest in conditional generation problems. perform conditional video generation from a video and a discrete class. uses an image and a continuous value as input to produce a new image. introduce causality in their generation process, to produce images from classes. **Contributions**: In this paper (1) We extend the causal inference methodology known as Deep Twin Networks into a novel generative modelling method ([dtgn]{acronym-label="dtgn" acronym-form="singular+short"} [^1]) able to handle counterfactual queries. (2) We apply our framework on the synthetic MorphoMNIST and real-world EchoNet-Dynamic datasets, demonstrating that our method can perform well in both fully controlled environments and on real medical cases. To the best of our knowledge, this is an entirely novel approach and task, and thus the first time such an approach is explored for medical image analysis and computer vision. Our work differentiates itself from all other generative methods by combining video generation with continuous conditional input in a new causal framework. This setup supports counterfactual queries to produce counterfactual videos using a semi-supervised approach which allows most standard labelled datasets to be used.
# Preliminaries
**Structural causal models** [\[Section: Structural causal models\]]{#Section: Structural causal models label="Section: Structural causal models"} We work in the Structural Causal Models (SCM) framework. Chapter \(7\) of gives an in-depth discussion. For an up-to-date, review of counterfactual inference and Pearl's Causal Hierarchy, see. As a brief reminder, we define a structural causal model (SCM) as a set of latent noise variables \(U=\{u_1,\dots,u_n\}\) distributed as \(P(U)\), a set of observable variables \(V=\{v_1,\dots, v_m\}\) that is the superset of treatment variables \(X\) and confounders \(Z\), *i.e.* \(X,Z \subseteq V\). Moreover, we require a directed acyclic graph (DAG), called the *causal structure* of the model, whose nodes are the variables \(U\cup V\), and its edges represent a collection of functions \(F=\{f_1,\dots, f_n\}\), such that \(v_i = f_i(PA_i, u_i), \text{ for } i=1,\dots, n,\) where \(PA\) denotes the parent observed nodes of an observed variable. These are used to induce a distribution over the observable variables and assign uncertainty over them. Finally, the \(do\)-operator forces variables to take certain values, regardless of the original causal mechanism. Graphically, \(do(X=x)\) means deleting edges incoming to \(X\) and setting \(X=x\). Probabilities involving \(do(x)\) are normal probabilities in submodel \(M_x\): \(P(Y=y \mid \text{do}(X=x)) = P_{M_x} (y)\). **Counterfactual inference** [\[Section: two approaches to counterfactual inference\]]{#Section: two approaches to counterfactual inference label="Section: two approaches to counterfactual inference"} The latent distribution \(P(U)\) allows us to define probabilities of counterfactual queries, \(P(Y_{y}=y) = \sum_{u \mid Y_{x}(u)=y} P(u)\). Moreover, one can define a counterfactual distribution given seemingly contradictory evidence and thereby state the counterfactual sentence "\(Y\) would be \(y\) (in situation \(U=u\)), had \(X\) been \(x\)". Despite the fact that this query may involve interventions that contradict the evidence, it is well-defined, as the intervention specifies a new submodel. Indeed, \(P(Y_{x}={y}' \mid E = {e})\) is given by as \(\sum_{{u}} P(Y_{{x}}({u})={y}')P({u}|{e})\.\) **Twin Networks** As opposed to the standard *Abduction-Action-Prediction* paradigm, we will be operating under the Twin Model methodology. Originally proposed by Balke and Pearl in, this method allows efficient counterfactual inference to be performed as a feed forward Bayesian process. It has also been shown empirically to offer computational savings relative to abduction-action-prediction . A twin network consists of two interlinked networks, one representing the real world and the other the counterfactual world being queried. Given a structural causal model, a twin network can be constructed and used to compute a counterfactual query through the following steps: First, duplicate the given causal model, denoting nodes in the duplicated model via superscript \(^*\). Let \(V\) be observable nodes that include the treatment variables \(X\) and the confounders \(Z\), and let \(X^*,Z^* \subseteq V^{*}\) be the duplication of these. Also let \(U\) be the unobserved latent noise. Then, for every node \(v_i^{*}\) in the duplicated, or "counterfactual" model, its latent parent \(u_i^{*}\) is replaced with the original latent parent \(u_i\) in the original, or "factual", model, such that the original latent variables are now a parent of two nodes, \(v_i\) and \(v_i^{*}\). The two graphs are linked only by common latent parents, but share the same node structure and generating mechanisms. To compute a general counterfactual query \(P(Y=y \mid E=e, \text{do}(X=x))\), modify the structure of the counterfactual network by dropping arrows from parents of \(X^*\) and setting them to value \(X^*=x\). Then, in the twin network with this modified structure, compute the probability \(P(Y^*=y \mid E=e, X^*=x)\) via standard Bayesian inference techniques, where \(E\) are factual nodes.
# Method {#sec:dtgn}
**Deep Twin Networks** The methodology we propose is based on Deep Twin Networks. The training procedure and parametrization are borrowed from, which sets the foundation for our causal framework. Deep Twin Networks use two branches, the factual and the counterfactual branch. We denote our factual and counterfactual treatments (inputs unique to each branch) as \(X\) and \(X^*\), while the confounder (input shared between both branches) is denoted by \(Z\). We denote the factual and counterfactual outcomes as \(\hat{Y}, \hat{Y}^*\) while the noise, injected midway through the network, and shared by both branches, is denoted by \(U_Y\). This information flow sets \(Z\) as the data we want to query with \(X\) and \(X^*\), to produce the outcomes \(\hat{Y}\) and \(\hat{Y}^*\). These variables will be detailed on a case-specific basis in . See Fig.[\[figure: deep twin netork architectures_TN\]](#figure: deep twin netork architectures_TN){reference-type="ref" reference="figure: deep twin netork architectures_TN"} for a visual representation of the information flow. **Synthetic data** Synthetic data allows for full control over the generation of both ground truth outcomes \(Y\) and \(Y^*\) along with their corresponding inputs \(X\), \(X^*\) and \(Z\). This makes training the Deep Twin Network trivial in a fully supervised fashion, as demonstrated in our first experiment. **Real-world data** Theoretically, our approach requires labelled data pairs, but very few datasets are arranged as such. To overcome this limitation and support most standard labeled imaging datasets, we establish a list of features that our model must possess to generate counterfactual videos for the medical domain: (1) produce a factual and counterfactual output that share general visual features, such as style and anatomy, (2) produce accurate factual and counterfactual videos with respect to the treatment variable, and (3) the produced videos must be visually indistinguishable from real ones. In the following, we use the Echonet-Dynamic dataset to illustrate the method, see for details. We solve feature (1) by sharing the weights of the branches in the network, such that we virtually train a single branch on two tasks in parallel. To do so, we set the confounder \(Z\) as the input video and the treatment \(X\) as the labelled [ef]{acronym-label="ef" acronym-form="singular+short"}. We train the network to match the factual outcome \(\hat{Y}\) with the input video. By doing so, the model learns to retain the style and anatomical structure of the echocardiogram from the confounder. This is presented as *Loss 1* in . For feature (2) we pre-train an expert network to regress the treatment values from the videos produces by the counterfactual branch, and compare them to the counterfactual treatment. The expert network's weights are frozen when training the Twin model, and alleviates the need for labels to train the counterfactual branch. This loss is denoted as *Loss 2* in . Finally, feature (3) calls for the well-known [gan]{acronym-label="gan" acronym-form="singular+short"} framework, where we train a neural network to discriminate between real and fake images or videos, while training the Twin Network. This constitutes the adversarial *Loss 3* in and ensures that the counterfactual branch produces realistic-looking videos. With those 3 losses, we can train the generator (i.e. factual and counterfactual branches) to produce pairs of visually accurate and anatomically matching videos that respect their factual and counterfactual [ef]{acronym-label="ef" acronym-form="singular+short"}s treatment. To learn the **noise distribution** \(U_Y\), we follow and without loss of generality we can write \(Y=f(X,Z,g(U_Y'))\) with \(U_Y'\sim \mathcal{E}\) and \(U_Y = g(U_Y')\), where \(\mathcal{E}\) is some easy-to-sample-from distribution, such as a Gaussian or Uniform. Effectively, we cast the problem of determining \(U_Y\) to learning the appropriate transformation from \(\mathcal{E}\) to \(U_Y\). For ease of understanding, we will be using \(U_Y\) henceforth to signify our approximation \(g(U_Y')\) of the true unobserved parameter \(U_Y\). In addition to specifying the causal structure, the following standard assumptions are needed to correctly estimate \(\mathbb{E}(Y | do(X), Z)\): (1) *Ignorability:* there are no unmeasured confounders; (2) *Overlap:* every unit has non-zero probability of receiving all treatments given their observed covariates.
# Experimentation {#sec:experiments}
**Datasets** To evaluate [dtgn]{acronym-label="dtgn" acronym-form="singular+short"}, we use two publicly available datasets, the synthetic MorphoMNIST and the clinical Echonet-Dynamic dataset. MorphoMNIST is a set of tools that enable fine-grained perturbations of the MNIST digits through four morphological functions, as well as five measurements of the digits. To train our model, we need five elements: an original image, a (counter-)factual treatment \(X\)(\(X^*\)) and a corresponding (counter-)factual label \(Y\)(\(Y^*\)). To generate this data, we take 60,000 MNIST images \(I_i\) and sample 40 perturbation vectors \(p_{i,j}\) for the five possible perturbations, including identity, thus generating 2.4 million images \(I_{p_{i,j}}\). The perturbation vectors also encode the relative positions of the perturbations, when applicable. We *measure* the original images to produce vectors \(m_i\) and one-hot encode the labels into vectors \(l_i\). We decided to perform the causal reasoning over a latent space, rather than image space. To do so, we train a [vqvae]{acronym-label="vqvae" acronym-form="singular+short"} to project the MorphoMNIST images to a latent space \(\mathcal{H} \in \mathbb{R}^{(q \times h \times w)}\) and reconstruct them. Once trained, the [vqvae]{acronym-label="vqvae" acronym-form="singular+short"} weights are frozen, and we encode all the ground-truth perturbed images \(I_{p_{i,j}}\) into a latent embedding \(\mathcal{H}_{i,j}\). Afterwards, the [vqvae]{acronym-label="vqvae" acronym-form="singular+short"} is used to reconstruct the generated latent embeddings \(\hat{Y}, \hat{Y}^*\) for qualitative evaluation purposes. The clinical dataset Echonet-Dynamic consists of 10,030 4-chamber echocardiography videos with \(112 \times 112\) pixels resolution and various length, frame rates, image quality and cardiac conditions. Each video contains a single pair of consecutively labelled [es]{acronym-label="es" acronym-form="singular+short"} and [ed]{acronym-label="ed" acronym-form="singular+short"} frames. Each video also comes with a manually measured [ef]{acronym-label="ef" acronym-form="singular+short"}. For our use case, all videos are greyscaled and resampled to 64 frames, with a frame rate of 32 images per second. All videos shorter than two seconds are discarded, and we make sure to keep the labelled frames. For the resulting 9724 videos dataset, the original split is kept, with 7227 training, 1264 validation and 1233 testing videos. **MorphoMNIST** For our synthetic experiment, we define a deep twin network as in and follow the process we described in the Methods (). Regarding data organization, we use the elements defined in the section above. We set our confounder \(Z_i = [l_i, m_i]\) to contain the one-hot encoded labels as well as the measurement of the original image. We sample two perturbation vectors \(p_{i,m}\) and \(p_{i,n}\) and their corresponding latent embeddings \(\mathcal{H}_{i,m}\) and \(\mathcal{H}_{i,n}\), where \(n, m \in \llbracket 0, 40\llbracket\), \(n \neq m\). We set our input treatments as the perturbations vectors (\(X = p_{i,m}\), \(X^* = p_{i,n}\)) and our ground-truth outcomes as the corresponding latent embeddings of the perturbed images (\(Y=\mathcal{H}_{i,m}\), \(Y^*=\mathcal{H}_{i,n}\)). We sample \(U_Y \sim [ \mathcal{N}(0 ,0.25)\bmod{1} + 1 ]\) and disturb the output of the branches of the neural networks that combine \(X\) and \(X^*\) with \(Z\) by multiplying the outputs of both with the same \(U_Y\). With this setup, we generate factual and counterfactual perturbed MNIST embeddings from a latent description. We assess the quality of the results by three means: (1) *Embeddings' MSE:* We sample a quintuplet (\(Z\), \(X\), \(X^*\), \(Y\), \(Y^*\)) and 1000 \(U_Y\). The MSE between all \(\hat{Y}_{i}\) and \(Y\) are computed and used to order the pairs (\(\hat{Y}_i\), \(\hat{Y}^*_i\)) in ascending order. We keep the sample with the lowest MSE as our factual estimate and compute the MSE between \(\hat{Y}^*_0\) and \(Y^*\) to get our counterfactual MSE score. (2) *SSIM:* We use the Structural SIMilarity metric between the perturbed images \(I_{gt} = I_{p_{i,j}}\), the images reconstructed by the VQVAE \(I_{rec}\) and the images reconstructed from the latent embedding produced by the twin network \(I_{pred}\) to get a quantitative score over the images. (3) *Images:* We sample some images to qualitatively assess best and worst cases scenarios for this framework. We show the quantitative results in (a), and qualitative results in and in the appendix.
:::
[\[tab:metrics\]]{#tab:metrics label="tab:metrics"}
**Echonet Dynamic** As stated in , our methodology requires an **Expert Model** to predict the [ef]{acronym-label="ef" acronym-form="singular+short"} of any [us]{acronym-label="us" acronym-form="singular+short"} video. To do so, we re-implement the ResNet 2+1D network as it was shown to be the best option for [ef]{acronym-label="ef" acronym-form="singular+short"} regression in. We opt not to use transformers as they do not supersede convolutions for managing the temporal dimension, as shown in. We purposefully keep this model as small as possible in order to minimize its memory footprint, as it will be operating together with the generator and the frame discriminator. The expert network is trained first, and frozen while we train the rest of [dtgn]{acronym-label="dtgn" acronym-form="singular+short"}. Metrics for this model are presented in the Appendix. **[dtgn]{acronym-label="dtgn" acronym-form="singular+short"}** We implement the generator as described in . We define a single deep network to represent both the factual and counterfactual paths. By doing so, we can meet the objectives listed in . The branch is implemented as a modified ResNet 2+1D to generate videos. It takes two inputs: a continuous value and a video, where the video determines the size of the output. For additional details, please refer to and the code. **Discriminator** We build a custom discriminator architecture using five "residual multiscale convolutional blocks\", with kernel sizes 3, 5, 7 and appropriate padding at each step, followed by a max-pooling layer. Using multiscale blocks enables the discriminator to look at both local and global features. This is extremely important in [us]{acronym-label="us" acronym-form="singular+short"} images because of the noise in the data, that needs to be both ignored, for anatomical identification, and accounted for to ensure that counterfactual [us]{acronym-label="us" acronym-form="singular+short"} images look real. We test this discriminator both as a frame-based discriminator and a video-based discriminator, by changing the 2D layers to 3D layers where appropriate. We note that, given our architecture, the 3D version of the model requires slightly less memory but doubles the processing power compared to the 2D model. **Training the framework** At each training step, we sample an [us]{acronym-label="us" acronym-form="singular+short"} video (\(V\)) and its [ef]{acronym-label="ef" acronym-form="singular+short"} (\(\psi\)). We set our factual treatment, \(X = \psi\) and our counterfactual treatment \(X^* \sim \mathcal{U}(0,\psi-0.1) \cup \mathcal{U}(\psi+0.1,1)\). The confounder \(Z\) and the factual ground truth \(Y\) are set to the sampled video such that \(Z = Y = V\). We compute an L1 reconstruction loss (loss 1) between \(\hat{Y}\) and \(Y = V\). As we do not have ground truth for the counterfactual branch, we use the frame discriminator and the expert model to train it. Both models take as input the counterfactual prediction \(\hat{Y}^*\). The expert model predicts an [ef]{acronym-label="ef" acronym-form="singular+short"} \(\hat{\psi}\) that is trained to match the counterfactual input \(X^*\) with an L1 loss (loss 2). The discriminator is trained as a [gan]{acronym-label="gan" acronym-form="singular+short"} discriminator, with \(\hat{Y}^*\) as fake samples and \(V\) as real samples. It trains the generator with L1 loss (loss 3). The discriminator and expert model losses are offset respectively by three and five epochs, leaving time for the generator to learn to reconstruct \(V\), thus maintaining the anatomy and style of the confounder. Our experiments show that doing so increases the speed at which the network is capable of outputting realistic-looking videos, thus speeding up the training of the discriminator that sees accurate fake videos sooner. Once the generator and discriminator are capable of generating and discriminating realistic-looking videos, the expert network loss is activated and forces the generator to take into account the counterfactual treatment, while the factual treatment is enforced by the reconstruction loss. The losses are also scaled, such that the discriminator loss has a relative weight of 3 compared to the reconstruction and expert loss. **Metrics** To match our three objectives, we evaluate this new task on 1) the accuracy of the anatomy and 2) the accuracy of the regressed [ef]{acronym-label="ef" acronym-form="singular+short"}. We obtain the best possible video by sampling 100 \(U_Y\) and keeping the \(\hat{Y}^*\) with \(\hat{\phi}^*\) closest to \(X^*\). We then evaluate our metrics over those "best" videos. The anatomical accuracy is measured with SSIM and the [ef]{acronym-label="ef" acronym-form="singular+short"} precision is measured using R2, MAE and RMSE scores. Results are shown in . **Qualitative** In we show video frame examples to showcase the quality of the reconstructed frames, as well as how the anatomy and style are maintained. **Discussion** The predicted [ef]{acronym-label="ef" acronym-form="singular+short"} has an MAE of 15.7% which is not optimal. This can come from many factors such as the use of the same dataset to train the Expert model and [dtgn]{acronym-label="dtgn" acronym-form="singular+short"}, the limited number of real videos, or the limited size of the networks due to the necessity of working with three models at once. Those problems could be addressed with hyperparameter search, larger models, as well as additional medical data. For completeness, we compare the performance of [dtgn]{acronym-label="dtgn" acronym-form="singular+short"} with the literature , and run additional experiments (see Appendix) with an ablated version of our model for conditional video generation, where it achieves the best SSIM score of 0.72.
# Conclusion
In this paper we introduce [dtgn]{acronym-label="dtgn" acronym-form="singular+short"}, a Deep Twin Generative Network able to produce counterfactual images and videos. We showcase its performance in both a synthetic and a real world medical dataset and achieve visually accurate results and high quantitative scores. In future work, we will explore other treatments, such as changing the heartbeat, as well as less constrained confounders, e.g. image segmentations.\
**Acknowledgements:** This work was supported by Ultromics Ltd., the UKRI Centre for Doctoral Training in Artificial Intelligence for Healthcare (EP/S023283/1), and the UK Research and Innovation London Medical Imaging and Artificial Intelligence Centre for Value Based Healthcare. We thank the NVIDIA corporation for their GPU donations used in this work.
[^1]: D'Artagnan is the fourth Musketeer from the French tale "The three Musketeers". | {'timestamp': '2022-07-01T02:18:23', 'yymm': '2206', 'arxiv_id': '2206.01651', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01651'} |
null | null |
# Introduction
The core and one of the most expensive operations in many machine learning applications is matrix multiplication. Performing such operations locally on a single computer takes a long time. Users would consider outsourcing their matrices to a distributed system for time-sensitive applications to carry out demanding computation tasks. Efficient methods require coding over the input matrices to speed up the computational time, yielding a trade-off among the number of workers needed, tasks performed at each worker, and the total amount of data transmitted. While outsourcing disrupts the computational delays, it leads to data security concerns. This paper aims to develop an efficient, secure distributed matrix multiplication (SDMM) scheme which keeps matrices secure from the potentially colluding servers. We consider the problem of secure distributed matrix multiplication (SDMM), where a user has two matrices, \(A \in \mathbb{F}_q^{a \times b}\), \(B \in \mathbb{F}_q^{b \times c}\) and wishes to compute their product, \(AB \in \mathbb{F}_q^{a \times c}\), with the assistance of \(N\) servers, without leaking any information about either \(A\) or \(B\) to any server. We assume that all servers are honest but curious (passive) in that they are not malicious and will follow the pre-agreed upon protocol. However, any \(T\) of them may collude to eavesdrop and extrapolate information regarding \(A\) or \(B\). The setting considered in this paper is proposed in , with many follow-up works . The initial performance metric used was the download cost, meaning the total amount of data downloaded by the users from the server. Following papers have also considered the upload costs, the total communication costs, and computational costs. Different partitions of the matrices lead to different trade-offs between upload and download costs. In this paper, we consider the most general one, namely, the grid product partition given by \[A = \begin{bmatrix} A_{i,j} \end{bmatrix}_{\substack{1\leq i\leq t\\ 1\leq j\leq s}}, B = \begin{bmatrix} B_{i,j} \end{bmatrix}_{\substack{1\leq i\leq s\\ 1\leq j\leq d}}\] such that, \[\displaystyle AB = \begin{bmatrix} M_{1,1} & \cdots & M_{1,d}\\ \vdots & \ddots & \vdots\\ M_{t,1} & \cdots & M_{t,d}\\\end{bmatrix}=\begin{bmatrix} \sum_{\ell=1}^s A_{i,\ell} B_{\ell,j}\end{bmatrix}_{\substack{1\leq i\leq t\\ 1\leq j\leq d}},\] where the products \(A_{i,\ell} B_{\ell, j}\) are well-defined and of the same size. Under this partition, a polynomial code is a polynomial \(h(x)=f_A(x) \cdot f_B(x)\), whose coefficients encode the submatrices \(A_{i, j}B_{k,\ell}\). The next step is where the scheme we propose differs from previous works that use the grid product partition. The user evaluates polynomials \(f_A(x)\) (encoding matrix \(A\)) and \(f_B(x)\) (encoding matrix \(B\)) at powers \(\alpha_N, \alpha^2_N, \ldots, \alpha^N_N=1\) of an \(N\)-th root of unity \(\alpha_N\). The \(N\) servers compute the product \(h(\alpha^i)=f_A(\alpha^i)f_B(\alpha^i)\) for \(i\in\{1,\ldots,N\}\). The polynomial \(h(x)\) is constructed so that no \(T\)-subset of evaluations reveals any information about \(A\) or \(B\) (\(T\)-security), but so that the user can reconstruct \(AB\) given all \(N\) evaluations (decodability). An example of a polynomial scheme for the grid product partition is the secure MatDot codes in and the entangled polynomial codes in . In Theorem [\[theo:scheme\]](#theo:scheme){reference-type="ref" reference="theo:scheme"}, we characterize the total communication rate achieved by our proposed scheme:
In Theorem [\[theo:comparison\]](#theo:comparison){reference-type="ref" reference="theo:comparison"}, we show that our proposed code matches the recovery threshold of the best-known scheme for inner product partition, and also matches GASP codes for outer product partition when \(T<t\).
## Related Work
For distributed computations, polynomial codes were initially introduced in to mitigate stragglers in distributed matrix multiplication. A series of works followed this, . The literature on SDMM has also studied different variations of the model we focus on here. For instance, in the encoder and decoder are considered to be separate, in servers are allowed to cooperate. In the authors consider a hybrid setup between SDMM and private information retrieval where the user has a matrix \(A\) and wants to multiply it with a matrix \(B\) belonging to some public list privately.
## Main Contributions
Our main contributions are summarized below.
- We present a generalization for polynomial coding in the context of the secure distributed matrix multiplication problem, considering the grid product partition. This partition allows extending the use of techniques to reduce upload costs.
- By adapting the Fourier Discrete Transform used in, we present a new scheme for SDMM. It reduces the communication load by lowering the recovery threshold. We show that they are secure, decodable, and present their total communication rate in Theorem [\[theo:scheme\]](#theo:scheme){reference-type="ref" reference="theo:scheme"}.
- In Theorem [\[theo:comparison\]](#theo:comparison){reference-type="ref" reference="theo:comparison"}, we show that *the proposed scheme* matches the recovery threshold of the best-known scheme for inner product partition,, and also matches the GASP scheme for outer product partition when \(T<t\).
# A Motivating Example: \(d=t=2\) and \(T=1\) {#sec2}
We begin the description of our proposed scheme with an example, which we present to showcase our scheme. At the end of the section, we assume that each server can compute \(\frac{abc}{4}\) scalar operations, meaning additions or multiplications in \(\mathbb{F}_q\). Finally, we compare the proposed method with GASP codes that use an outer partition product and the one using an inner partition product. In this example, a user wishes to multiply two matrices \(A \in \mathbb{F}_{q}^{a \times b}\) and \(B \in \mathbb{F}_{q}^{b \times c}\) with the assistance of non-colluding helper servers. Consider the following matrices are partitioned as follows \[A=\begin{bmatrix} A_{1,1} & A_{1,2}\\mathcal{A}_{2,1} & A_{2,2} \end{bmatrix}\in \mathbb{F}_q^{a\times b}, \ B=\begin{bmatrix} B_{1,1} & B_{1,2}\\B_{2,1} & B_{2,2} \end{bmatrix}\in \mathbb{F}_q^{b\times c}\] By multiplying the matrices we obtain \[M=AB=\begin{bmatrix} A_{1,1}B_{1,1}+A_{1,2}B_{2,1} & A_{1,1}B_{1,2}+A_{1,2}B_{2,2}\\mathcal{A}_{2,1}B_{1,1}+A_{2,2}B_{2,1}&A_{2,1}B_{1,2}+A_{2,2}B_{2,2}. \end{bmatrix}\] Since we assume non-colluding servers, i.e, \(T=1\), it involves picking two random matrices \(R \in \mathbb{F}_{q}^{\frac{a}{2} \times \frac{b}{2}}\) and \(S \in \mathbb{F}_{q}^{\frac{b}{2} \times \frac{c}{2}}\). Consider the (Laurent) polynomials \[\begin{aligned}
f_A(x)&=A_{1,1} + A_{1,2}x + A_{2,1}x^2+ A_{2,2}x^3 + R x^4,\\ &\mbox{ and } \\ f_B(x)&=B_{1,1} + B_{2,1}x^{-1} + B_{1,2}x^{-5}+ B_{2,2}x^{-6} + S x^{-10} \end{aligned}\] We obtain the following degree table for polynomial \(h(x) = f_A(x)f_B(x)\): \[\begin{array}{c|ccccc} + & 0 & 1 & 2 &3& 4 \\ \hline 0& {\color{blue}0 }& 1 & {\color{green}2} & 3 & 4 \\-1 &-1 & {\color{blue}0} & 1 & {\color{green}2} & 3 \\-5 & {\color{red}-5} &-4 & {\color{orange}-3} &-2&-1 \\-6 &-6 & {\color{red}-5} &-4 & \color{orange}{-3} &-2 \\-10 &-10 &-9 &-8 &-7 &-6 \\\end{array}\] leading to a problem to find evaluations points \(\mathbb{F}_q\) that minimizes the set \(\{\alpha^{i}: i\in \{-10,-9, \ldots, 4\} \}\) under the following conditions:
- \(|\{\alpha^{-5}, \alpha^{-3}, \alpha^{0}, \alpha^{2}\}|=4\)
- \(\{\alpha^{-5}, \alpha^{-3}, \alpha^{0}, \alpha^{2}\} \cap \{\alpha^{-10},\ldots, \alpha^{-6}, \alpha^{-4}, \\ \mbox{\hspace{4.5cm}}\alpha^{-2}, \alpha^{-1}, \alpha, \alpha^{3}, \alpha^{4}\}=\emptyset\)
Consequently, if \(\alpha=\alpha_{13}\) is an \(13\)-th root of unity, such a condition is satisfied.
## Computational complexity
Let \(\alpha_{13}\in \mathbb{F}_q\) be a primitive root of unity. The algorithm for computing the multiplication is as follows
1. **Encode.** For \(i=1,\ldots,13\), the user computes \(f_A(\alpha_{13}^i)\) and \(f_B(\alpha_{13}^i)\).
2. **Upload.** The user sends matrices \(f_A(\alpha_{13}^i)\) and \(f_B(\alpha_{13}^i)\) to Server \(i\).
3. **Server multiplication.** Servers multiply together the received matrices.
4. **Download.** Servers send the result \(f_A(\alpha_{13}^i)\cdot f_B(\alpha_{13}^i)\) back to the user.
5. **Decode.** The user uses Equation [\[prop:rootofunity\]](#prop:rootofunity){reference-type="ref" reference="prop:rootofunity"} to obtain the coefficients with degree \(-5\), \(-3\), \(0\), and \(2\) or, equivalently, \(0\), \(2\), \(8\), and \(10\) since polynomials are evaluated at an 13-th root of unity \(\alpha_{13}\). Therefore, \[\begin{aligned}
A_{1,1}B_{1,1}+A_{1,2}B_{2,1}&=\frac{1}{13}\sum_{i=1}^{13}f_A(\alpha_{13}^i)\cdot f_B(\alpha_{13}^i)\\ A_{1,1}B_{1,2}+A_{1,2}B_{2,2}&=\frac{1}{13}\sum_{i=1}^{13}(\alpha_{13}^i)^5f_A(\alpha_{31}^i)\cdot f_B(\alpha_{13}^i)\\ A_{2,1}B_{1,1}+A_{2,2}B_{2,1}&=\frac{1}{13}\sum_{i=1}^{13}(\alpha_{13}^i)^{11}f_A(\alpha_{13}^i)\cdot f_B(\alpha_{13}^i)\\ A_{2,1}B_{1,2}+A_{2,2}B_{2,2}&=\frac{1}{13}\sum_{i=1}^{13}(\alpha_{13}^i)^3f_A(\alpha_{13}^i)\cdot f_B(\alpha_{13}^i) \end{aligned}\]
We start with the assumption that addition and multiplication in \(\mathbb{F}_q\) take constant time. We consider for simplicity that parameters \(a\), \(b\), and \(c\) are divisible by \(2\). We describe below here the complexities of each step:
1. Computing \(f_A(\alpha_{13}^i)\) and \(f_B(\alpha_{13}^i)\) requires \(2ab\) and \(2bc\) \(\mathbb{F}_q\)-operations, respectively. This translates to \(26(ab+bc)\) \(\mathbb{F}_q\)-operations to compute the \(13\) evaluations.
2. The user sends \(\frac{13}{4}(ab+bc)\) \(\mathbb{F}_q\)-elements to the servers.
3. The computational cost to perform the product \(f_A(\alpha_{13}^i)f_B(\alpha_{13}^i)\) on each server is \(\frac{ac(b-1)}{4}\).
4. Each server sends \(\frac{ac}4\) \(\mathbb{F}_q\)-elements to the user.
5. The decoding step requires up to \(\frac{13ac}{2}\) \(\mathbb{F}_q\)-operations to obtain each coefficient of interest. Since \(3\) of those requires exactly \(\frac{13ac}{2}\) and one requires \(\frac{14ac}{4}\), then in total, we need \(23ac\) \(\mathbb{F}_q\)-operations are required to retrieve the desired product \(AB\).
[\[tab:compare\]]{#tab:compare label="tab:compare"}
In Table [\[tab:compare\]](#tab:compare){reference-type="ref" reference="tab:compare"}, we present a comparative summary for this example among the proposed method, GASP (which uses outer product partition), and the scheme shown in (which inner product partition). For this comparison, we fixed the amount of \(\mathbb{F}_q\)-operations in each server by \(\frac{ac(b-1)}{4}\); therefore, we shall assume parameters \(a\), \(b\) and \(c\) are divisible by \(4\).
# Proposed Scheme
This section is devoted to presenting the general construction of the proposed scheme. We perform the same technique as in Section [2](#sec2){reference-type="ref" reference="sec2"} retrieving the \(dt\) matrices \(\sum_{\ell=1}^sA_{i,\ell} B_{\ell,j}\) from the polynomial \(h(x) = f_A(x) \cdot f_B(x)\). **Choosing the Polynomials:** As described in the introduction, the user partitions the matrices \(A \in \mathbb{F}_q^{a \times b}\) and \(B \in \mathbb{F}_q^{b \times c}\) as \(A = \begin{bmatrix} A_{i,j} \end{bmatrix}_{\substack{1\leq i\leq t\\ 1\leq j\leq s}}\) \(B = \begin{bmatrix} B_{i,j} \end{bmatrix}_{\substack{1\leq i\leq s\\ 1\leq j\leq d}}\) with the purpose of getting the matrix multiplication expressed as \[\displaystyle AB =\begin{bmatrix} \sum_{\ell=1}^s A_{i,\ell} B_{\ell,j}\end{bmatrix}_{\substack{1\leq i\leq t\\ 1\leq j\leq d}},\] where \(A_{i,j}\in\mathbb{F}_q^{\frac{a}{t} \times \frac{b}{s}}\) and \(B_{i,j}\in\mathbb{F}_q^{\frac{b}{s}\times \frac{c}{t}}\). To obtain \(T\)-security \(R_1, \ldots, R_T \in \mathbb{F}_q^{\frac{a}{t} \times \frac{b}{s}}\) and \(S_1, \ldots, S_T \in \mathbb{F}_q^{\frac{b}{s}\times \frac{c}{t}}\) are chosen independently and uniformly at random. We then define \(f_A\in \mathbb{F}_q^{\frac{a}{t} \times \frac{b}{s}}[x,x^{-1}]\) and \(f_B\in \mathbb{F}_q^{\frac{b}{s}\times \frac{c}{t}}[x,x^{-1}]\) as the following polynomials \[\label{eq:encfunction} f_A(x) = \sum_{i=1}^t\sum_{j=1}^s A_{i,j}x^{(i-1)s+j-1} + \sum_{k=1}^{T} R_kx^{ts+k-1},\] \[\begin{gathered}
f_B(x) = \sum_{i=1}^s\sum_{j=1}^d B_{i,j}x^{(1-j)(ts+T)+(1-i)}+ \\+\sum_{k=1}^{T} S_kx^{(-d)(ts+T)-k+1}. \end{gathered}\] **Choosing the Field and Evaluation Points:** Let \(J\subset \mathbb{Z}\) be a finite set and \(p(x)=\sum_{i\in J}M_ix^i\in \mathbb{F}_q^{m_1\times m_2}[x,x^{-1}]\). Define the support set of \(p(x)\) to be \[\mathsf{COEF}_{+} (p)= \{i\in J: M_i\neq 0\}.\] To choose the evaluation points in \(\mathbb{F}_q\), we need to look for the \(N\)-th primitive root of unity \(\alpha_N\) that minimizes the set of exponents \(\mathsf{COEF}_{+} (f_A) + \mathsf{COEF}_{+} (f_B) = \mathsf{COEF}_{+} (h)\) under the following conditions:
- \(|\{\alpha_N^i: i\in \mathsf{COEF}_{+} (f_A)\}| = ts + T\)
- \(|\{\alpha_N^i: i\in \mathsf{COEF}_{+} (f_B)\}| = ds + T\)
- \(|\mathcal{I}|=|\{\alpha_N^{(i-1)s + (1-j)(ts+T)}: 1\leq i\leq t, 1\leq j \leq d\}|=td\)
- \(\alpha_N^{z}\notin \mathcal{I}\), for any power \(z\) of polynomial \(h(x)\) associated to coefficients \(A_{i,k_1}B_{k_2,j}\) with \(k_1\neq k_2\), and any coefficient multiple of \(R_k\) or \(S_k\).
If \(s=1\), then a \(N=((d+1)(t+T)-1)\)-th primitive root of unity satisfies the conditions. Otherwise, if \(s>1\), a \(N=(dst+dT+ts-1+T+1)\)-th primitive root of unity will do so. Therefore, Remark [\[rem:fieldsize\]](#rem:fieldsize){reference-type="ref" reference="rem:fieldsize"} establishes the following condition on the size \(q\) of the finite field \(((d+1)(t+T)-1)\mid q\) if \(s=1\), or \((dst+dT+ts-1+T+1)\mid q\), otherwise. **Upload Phase:** The proposed scheme uses \(N\) servers, as determined in the previous item. The user uploads \(f_A(\alpha^i_N)\) and \(f_B(\alpha^i_N)\) to each Server \(i\). **Download Phase:** The \(i\)-th server computes the matrix multiplication \(f_A(\alpha^i_N)\cdot f_B(\alpha^i_N)\) and sends its result back to the user. **User Decoding:** In Lemma [\[lem:decodability\]](#lem:decodability){reference-type="ref" reference="lem:decodability"}, we show that the user is able to retrieve \(\sum_{\ell=1}^sA_{i,\ell} B_{\ell,j}\) from \(\{h(\alpha^i_N): i\in\{1, \dots, N\}\}\). Combining these, the user can decode \[\displaystyle AB =\begin{bmatrix} \sum_{\ell=1}^s A_{i,\ell} B_{\ell,j}\end{bmatrix}_{\substack{1\leq i\leq t\\ 1\leq j\leq d}}.\]
# Proof of Theorem [\[theo:scheme\]](#theo:scheme){reference-type="ref" reference="theo:scheme"}
We break the proof into different Lemmas. We show that the proposed scheme is decodable, in Lemma [\[lem:decodability\]](#lem:decodability){reference-type="ref" reference="lem:decodability"}, \(T\)-secure, in Lemma [\[lem:tsecure\]](#lem:tsecure){reference-type="ref" reference="lem:tsecure"}, and characterize their performance, in Lemma [\[lem:comcosts\]](#lem:comcosts){reference-type="ref" reference="lem:comcosts"}. These statements combined prove Theorem [\[theo:scheme\]](#theo:scheme){reference-type="ref" reference="theo:scheme"}.
Next, we show that the proposed scheme is \(T\)-secure.
We now characterize the total communication rate. | {'timestamp': '2022-06-06T02:14:19', 'yymm': '2206', 'arxiv_id': '2206.01559', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01559'} |
null | null |
# Introduction {#sec:intro}
There is no doubt that modern deep learning techniques have dramatically changed the way engineers approach computer vision problems, replacing hand-engineered image processing algorithms with large neural networks trained end-to-end. While this has improved model performances on a range of tasks, it leaves the underlying algorithms opaque. As deep-learned models become ubiquitous for computer-vision applications, this poses a notable concern, sparking a general interest in neural network *interpretability*.
Work in neural network interpretability can be divided into two broad categories, that of data *representation*, and data *processing*. Much of the interpretability work in computer vision concerns data representation, or *what* the network represents, in the conceptual sense. The focus on representation usually leads to methods for mapping network features back into pixel space, where the researcher can employ their own visual capacities to make inferences about the latent space of the network. Examples of such representational interpretability methods include viewing the output activation maps of a cnn filter, the attention map in a visual transformer, viewing dataset examples or feature visualizations that maximally excite/inhibit a latent feature, or attribution methods that highlight pixel regions most responsible for a model's response. Other work ascribes semantic labels to latent features that align with the annotations in densely segmented image sets. Much less common are accounts of data processing, e.g. an interpretable sequence of computations that generate representations. So, while a researcher might be confident a feature in their model acts as, say, a brown fur detector, this is not necessarily coupled with a formal understanding of how a brown fur detector is built. In general, while it is often possible to trace how the first layer of features are constructed from weights over the input in a deep net, the computation graph becomes increasingly unwieldy when considering features in later layers.
The architectural constraint of convolutional neural networks allows for greater *data processing* interpretability than many other deep models. The kernel convolution operation not only reduces the number of parameters to contend with by replicating weights across space, but also constrains the function space learnable by the network to a familiar subspace. Kernel convolution operations are typical in hand-engineered image processing algorithms, where single kernels can enable filters for edge detection, sharpening, blurring, spatial frequency filtering, etc. Composing filters in a hierarchy enables the construction of more complex detectors, with each kernel acting as a spatial template for the arrangement of features from the previous layer. Cammarata et al (2020) show that viewing feature visualizations in conjunction with the kernels connecting them provides strong visual intuition for how simple features are combined into complex ones. The authors describe many interesting 'circuits' they've discovered in InceptionV1, from early circuits encoding curves and basic shapes, to late-layer circuits encoding real-world objects. A prototypical circuit from their work shows how a car feature is constructed with three kernels spatially arranging features in the previous layer for windows, wheels, and car bodies.
Circuits provide a promising framework for interpreting the image filtering processes latent in CNNs, and this work aims to bolster the approach by addressing two related issues. Firstly, when considering the aforementioned 'car circuit' composed of wheel and window features, the immediate questions arise, 'But what then is the wheel circuit? What is the window circuit? I can't fully grasp the car feature without grasping these constituent parts.' As such, it would be useful to *extend* circuit diagrams backwards layer-wise towards the inputs, as diagrammed in Fig [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}. Such a diagram runs the risk of being massive and unparsable, presenting our second issue; which kernels (edges) do we include in the diagram, and how do we avoid over-simplifying our target feature given our exclusions? This concern with over-simplification is also present in the original circuit work; diagramming the car feature with 3 kernels is certainly parsimonious, but ignoring the effects of the hundreds of other kernels in the original filter may result in a much degraded feature.
In the present work, we address these issues by re-purposing techniques from neural network pruning for *circuit extraction*. We'll compare the efficacy of several saliency-based pruning methods for extracting sparse sub-networks that preserve a target feature's responses to input images *without fine-tuning*. The differences in responses returned by such a circuit formalize the effect of an interpretability researcher *zooming in* (to use the term from ), focusing on some computations while ignoring others, as is a pragmatic necessity. Further, our method allows the researcher to zoom-in in a principled way, by identifying small subsets of computations that approximate the original feature when considered in isolation. In a follow-up set of experiments, we will show how our method can be used to extract *sub-feature* circuits, which preserve feature responses only to a target set of images, dividing complex features into component parts. Lastly, we will explain how sparse circuits can be visualized with our circuit diagramming tool (Fig [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}).
# Methods {#sec:methods}
## Notation {#sec:notation}
This work concerns convolutional layers in deep neural networks. The \(l\)th convolutional layer takes a stack of *activation maps* as input, \(\boldsymbol{A}^{l-1} \in \mathbb{R}^{C_{\text {in}} \times H_{\text {in }} \times W_{\text {in }}}\) and outputs the activation maps \(\mathbf{A}^{l} \in \mathbb{R}^{C_{\text {out}}\times H_{\text {out}} \times W_{\text {out}}}\). The layer is composed of \(C_{\text {out}}\) *filters* (one for each output channel), each parameterized by weights \(\mathbf{w}^{l}_{c_{\text{out}}} \in \mathbb{R}^{C_{\text {in}} \times K_{h} \times K_{w}}\) and bias \(b^{l}_{c_{\text{out}}} \in \mathbb{R}\). Each filter transforms input activation maps to output maps by; \[\label{eq:filter eq} \begin{gathered} \mathbf{A}^{l}_{c_{\text{out}}} = \sum_{c_{\text{in}=1}}^{C_{\text {in}}} \mathbf{w}^{l}_{c_{\text{out}}, c_{\text{in}}} * \boldsymbol{A}^{l-1}_{c_{\text{in}}}+b^{l}_{c_{\text{out}}}, \end{gathered}\] where \[ denotes the convolution. A filter is composed of \(C_{\text{in}}\) 2D *kernels* with weights \(\mathbf{w}^{l}_{c_{\text{out}},c_{\text{in}}} \in \mathbb{R}^{K_{h} \times K_{w}}\) which compute the kernel-wise activation map \(\mathbf{A}^{l}_{c_{\text{out}},c_{\text{in}}} \in \mathbb{R}^{H_{\text {out}}\times W_{\text {out}}}\), before the filter sums these maps element-wise and adds its bias. Individual activations in any activation map are denoted as \(a^{l}_{c_{\text{out}},c_{\text{in}},h,w}\). A *feature* refers to the full function from pixel-space that returns a particular activation, denoted \(\mathcal{F}(x) = a\). An activation map returns the same feature \(H_{\text {out }} \times W_{\text {out }}\) times with respect to the *image patches* in each activation's effective receptive field. For the pruning methods section, \(\boldsymbol{\theta} = \left\{\theta_{j}\right\}_{j=1}^{m}\) refers to the flattened vector of all network parameters.
## Problem Statement {#sec:problem}
We endeavor to simplify the data processing underlying a particular feature by removing parameters from the network, while minimally affecting the feature's activations. We can formulate our *circuit pruning* problem similarly to conventional pruning; given a latent feature \(\mathcal{F}\) in a network and a set of input images (or image patches) \(\mathcal{D}=\left\{\mathbf{x}_{i}\right\}_{i=1}^{n}\), we want to identify a subnetwork parameterized by \(\boldsymbol{\bar{\theta}}\) under a sparsity constraint, such that the change in \(\mathcal{F}\)'s responses to \(\mathcal{D}\) is minimized: \]\label{eq:problem form 1} \begin{gathered} \underset{\boldsymbol{\bar{\theta}}}{\arg \min } \Delta\mathcal{F}(\boldsymbol{\bar{\theta}},\boldsymbol{\theta}; \mathcal{D}):= \sum_{i=1}^{n}\left|\mathcal{F}\left(\boldsymbol{\theta} ;\mathbf{x}_{i}\right)-\mathcal{F}\left(\boldsymbol{\bar{\theta}} ;\mathbf{x}_{i}\right)\right| \\ \text { s.t. } \quad \left\|\boldsymbol{\bar{\theta}}\right\|_{0} \leq \kappa, \quad \bar{\theta}_{j} \in \{\theta_{j},0\} \,\, \forall j \in \{1 \ldots m\} \end{gathered}\[ \(\boldsymbol{\bar{\theta}}\) corresponds to a sparse path of computations through the full network that generates an approximation of the target feature, with no more than \(\kappa\) parameters. The constraint that \(\bar{\theta}_{j} \in \{\theta_{j},0\} \,\, \forall j \in \{1 \ldots m\}\) ensures that we cannot fine-tune the network, that circuits we extract are latent within the network *as is*.
As is typical with pruning problems, finding an optimal \(\boldsymbol{\bar{\theta}}\) cannot be brute forced, as there exist \(m\choose\kappa\) possible parameterizations. Fortunately, many saliency-based approaches to pruning have a similar problem statement to (eq [\[eq:problem form 1\]](#eq:problem form 1){reference-type="ref" reference="eq:problem form 1"}) in that they endeavor to find a sparse \(\boldsymbol{\bar{\theta}}\) that minimally changes the loss function \(\mathcal{L}\) of the original model. Such methods consider quickly-computable, parameter-wise *saliency criteria* \(\boldsymbol{S}(\theta_{j})\) meant to approximate \(|\Delta\mathcal{L}|\) induced by removing individual parameters. With a simple substitution of \(\mathcal{F}\) for \(\mathcal{L}\), these saliency criteria should be well suited for our problem. We can then replace (eq [\[eq:problem form 1\]](#eq:problem form 1){reference-type="ref" reference="eq:problem form 1"}) with the simpler problem of identifying a parameterization \(\boldsymbol{\bar{\theta}}\) with maximal cumulative saliency; \]\label{eq:problem form saliency} \begin{gathered} \underset{\boldsymbol{\bar{\theta}}}{\arg \max } \boldsymbol{S}_{sum}(\boldsymbol{\bar{\theta}})\coloneqq \sum_{j\in supp(\boldsymbol{\bar{\theta}})}\boldsymbol{S}(\theta_{j}) \\ \text { s.t. } \quad \left\|\boldsymbol{\bar{\theta}}\right\|_{0} \leq \kappa, \quad \bar{\theta}_{j} \in \{\theta_{j},0\} \,\, \forall j \in \{1 \ldots m\} \end{gathered}\[ Where solving problem (eq [\[eq:problem form 1\]](#eq:problem form 1){reference-type="ref" reference="eq:problem form 1"}) is hard, solving problem (eq [\[eq:problem form saliency\]](#eq:problem form saliency){reference-type="ref" reference="eq:problem form saliency"}) is easy, just choose paramaters \(\theta_{j}\) with the top-\(\kappa\) \(\boldsymbol{S}(\theta_{j})\) for inclusion in \(\boldsymbol{\bar{\theta}}\).
In this work we test the efficacy of 3 saliency criteria from the literature on our circuit pruning problem, *SNIP*, *FORCE*, and *actgrad*. These are all first order saliency criteria, utilizing gradients. While other methods exist that utilize the Hessian, these are motivated by the observation that training a network with gradient descent should push \(\frac{{\delta\mathcal{L}}}{\delta\boldsymbol{\theta}}\) to 0, but this does not necessarily hold for \(\frac{{\delta\mathcal{F}}}{\delta\boldsymbol{\theta}}\).
## Circuit SNIP {#sec:SNIP}
While the SNIP saliency criterion was originally introduced in, Lee et al (2018) demonstrated the criterion's efficacy for *single-shot* pruning at initialization, which utilizes no fine-tuning, making the criterion promising for our application. In SNIP, a binary mask \(\mathbf{c} \in\{0,1\}^{m}\) is inserted into the network, such that the new network is parameterized by \(\mathbf{c} \odot \boldsymbol{\theta}\), where \(\odot\) denotes the Hadamard product. In the original un-pruned network, \(\mathbf{c} = \{1\}^{m}\). When \(c_{j} = 0\), this is equivalent to parameter \(\theta_{j}\) being pruned from the network. The authors reason the importance of parameter \(\theta_{j}\) to be proportional to the effect on the loss were it to be removed, or equivalently \(c_{j} = 0\): \]\Delta \mathcal{L}_{j}(c_{j}; \mathcal{D})=\mathcal{L}(c_{j} = 1; \mathcal{D})-\mathcal{L}\left(c_{j} = 0; \mathcal{D}\right),\[ The authors' *connection sensitivity* saliency criterion considers the gradient of \(\mathcal{L}\) with respect to \(c_{j}\). \(|\partial \mathcal{L} / \partial c_{j}|\) measures the sensitivity of \(\mathcal{L}\) to perturbations of \(c_{j}\), which they regard as an approximation of the \(|\Delta \mathcal{L}_{j}|\) induced by setting \(c_{j}=0\). Their criterion can be equivalently calculated as the magnitude of parameters multiplied by their gradients: \]\boldsymbol{S}_{snip}(\boldsymbol{\theta}):=\left|\partial \mathcal{L} / \partial \mathbf{c} \right|=\left|\frac{\partial \mathcal{L}(\boldsymbol{\theta} \odot \boldsymbol{c})}{\partial \boldsymbol{c}}\right|_{\mathbf{c}=\mathbf{1}}=\left|\frac{\partial \mathcal{L}}{\partial \boldsymbol{\theta}} \odot \boldsymbol{\theta}\right|\[ We are interested in identifying network parameters that will affect some target feature's average response when removed, rather than the loss. We thus formulate the circuit SNIP saliency criterion with the simple \(\mathcal{F}\) for \(\mathcal{L}\) substitution proposed above: \]\boldsymbol{S}_{snip}(\boldsymbol{\theta};\mathcal{F}):=\left|\partial \mathcal{F} / \partial \mathbf{c} \right|=\left|\frac{\partial \mathcal{F}}{\partial \boldsymbol{\theta}} \odot \boldsymbol{\theta}\right|\[
## Circuit Force {#sec:FORCE}
The above SNIP criterion considers the effects of removing parameters in isolation, where in actuality we are endeavoring to remove many parameters from the network. Removal of a particular \(\theta_{j}\) may have a very different effect when the network is fully-intact versus removal when other parameters are also pruned. Suppose our sparse network is parameterized by \(\boldsymbol{\bar{\theta}}\), then \(\boldsymbol{S}_{snip}(\boldsymbol{\bar{\theta}})\) will certainly be different than \(\boldsymbol{S}_{snip}(\boldsymbol{\theta})\), as the gradients passing through the network will be different. Motivated by this observation, de Jorge et al (2020) (and similarly ) developed the *FORCE* saliency criterion (foresight connection sensitivity), which attempts to identify important parameters for the resultant sparse network, despite the unfortunate circularity that this sparse network is supposed to be identified *by way of* the saliency criterion. The author's make a subtle adjustment to the saliency-based pruning problem statement (eq [\[eq:problem form saliency\]](#eq:problem form saliency){reference-type="ref" reference="eq:problem form saliency"}); \]\label{eq:problem form force} \begin{gathered} \underset{\boldsymbol{\bar{\theta}}}{\arg \max } \boldsymbol{S}_{force\_sum}(\boldsymbol{\bar{\theta}})\coloneqq \sum_{j\in supp(\boldsymbol{\bar{\theta}})}\boldsymbol{S}_{snip}(\bar{\theta_{j}}) \\ \text { s.t. } \quad \left\|\boldsymbol{\bar{\theta}}\right\|_{0} \leq \kappa, \quad \bar{\theta}_{j} \in \{\theta_{j},0\} \,\, \forall j \in \{1 \ldots m\}. \end{gathered}\[ They still endeavor to find a \(\boldsymbol{\bar{\theta}}\) with maximum cumulative saliency, but when parameter-wise saliency is measure in the pruned network, rather than the original network as in (eq [\[eq:problem form saliency\]](#eq:problem form saliency){reference-type="ref" reference="eq:problem form saliency"}).
Unfortunately (eq [\[eq:problem form force\]](#eq:problem form force){reference-type="ref" reference="eq:problem form force"}) is not trivial to solve as is (eq [\[eq:problem form saliency\]](#eq:problem form saliency){reference-type="ref" reference="eq:problem form saliency"}), but the authors reason an approximate solution can be obtained by iteratively masking an increasing number of parameters, recomputing \(\boldsymbol{S}_{snip}(\boldsymbol{\bar{\theta}})\) at each step to get the next mask by way of (eq [\[eq:problem form saliency\]](#eq:problem form saliency){reference-type="ref" reference="eq:problem form saliency"}). Let's call the network parameterization after the first masking iteration \(\boldsymbol{\bar{\theta}}_{1}\), the second \(\boldsymbol{\bar{\theta}}_{2}\), up to the final iteration \(T\) with a parameterization at the desired sparsity \(\boldsymbol{\bar{\theta}}_{T}\). The motivating intuition for this iterative sparsification follows from the observation that the fewer parameters are masked from \(\boldsymbol{\theta}\) to obtain \(\boldsymbol{\bar{\theta}}_{1}\), the better the *gradient approximation* \(\frac{\delta\mathcal{L}(\boldsymbol{\theta})}{\delta\boldsymbol{\theta}} \approx \frac{\delta\mathcal{L}(\boldsymbol{\bar{\theta}}_{1})}{\delta\boldsymbol{\bar{\theta}}}_{1}\) should hold, and we can use \(\boldsymbol{S}_{snip}(\boldsymbol{\theta})\) to approximate \(\boldsymbol{S}_{snip}(\boldsymbol{\bar{\theta}}_{1})\). This logic is inductive; \(\boldsymbol{S}_{snip}(\boldsymbol{\bar{\theta}}_{1})\) is a better approximation of \(\boldsymbol{S}_{snip}(\boldsymbol{\bar{\theta}}_{2})\) than \(\boldsymbol{S}_{snip}(\boldsymbol{\theta})\) (as there are fewer parameters masked between \(\boldsymbol{\bar{\theta}}_{1}\) and \(\boldsymbol{\bar{\theta}}_{2}\)), \(\boldsymbol{S}_{snip}(\boldsymbol{\bar{\theta}}_{3})\) is better approximated by \(\boldsymbol{S}_{snip}(\boldsymbol{\bar{\theta}}_{2})\) than any of the previously computed scores, etc. In general \(S(\boldsymbol{\theta}_{T})\) should be calculated with many iterations of \(\boldsymbol{S}_{snip}\), resulting in a small difference in the masks between iterations, to optimize (eq [\[eq:problem form force\]](#eq:problem form force){reference-type="ref" reference="eq:problem form force"}). In this work we implement *FORCE* with 10 iterations and the exponential decay scheduler from, which were shown to be effective hyper-parameters in the original work (*appendix*). We note that while optimizing (eq [\[eq:problem form force\]](#eq:problem form force){reference-type="ref" reference="eq:problem form force"}) may lead to a *trainable* model parameterization with high connection sensitivity (as was the intention in the original work), we explicitely want to minimize \(\Delta\mathcal{F}(\boldsymbol{\theta},\boldsymbol{\theta}_{T})\). With *FORCE*, at each iteration we are approximating *an approximation*, \(\Delta\mathcal{F}(\boldsymbol{\theta}_{i-1},\boldsymbol{\theta}_{i})\). This could lead to divergence from \(\mathcal{F}\), depending on the accuracy of the iterative approximations.
## Circuits by Taylor Expansion {#sec:taylor}
Another way of deriving an approximation for \(\Delta \mathcal{L}\) given the removal of a parameter is by Taylor approximation. This approach was employed successfully by Molchanov et al (2019) for the pruning of convolutional filters. Their saliency criterion concerns the activations in the network, rather than the weights themselves. Suppose some particular activation \(a_{j}\) in the network were to be masked (set to 0), then the first order Taylor approximation of the loss is: \]\label{eq:taylor 1} \mathcal{L}(a_{i}=0) = \mathcal{L}(a_{i})-\frac{\delta \mathcal{L}}{\delta a_{i}}a_{i}\[ As per usual, the authors wanted their salience criteria for an activation to approximate the magnitude of the change in loss when that activation is masked, which can be approximated with (eq [\[eq:taylor 1\]](#eq:taylor 1){reference-type="ref" reference="eq:taylor 1"}); \]\label{eq:taylor orig} \boldsymbol{S}_{actgrad}(a_{i}):=\left| \mathcal{L}(a_{i})-\mathcal{L}(a_{i}=0) \right| = \left|\frac{\partial \mathcal{L}}{\partial a_{i}} \odot a_{i}\right|\[ As we did for SNIP, we can formulate an equivalent circuit saliency criterion by substituting \(\mathcal{L}\) with \(\mathcal{F}\): \]\label{eq:taylor circuit} \boldsymbol{S}_{actgrad}(\boldsymbol{a};\mathcal{F}):= \left|\frac{\partial \mathcal{F}}{\partial \boldsymbol{a}} \odot \boldsymbol{a}\right|\[ Notice that this criterion is equivalent to \(\mathbf{S}_{snip}\), applied to activations as opposed to weights. Relatedly, the above taylor derivation is applicable to \(\mathbf{S}_{snip}\), and as such we denote this saliency criteria \(\mathbf{S}_{actgrad}\), rather than \(\mathbf{S}_{taylor}\) as in.
## Structured Pruning of Kernels
It is sometimes practical to prune *structurally*, i.e. with respect to architecturally related groups of parameters. In the case of CNNs, this has typical meant pruning entire convolutional filters, as this leads to models that are resource efficient when implemented with BLAS libraries. Our goal is not resource efficiency, but rather to extract interpretable circuits from the model, and as such we will prune with respect to individual convolutional *kernels*. Kernels constitute the edges of our desired circuit diagrams (Fig [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}), thus an extracted circuit with kernel sparsity should make for a maximally parsable diagram. To compute *kernel-wise* saliency scores, we simply average the saliency scores of the parameters within each kernel. For \(\boldsymbol{S}_{snip}\) (and by extension \(\boldsymbol{S}_{force}\)), this is the average saliency of each of the kernel's weights. For \(\boldsymbol{S}_{actgrad}\), we average activations across the kernel-wise activation map. For a kernel \(k\) with weights \(\mathbf{w}\) and kernel-wise activation map \(\boldsymbol{A}\); \]\label{eq:kernel snip} \boldsymbol{S}_{snip}(k) \coloneqq \frac{1}{K_{w}K_{h}}\sum_{w=1}^{K_{w}}\sum_{h=1}^{K_{h}}\boldsymbol{S}_{snip}({w_{h,w}})\[ \]\label{eq:kernel actgrad} \boldsymbol{S}_{actgrad}(k) \coloneqq \frac{1}{H_{\text{out}}W_{\text{out}}}\sum_{w=1}^{W_{\text{out}}}\sum_{h=1}^{H_{\text{out}}}\boldsymbol{S}_{actgrad}({a_{h,w}})\[ Pruning kernels to sparsity \(\kappa\) requires sorting the kernel-wise saliency scores and keeping the top-\(\kappa\) kernels, as before.
# Experiments {#sec:experiments}
All of the following experiments were conducted with a single NVIDIA *GeForce RTX 2080 Ti* GPU. All feature visualizations in this work were generated using the *Lucent* library (under an Apache License 2.0), a PyTorch implementation of *Lucid*, using the default hyper-parameters.
## Methods Comparison {#sec:methodcompare}
To test our circuit pruning methods for feature preservation, we pruned circuits for features in an Imagenet trained Alexnet. We tested two versions of this model, one trained in the typical manner, as is available through the Pytorch modelzoo, and one we trained with hierarchical group sparsity regularization. This regularization encourages groups of parameters (kernels and filters) towards zero magnitude during training *(appendix)*. If we endeavor to find sparse circuits within a model, we reason it is useful to start with a model that already sparsely achieves the objective. We selected 20 random features from the last 3 of the model's convolutional layers on which to test our methods. For each feature, we average the activations it returns to a subset of 2000 random images from Imagenet (2 images per category). This is the value for \(\mathcal{F}\) we use to compute kernel-wise saliency scores for each pruning method. As a baseline method, we additionally compute a *magnitude* kernel-wise saliency score, equal to the average magnitude of each weight in the kernel; \]\label{eq:magnitude} \boldsymbol{S}_{magnitude}(k) \coloneqq \frac{1}{K_{w}K_{h}}\sum_{w=1}^{K_{w}}\sum_{h=1}^{K_{h}}|w_{h,w}|\[ Using these saliency scores, we then extracted circuits at 13 sparsities, preserving 99%-.1% of the *relevant* kernels with highest saliency. By *relevant*, we mean kernels that could possibly influence the target feature's activations. These include all the kernels in the preceding layers and those kernels that belong to the feature's filter itself, but no other kernels from that filter's layer or later layers. For each of these extracted circuits, we computed the absolute value of the Pearson correlation between the original, un-pruned features' activations to the 2000 images and the corresponding circuit features' activations. This metric measures the linearity of the relationship between the circuit feature and the original feature.
Fig [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"} shows how well feature activations were preserved (\|Pearson's R\|), as a function of circuit sparsity, with separate lines for the three pruning methods, as well as the magnitude baseline method. The results show that *actgrad* is generally most effective for the Alexnet model, while *actgrad*, *SNIP* and *FORCE* perform similarly on the regularized Alexnet. *FORCE* shows a slight improvement with respect to the other metrics at the highest sparsities measured (though this effect is minimal, with a higher correlation of only \(\Delta R \approx.08\). The results in showed *FORCE* is most effective in the high sparsity regime, and our results converge with this finding. Overall, all methods are able to extract sparse feature-preserving circuits well, outperforming the baseline magnitude saliency criterion.
These circuit extraction experiments also show that features in Alexnet are not homogenously sparse. Some features can maintain a high correlation with the full-model feature counterpart, even at extreme sparsities, while other features degrade quickly as more kernals are pruned. To get the full picture of the distribution of feature sparsities, we moved beyond the 60 features tested for methods comparison and next extracted circuits for every feature in the last 3 convolutional layers of Alexnet (896 features), using the *actgrad* method. We did not test all these features in the previous methods comparison experiment due to the long compute times for the *FORCE* method. Fig [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"} shows these distributions of correlations for both models. Nearly all features tested are preserved (Pearson's R \>.99) at sparsities up to 50% in the regularized model. At higher sparsities (\(\geq 20\%\)) there is increasingly high variance in the correlations across features, but individual features can be identified that are well preserved at even the highest sparsity tested (.1% of relevant kernels kept).
## Sub-Feature Circuits {#sec:subfeature}
The circuits we've extracted so far attempt to capture features as a whole, maintaining their responses to a broad sample of image patch inputs, but it may be useful for an interpretability researcher to extract *sub-feature* circuits, which preserve responses to a targeted selection of image patches. We are motivated here by the observation of *poly-semantic* features in CNNs, which respond selectively and highly to seemingly disparate semantic categories. Poly-semantic features may be composed of multiple, separable circuits with their own constituent kernels, with each circuit responsible for detecting only one of the semantic categories. The *general* circuit for a poly-semantic feature would then be the union of these category-wise circuits. In this case, an interpretability researcher should study these sub-feature circuits independently, as they constitute distinct image filtering processes. These sub-feature circuits would have the additional benefit of being ostensibly sparser than the general circuit for a feature.
To identify candidate poly-semantic features for our *sub-feature* extraction experiments, we took a data-driven approach. For each feature \(\mathcal{F}^{l}_{j}\) in the regularized Alexnet convolutional layers, we identified the top 300 individual activations it produced in response to the ImageNet training set, under the constraint that no two activations belong to the same activation map (are responses to different parts of the same image). Let \(\mathbf{a}^{l}_{j} = \{a_{j,i}\}_{i=1}^{300}\) refer to these activations, and \(\mathcal{D}^{l}_{j} = \{\mathbf{x}^{l}_{j,i}\}_{i=1}^{300}\) refer to the corresponding image patches in each activation's effective receptive field. We then consider layer \(l\)'s *population* response to \(\mathcal{D}^{l}_{j}\), a set of 300 *activation vectors* \(\mathcal{P}^{l}_{j}\): \]\label{eq:population vecs} \mathcal{P}^{l}_{j} := \{p^{l}_{j,i}\}_{i=1}^{300} := \{\{\mathcal{F}_{h}^{l}(\mathbf{x}^{l}_{j,i})\}_{h=1}^{C_{\text{out}}}\}_{i=1}^{300}\[ We reason that for a candidate poly-semantic feature \(\mathcal{F}_{j}^{l}\), the set \(\mathcal{P}^{l}_{j}\) should be clusterable. This would suggest \(\mathcal{D}^{l}_{j}\) is represented as distinct groups in layer \(l\)'s representation space, while the supposed poly-semantic feature \(\mathcal{F}_{j}^{l}\) makes no such distinction, as its activations \(\mathbf{a}^{l}_{j}\) are all very high, given our selection criteria for \(\mathbf{a}^{l}_{j}\).
We used HDBScan clustering to search for features with clusterable \(\mathcal{P}^{l}_{j}\) (under a BSD 3-Clause license). HDBScan works well in high-dimensions and requires only a minimum cluster size as its hyperparameter (we used 10). It will return anywhere from 0 to 30 clusters (300/10), allowing us to ignore filters that don't show good clustering, and dissect other filters with respect to potentially many semantic groups (although in practice this approach rarely returned more than 2 clusters) (*appendix*).
Fig [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"} shows the results of a pruning experiment on feature *Conv5:9* in regularized Alexnet. Our clustering analysis revealed this as a candidate poly-semantic filter, returning two distinct clusters, \(\mathcal{P}_{1}\) and \(\mathcal{P}_{2}\). Qualitatively, the image patches corresponding to \(\mathcal{P}_{1}\) and \(\mathcal{P}_{2}\) (\(\mathcal{D}_{1}\) and \(\mathcal{D}_{2}\)) look distinct, with one cluster corresponding to terrier faces, and the other corresponding to metal cylinders with horizontal lines, as can be seen in the sample of image patches in the top right of Fig [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}. Next, we extracted different circuits for *Conv5:9*, which were intended to preserve *Conv5:9*'s responses to either \(\mathcal{D}_{1}\) or \(\mathcal{D}_{2}\), using saliency scores \(\boldsymbol{S}_{actgrad}(\mathbf{k};\mathcal{F};\mathcal{D}_{1})\) and \(\boldsymbol{S}_{actgrad}(\mathbf{k};\mathcal{F};\mathcal{D}_{2})\). Rather than prune circuits from the full model, we start pruning from the *general circuit*, extracted in the previous experiment (sec [3.1](#sec:methodcompare){reference-type="ref" reference="sec:methodcompare"}), at 50% sparsity. This general circuit shows a very high correlation to the original filter's activations (\>.999 Pearson R). We extract the sub-filter circuits at 70 linearly spaced sparsities, preserving 49.99%-.5% of the relevant kernels in the original model. The line plot in Fig [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"} shows how these sub-filter circuits' activations deviate from the general circuit's activations as a function of sparsity. The Y-axis shows a normalized measure of this deviation. For a sub-feature circuit parameterized by \(\boldsymbol{\bar{\theta}}\) and the general circuit by \(\boldsymbol{\theta}_{gen}\); \]\label{eq:F norm} \Delta\mathcal{F}_{norm}(\boldsymbol{\theta}_{gen},\boldsymbol{\bar{\theta}};\mathcal{D}) := \frac{\Delta\mathcal{F}(\boldsymbol{\theta}_{gen},\boldsymbol{\bar{\theta}};\mathcal{D})}{\frac{1}{n}\sum_{i=1}^{n}\mathcal{F}(\boldsymbol{\theta}_{gen};\mathbf{x}_{i})}\(\) With this metric, we can see how the sub-circuit's activations deviate from the general circuit's activations as a proportion of the general circuit's activation magnitude. The solid lines show how the activations the circuit was meant to preserve deviate from the respective general circuit activations. Overall, we were able to extract sub-feature preserving circuits to high levels of sparsity.
Critically, we next examined how well these sub-feature circuits responded to the other cluster's image patches (dotted lines). If the circuits act as distinct feature detectors, these responses should be far less preserved. The dotted lines in this plot clearly deviate at lower sparsities than the solid lines; we've identified circuits that preserve *Conv5:9* responses to only terriers but not metal cylinders, and visa-versa. To quantify the extent to which these specific sub-feature circuits involve different computations, we computed the intersection over union (IoU) of the sets of kernels belonging to each circuit. We compare the circuits at the last sparsity for which our \(\Delta\mathcal{F}_{norm}\) remains below.15 for each circuits' target activations. This intersection over union is shown as a function of layer depth in the bottom-right of Fig [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}. It seems the two sub-feature circuits utilize the same layer 1 computations (gabors, center-surround filters etc.) but deviate from one another in subsequent layers. We also render feature visualizations for these extracted circuits, shown above the IoU plot. These show a clear distinction between the sub-feature circuits as well. Fig [\[fig:6\]](#fig:6){reference-type="ref" reference="fig:6"} shows the same analyses again, but for feature conv5:161, which is poly-semantic across monkey faces and written text.
As a control analysis, we extracted two new sub-feature circuits, this time with respect to a random 50/50 split of image patches from \(\mathcal{D}_{1}\) and \(\mathcal{D}_{2}\). We hypothesized that these sub-feature circuits would not be separable. The results are shown in Fig [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"} and confirm our hypothesis, indicated by the overlap between the solid and dashed lines, and the high IoU plot across the layers. Thus, for this feature, while separable sub-feature circuits exist for processing terrier faces and metal cylinders, they do not exist for any arbitrary set of images.
Finally, to test whether separable sub-feature circuits could be extracted without utilizing our clustering procedure, we attempted to *split* feature *Conv4:53* in a hypothesis-driven way. The top 300 highest activating image patches for *Conv4:53* show cocentric circles (*appendix*), and we hypothesized there may be separable circuits responsible for the feature's response to large circles versus small circles, resulting in a scale-invariant circle detector. We attempt to isolate these circuits by extracting with respect to the two simple images in the top right of Fig [\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}, fit to the effective receptive field of *Conv4:53*. The pattern of data in Fig [\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"} reveals that this feature can indeed be split into separable circuits with respect to these images.
# Extended Circuit Diagrams {#sec:diagram}
Circuit pruning allows us to extract sparse sub-networks from a CNN that preserve the responses of targeted features. This reduces the size of features' computation graphs, and allows for detailed inspection of how rich features are built. To facilitate this inspection, we have developed a tool for *extended circuit diagrams*. Fig [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} shows an extended circuit diagram for *conv3:56*, a feature that can be preserved at high sparsities according to our pruning experiments (sec [3.1](#sec:methodcompare){reference-type="ref" reference="sec:methodcompare"}). Following Olah et al (2020), edges in this diagram represent convolution with a 2D kernel, and vertices represent filters. Edge lines are colored based on the inhibition/excitation of their corresponding kernel (red for excitation, blue for inhibition, and gray for a mixture). The width of each edge line is proportional to the *actgrad* saliency score it received when pruning the circuit. Feature visualizations are displayed over their corresponding vertex, rendered with respect to the circuit, *not the original model* (unlike ). This ensures the visualization is representative of the filtering operations displayed in the diagram. Additionally, our sparse circuit extraction allows for rendering of the *entire* circuit graph, with kernels/edges all the way back to the first convolutional layer (also unlike ). We visualize the first layer's features directly, by rendering the weights of each filter's 3 kernels as the visualization's 3 color channels. Regarding Fig [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}, where feature *conv3:56* had 12864 kernels preceding kernels in the original model, we've extracted a feature approximating circuit with only 210 kernels. Even at this high sparsity, Fig [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} is still a complex graph that's difficult to parse as a static image. To remedy this, our full circuit diagramming tool comes equipped with a point-click GUI that allows the user to isolate individual vertices in the graph, with their incoming and outgoing edges. (*appendix*).
# Conclusion & Future Work {#sec:conclusion}
In this work, we repurpose 3 saliency-based pruning methods for the extraction of feature preserving *circuits*. We compare the efficacy of these methods, finding they are similarly effective, and better than magnitude-based pruning. We then demonstrate how sparse circuits can be extracted that preserve features' responses to only a selection of images. These extracted circuits constitute sparse image-filtering algorithms, which can be analyzed with our circuit diagramming tool. We've limited the current work to analyses of Alexnet, as our circuit diagramming tool does not currently support architectures with branching or skip connections. In the future we will extend this work to other CNN models. Additionally, our current experiments in 'sub-filter' pruning are largely hypothesis-driven, in the future we would like to develop data-driven methods for identifying separable sub-filter circuits in a network. | {'timestamp': '2022-06-06T02:16:16', 'yymm': '2206', 'arxiv_id': '2206.01627', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01627'} |
# Encoders as projectors
A good *learned representation* has many desiderata. The perhaps most elementary constraint placed over most learned representations is that a given observation \(\vec{x}\) should have a *unique* representation \(\vec{z}\), at least in distribution. In practice this is ensured by letting the representation be given by the output of a function, \(\vec{z} = g(\vec{x})\), often represented with a neural network. Having a unique representation for a given observation is important for downstream tasks, as otherwise the downstream loss function is complicated by having to be defined on sets. Uniqueness of representation is also important when a human investigator seeks information about the phenomenon underlying data, e.g.through visualizations, as uniqueness is a ubiquitous assumption. The *autoencoder* is an example where uniqueness of representation is explicitly enforced, even if its basic construction does not suggest unique representations. In the most elementary form, the autoencoder consists of an *encoder* \(g_{\psi}: \R^D \rightarrow \R^d\) and a *decoder* \(f_{\phi}: \R^d \rightarrow \R^D\), parametrized by \(\psi\) and \(\phi\), respectively. These are trained by minimizing the *reconstruction error* of the training data \(\{ \vec{x}_1, \ldots, \vec{x}_N \}\), \[\begin{aligned}
\psi^*, \phi^* = \argmin_{\psi, \phi} \sum_{n=1}^N \| f(g(\vec{x}_n))-\vec{x}_n \|^2. \end{aligned}\] Here \(d\) is practically always smaller than \(D\), such that the output of the encoder is a low-dimensional latent representation of high-dimensional data. The data is assumed to lie near a \(d\)-dimensional manifold \(\M\) spanned by the decoder. For a given decoder, we see that the optimal choice of encoder is the projection onto \(\M\), i.e. \[\begin{aligned}
g_{\text{optimal}}(\vec{x}) &= \proj_{\M}(\vec{x}) = \argmin_{\vec{z}} \| \vec{x}-f(\vec{z}) \|^2. \end{aligned}\] For any *nonlinear* choice of decoder \(f\), this optimal encoder does *not* exists everywhere. That is, multiple best choices of latent representation may exist for a given point, as the projection is not unique everywhere. As the learned encoder enforces a unique representation, it will choose arbitrarily among the potential representations. In this case, any analysis of the latent representations can be misleading, as it does not contain the information that another choice of representation would be equally as good. **In this paper** we investigate the *reach* of the manifold \(\M\) spanned by the decoder \(f\). This concept, predominantly studied in geometric measure theory, informs us about regions of observation space where the projection onto \(\M\) is unique, such that trustworthy unique representations exist. If training data resides inside this region we may have hope that a suitable encoder can be estimated, leading to trustworthy representations. The classic *reach* construction is global in nature, so we develop a local generalization that gives a more fine-grained estimate of the uniqueness of a specific representation. We provide a new local, numerical, estimator of this reach, which allows us to determine which observations can be expected to have unique representations, thereby allowing investigations of the latent space to disregard observations with non-unique representations. Empirically we find that in large autoencoders, practically all data is outside the reach and risk not having a unique representation. To counter this, we design a reach-based regularizer that penalizes decoders for which unique representations of given data do not exist. Empirically, this significantly improves the guaranteed uniqueness of representations with only a small penalty in reconstruction error.
# Reach and uniqueness of representation
Our starting question is *which observations \(\vec{x}\) have a unique representation \(\vec{z}\) for a given decoder \(f\)?* To answer this, we first introduce the *reach* of the manifold spanned by decoder \(f\). This is a *global* scalar that quantifies how much points can deviate from the manifold while having a unique projection. Secondly, we contribute a generalization of this classic geometric construct to characterize the local uniqueness properties of the learned representation.
## Defining reach
The nearest point projection \(\proj_{\M}\) is a well-defined function on all points for which there exists a unique nearest point. We denote this set \[\begin{aligned}
\Unp(\M) = \{\vec{x}\in \R^D: \vec{x}\text{ has a unique nearest point in }\M\}, \end{aligned}\] where \(\M = f(\R^d)\) is the manifold spanned by mapping the entire latent space through the decoder. Observations that lie within \(\Unp(\M)\) are certain to have a unique optimal representation, but there is no guarantee that the encoder will recover this. With the objective of characterizing the uniqueness of representation, the set \(\Unp(\M)\) is a good starting point as here the encoder at least has a chance of finding a representation that is similar to that of a projection. However, for an arbitrary manifold \(\M\) it is generally not possible to explicitly find the set \(\Unp(\M)\). Introduced by, the *reach* of \(\M\) provides us with an implicit way to understand which points are in and outside \(\Unp(\M)\).
Hence, \(\reach(\M)\) is the greatest radius \(r\) such that any open \(r\)-ball centered on the manifold lies in \(\Unp(\M)\). In the existing literature, the *global reach* is referred to as the *reach*; we emphasize the global nature of this quantity as we will later develop local counterparts. Definition [\[def:reach\]](#def:reach){reference-type="ref" reference="def:reach"} does not immediately lend itself to computation. Fortunately, provides a step in this direction, through the following result. | {'timestamp': '2022-06-06T02:13:41', 'yymm': '2206', 'arxiv_id': '2206.01552', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01552'} |
null | null |
# Introduction {#sec:Introduction}
Machine maintenance constitutes an intricate operational problem. The challenge is to avoid machine failures and costly overhauls, while simultaneously minimizing the cost of preventive maintenance (PM). Moreover, maintenance is often imperfect in practice since it does not restore the machine to a state as good as new. In fact, a broad spectrum of maintenance effects have been studied in the literature, ranging from perfect maintenance, which restores the system to a state as good as new, to worst maintenance, where maintenance causes the machine to fail. Existing approaches in imperfect maintenance rely on strong assumptions regarding the effect of PM. First, the effect is modelled as either deterministic or stochastic assuming a certain probability distribution. These assumed effects, however, might not correspond to the actual effect. Second, the effect is typically assumed to be machine-independent, i.e., identical for all machines. In reality, the effect of the same type of PM intervention could be very different for different machines. For example, changing a gear would likely have a different impact on a brand new machine compared to the exact same maintenance intervention on an old, worn down machine. This work relaxes both assumptions by proposing a completely data-driven maintenance policy that learns the effect of maintenance conditional on a machine's characteristics. The benefit of this approach is that it allows (1) to flexibly learn the maintenance effects from observational data (instead of assuming a certain deterministic or stochastic effect based on expertise), and (2) to design a machine-specific PM schedule based on these learned effects. These benefits are achieved by framing maintenance as a problem of causal inference. We argue that the challenge in maintenance is that, for each specific machine, we only observe one outcome for the maintenance frequency that was administered in practice. We never observe the counterfactual outcomes--what would have happened if that machine received more or less maintenance in the past. Therefore, we never know whether the optimal maintenance frequency was prescribed. This is exactly the aim of causal inference, i.e. to predict each individual machine's potential outcomes in terms of failures and overhauls for different levels of PM. By learning a model that predicts the number of overhauls and failures given the PM frequency, we can optimize the PM schedule to minimize the total estimated cost. Essentially, we propose using observational data to learn a machine-specific digital twin for maintenance that predicts what would happen if a machine is prescribed a certain maintenance schedule. This work contributes by proposing a novel prescriptive framework for maintenance that prescribes maintenance based on the estimated effect of PM on the machine's number of overhauls and failures. To this aim, we frame maintenance as a problem of causal inference. Consequently, we leverage state-of-the-art machine learning methods for causal inference that learn models to estimate a machine's potential outcomes for different PM frequencies from observational data. Moreover, we formulate a prescriptive policy that uses the potential outcomes to decide on the optimal PM frequency so as to minimize the total cost of failures and interventions. Empirically, we contribute by demonstrating the excellent use of the presented prescriptive framework on a dataset consisting of more than 4,000 maintenance contracts of industrial equipment provided by an industrial partner.
# Related work {#sec:Related_work}
Machine maintenance has been studied extensively in operations research, with a wide variety of proposed maintenance policies. Although most existing work assumes that maintenance restores the system to a state that is as good as new, maintenance is typically imperfect in reality. In fact, different maintenance effects have been studied in the literature, ranging from maintenance that restores the system to a perfect state to maintenance that makes the system's state worse. Consequently, developing maintenance policies that incorporate imperfect maintenance is an important research problem.
## Imperfect maintenance
Existing work models the effect of imperfect maintenance as either stochastic (based on a known probability distribution) or deterministic. Stochastic effects include the \((p,q)\) rule, where maintenance is perfect with probability \(p\) and minimal with probability \(q=1-p\), as well as its age-dependent variant \((p(t), q(t))\). Other work assumes a deterministic effect. Improvement factor models assume that maintenance decreases the system's failure rate by a deterministic improvement factor . Similarly, in virtual age models, imperfect maintenance decreases the system's age or its failure intensity with a deterministic factor \(q\) where \(0 < q < 1\). Conversely, instead of making assumptions regarding the effect of maintenance, our work proposes to learn the effect of PM from data. Data-driven approaches have recently gained importance in the maintenance literature. Condition-based maintenance is a recent paradigm where maintenance is optimized based on the machine's state or its characteristics. Especially relevant to our work are recent, predictive maintenance approaches that learn a predictive model from data to decide on the appropriate maintenance interventions. Similar to the general literature on imperfect maintenance, existing condition-based approaches that do consider imperfect maintenance assume either a deterministic or stochastic maintenance effect. There exist three broad categories of condition-based approaches that account for imperfect maintenance. A first category considers minimal maintenance with a *deterministic* effect, in which a system has several deterioration stages and imperfect maintenance returns the system to the previous stage. A second category considers *stochastic* effects where the maintenance effect is governed by an assumed probability distribution. Finally, in *improvement factor* models, imperfect maintenance decreases the system's hazard rate with a (deterministic) factor between zero and one. To the best of our knowledge, no existing condition-based approaches aim to *learn* the effect of maintenance from data. Finally, this work focuses on a provider of full-service contracts
## Prescriptive analytics and causal inference
Instead of assuming a certain PM effect, this work uses data-driven models to learn the effect of maintenance using techniques from causal inference. Causal inference aims to estimate the effect of a certain cause from data, e.g., the number of failures resulting from a certain maintenance frequency compared to not applying any maintenance. Ideally, estimating maintenance effects would be done by conducting a randomized controlled trial: assigning different levels of maintenance to a collection of (similar) machines and comparing the outcomes. However, in practice, this approach can be prohibitively expensive or even unfeasible. In maintenance specifically, it would be challenging to randomly assign various levels of PM to different machines. Therefore, we need to rely on historical, observational data of machines and their maintenance. The challenge of working with observational data is that this data is biased due to existing maintenance policies that were applied. For example, as a result of an existing policy, machines more prone to failure might have been more likely to receive maintenance in practice. This phenomenon, called selection bias or confounding bias, can result in biased estimates of the counterfactual outcomes if ignored. Therefore, specialized tools have been developed in the causal inference literature to tackle exactly this problem and learn causal effects from observational data, i.e., in the presence of selection bias. Specifically, our work is related to learning potential outcomes for continuous-valued interventions, e.g., the number of PM interventions per running period. Causal inference has been applied to a variety of applications, such as personalized medicine, economic policy design, or marketing. Moreover, it is related to prescriptive analytics, which has recently gained importance in operations research. In this work, causal inference is used to predict a machine's potential outcomes for different levels of maintenance and decide upon a personalized maintenance schedule. To the best of our knowledge, this is the first application of causal inference for maintenance optimization.
# Problem overview {#sec:Problem_overview}
This work aims to solve the problem faced by a provider of full-service maintenance contracts. The service provider is responsible for maintaining the client's asset at a predetermined price. Therefore, for each contract, the service provider needs to decide on a usage-based PM schedule, prior to contract start, based on information such as the type of machine it concerns and/or the machine's age at contract start. In this work, we assume the service provider conducts a single type of PM intervention and needs to decide on the frequency of these interventions. Planned PM aims to prevent two types of events. The first, overhauls, are unplanned, comprehensive maintenance interventions during which large parts of the machinery need to be replaced. From the viewpoint of the full-service maintenance provider, these are the most costly type of event. The second, machine failures, result in an urgent need for maintenance as the machine stops running until corrective maintenance occurs. This again incurs a cost to the service provider that is smaller than the cost of an overhaul, but larger than the cost of PM. The overall goal is to find each contract's optimal PM frequency that minimizes the combined cost of planned PM, overhauls and failures, from the perspective of the service provider. Although planning more PM interventions is likely to result in less overhauls and failures, it also comes at an increased maintenance cost. This means the PM frequency is a trade-off between costs resulting from planned PM on the one hand and costs resulting from overhauls and failures on the other hand. Due to heterogeneity in the contracts, maintenance might need to be planned more frequently for some machines. Therefore, it is important to consider the contract's characteristics when deciding on the PM frequency. To this aim, the service provider has access to information on past contracts including how often maintenance was applied as well as the number of overhauls and failures that were observed. Formally, each contract is defined as a tuple \((\mathbf{X}, T, O, F)\). \(\mathbf{X} \in \mathcal{X} \subset \mathbb{R}^d\) denotes the characteristics of the machine and contract. The treatment, the PM frequency or the number of preventive maintenance interventions that will be applied per running period, is denoted as \(T \in \mathcal{T} \subset \mathbb{R}^+\). \(O \in \mathcal{O} \subset \mathbb{R}^+\) and \(F \in \mathcal{F} \subset \mathbb{R}^+\) are the contract's number of overhauls and failures per running period. We adopt the Rubin--Neyman potential outcomes framework and denote the overhauls \(O\) and failures \(F\) per running period given maintenance frequency \(t\) as \(O(t)\) and \(F(t)\). The objective is to decide on the optimal maintenance frequency \(t_i^*\) that minimizes the total cost per running period. We assume a usage-based maintenance cost similar to. A machine \(i\)'s cost per running period given PM frequency consists of the combined costs of PM, overhauls and failures: \[\label{eq:total_cost} c_i(t_i) = \underbrace{c_t \, t_i \, \vphantom{c_f}}_{\text{PM}} + \underbrace{c_o \, o_i \, \vphantom{c_f}}_{\text{Overhauls}} + \underbrace{c_f \, f_i}_{\text{Failures}}.\] Here, we assume that the individual costs of preventive maintenance, overhauls and failures (\(c_t, c_o, c_f \in \mathbb{R}^+\)) are deterministic and known. To assist the full service-provider's decision-making, a data set is available with information on \(n\) past contracts \(\mathcal{D} = \left\{(\mathbf{x}_i, t_i, o_i, f_i)\right\}^{n}_{i=1}\). For each of the past contracts, only one potential outcome was observed for \(O\) and \(F\): \(o_i(t)\) and \(f_i(t)\). The other, counterfactual outcomes are never observed. This is known as the fundamental problem of causal inference. The challenge in causal inference is to predict, for a new contract, all potential outcomes by learning from this historical data. Because past decisions regarding the PM frequency were made according to an unknown existing policy, there is selection bias in the data. This means that contracts that were likely to receive relatively little PM are different from machines that were likely to receive relatively much PM. For example, the service provider might have known from experience that a certain type of machine would be likely to fail often when not receiving frequent PM and, because of this, prescribed more maintenance to those machines in the past. Therefore, learning a predictive model for estimating potential outcomes from observational data needs to adjust for selection bias in this data to obtain unbiased estimates.
# Methodology {#sec:Methodology}
Our methodology consists of a predict-then-optimize framework, see [\[fig:methodology_overview\]](#fig:methodology_overview){reference-type="ref" reference="fig:methodology_overview"} for a high-level overview. To estimate each contract's cost for a certain PM frequency \(c_i(t_i)\), each machine's potential outcomes need to be estimated, i.e., its number of overhauls \(o_i(t)\) and failures \(f_i(t)\) for a PM frequency \(t_i\), given its characteristics \(\mathbf{x}_i\). Therefore, the first step is to learn a machine learning model for estimating potential outcomes from historical, observational data on similar full-service contracts \(\mathcal{D}\). In a second phase, these estimated outcomes can be used to optimize the PM frequency and resulting total cost. The rest of this section is organized as follows. First, estimating potential outcomes from observational data requires two standard assumptions. These are put forward in [4.1](#ssec:assumptions){reference-type="ref" reference="ssec:assumptions"}. Second, we estimate the potential outcomes by learning a predictive model from observational data. For this, we use a state-of-the-art methodology called SCIGAN, which is described in [4.2](#ssec:predict){reference-type="ref" reference="ssec:predict"}. Third, in [4.3](#ssec:optimize){reference-type="ref" reference="ssec:optimize"}, these predictions are used to assign each machine's optimal PM frequency that minimizes the total estimated cost.
## Assumptions {#ssec:assumptions}
The challenge in estimating potential outcomes from observational data is dealing with selection bias. Learning unbiased estimates of the potential outcomes from observational data requires making three standard assumptions: consistency, overlap and unconfoundedness. Given these assumptions, adjusting for machine characteristics \(\mathbf{x}_i\) allows to account for selection bias in observational data and obtain unbiased estimates. The first assumption is consistency, i.e., a machine's potential outcome given observed treatment \(t\) is the observed outcome.
The second, overlap or positivity, ensures that each possible contract \(\mathbf{x}_i\) has a non-zero probability of receiving each frequency of PM interventions \(t_i\).
The third, unconfoundedness or no hidden confounders, ensures that there are no unobserved variables influencing both the treatment assignment \(T\) and a potential outcome \(O(t)\) or \(F(t)\).
## Predicting preventive maintenance effects {#ssec:predict}
First, we need to predict each machine's potential outcomes \(o_i(t)\) and \(f_i(t)\) given a PM frequency \(t_i\) based on characteristics \(\mathbf{x}_i\). Therefore, we aim to find models \(g_o: \mathcal{X}\times\mathcal{T} \to \mathcal{O}\) and \(g_f: \mathcal{X}\times\mathcal{T} \to \mathcal{F}\) defined by parameters \(\theta_o, \theta_g \in \Theta\) and obtain unbiased estimates of the potential outcomes \(g_o(t, \mathbf{x}) = \mathbb{E}\left[O(t) \; | \; \mathbf{X} = \mathbf{x}\right] \; \text{ and } \; g_f(t, \mathbf{x}) = \mathbb{E}\left[F(t) \; | \; \mathbf{X} = \mathbf{x}\right]\). In this work, \(g_o\) and \(g_f\) are learned using SCIGAN, a recently proposed machine learning approach for predicting potential outcomes given a continuously-valued treatments. SCIGAN achieved state-of-the-art performance across a variety of settings. \(g\) is learned in two steps. First, a generative adversarial network (GAN) is trained to model the distribution of the potential outcomes: the generator is trained to generate counterfactual contracts that cannot be distinguished from factual, observed contracts by the discriminator. In a second phase, the GAN is used to augment the observed training data with generated counterfactual samples. This way, the augmented data set contains all potential outcomes, including both the factual outcomes and the generated, counterfactual outcomes. Because of this, selection bias is no longer a problem and, using this augmented data set, a predictive model \(g_\theta\) can be trained to predict the potential outcomes in a supervised manner. For this, we use a neural network. More specificallly, we use a multilayer perceptron (MLP).
## Optimizing the maintenance cost {#ssec:optimize}
The optimal PM frequency is a trade-off between costs resulting from planned PM on the one hand and costs resulting from overhauls and failures on the other hand. However, using the potential outcomes \(o_i(t_i)\) and \(f_i(t_i)\), it can be seen that the overhauls and failures can be written as functions of the PM frequency \(t_i\). Therefore, the predicted potential outcomes can be used to directly estimate the costs incurred at different PM frequencies. This is achieved by rewriting all terms in [\[eq:total_cost\]](#eq:total_cost){reference-type="ref" reference="eq:total_cost"} (PM, overhauls and failures) as a function of the PM frequency \(t_i\): \[\label{eq:total_cost_t} c_i(t_i) = c_t \, t_i + c_o \, o_i(t_i) + c_f \, f_i(t_i).\] Each machine's optimal PM frequency \(t_i^*\) is found as the level that minimizes the expected cost: \(t_i^* = \argmin c_i(t)\). To account for heterogeneity in the contracts, this optimal PM frequency is optimized for each specific machine.
# Results {#sec:Results}
We validate our methodology empirically using real-world data on full-service maintenance contracts. The goal is to decide on the optimal PM frequency, prior to the contract start, to minimize the total cost resulting from preventive maintenance, overhauls and failures.
## Data {#ssec:Data}
Our data set contains more than 4,000 full-service maintenance contracts. For each contract \(i\), we have information \(\mathbf{x}_i\) on the machine, the contract itself, and maintenance-related events (see [\[tab:data_overview\]](#tab:data_overview){reference-type="ref" reference="tab:data_overview"}). Events are presented per running period, which is a set number of running hours. For reasons of confidentiality, the exact number of running hours per period is not presented. Costs are averaged over all events and re-scaled for reasons of confidentiality. The data is preprocessed as follows. Categorical variables are encoded with dummies and \(\mathbf{x}_i\) is standardized. The PM frequency, overhauls and failures that occurred throughout the contract are converted to the number of events per running period. Even though a contract's exact number of running hours is not known in advance, an estimate is typically available.
## Semi-synthetic setup {#ssec:Semi_synthetic_data}
A good estimator should accurately predict both the observed outcome, the number of failures that did occur at maintenance frequency \(t_i\), as well as the unobserved outcomes, the number of failures if the machine had received less or more maintenance. In practice however, not all potential outcomes are observed, which makes evaluation of causal models hard. Because of this, we rely on semi-synthetic data to evaluate our model. This approach is commonly used in both causal inference and maintenance. Potential outcomes \(o_i(t)\) and \(f_i(t)\) are generated based on the observed characteristics \(\mathbf{x}_i\). For the overhauls, we have: \[o_i(t) = 7 \, \sigma\Bigg( {\underbrace{\mathbf{v}_o^\intercal \mathbf{x}_i \vphantom{\frac{1}{10}}}_{\text{Base rate}}}-\, {\underbrace{\frac{1}{10} \, \sigma\left(\mathbf{w}_o^\intercal \mathbf{x}_i \right)}_{\text{PM effect}}} t \, + {\underbrace{\epsilon_o \vphantom{\frac{1}{10}}}_{\text{Noise}}} \Bigg)\] where \(\mathbf{v}_o, \mathbf{w}_o \sim \mathcal{U}\left((0, 1)^{d\times1}\right)\) and \(\epsilon_o \sim \mathcal{N}(0, 1)\). The 7 rescales the average number of overhauls to roughly same number in the original data. For failures, we similarly have: \[f_i(t) = 9 \, \sigma\bigg(\mathbf{v}_f^\intercal \mathbf{x}_i-\frac{1}{10} \, \sigma\left(\mathbf{w}_f^\intercal \mathbf{x}_i\right) t + \epsilon_f\bigg)\] with \(\mathbf{v}_f, \mathbf{w}_f \sim \mathcal{U}\left((0, 1)^{d\times1}\right)\) and \(\epsilon_f \sim \mathcal{N}(0, 1)\). Using the semi-synthetic setup, the test set contains the potential outcomes for all possible values of \(t_i \in \mathcal{T}\) using these equations. Conversely, the training and validation sets include only one observed outcome for one observed \(t_i\). The training, validation and test sets respectively consist out of 50%, 25% and 25% of the data. Hyperparameter optimization is based on the mean squared error on the observed outcomes in the validation set. An illustration of a generated data set is shown in [\[fig:semi_synthetic_data\]](#fig:semi_synthetic_data){reference-type="ref" reference="fig:semi_synthetic_data"}. On the one hand, we evaluate the prescribed maintenance frequencies using the maintenance frequency \(t_i\) that was observed in practice for the observed outcomes in the training and validation set. On the other hand, we want to evaluate our policy for different levels of selection bias. For this, we control the level of selection bias in the semi-synthetic data using an approach similar to. Selection bias is simulated by assigning PM frequencies from a beta distribution as follows: \[t_i \sim 20 \, \text{Beta}\left(1 + \frac{\lambda \delta_i}{10}, 1 + \lambda \delta_i\right) \label{eq:sim_selection_bias_lambda}\] where \(\delta_i = \sigma(\textbf{w}_b \mathbf{x}_i)\) with \(\textbf{w}_b \sim \mathcal{U}\left((0, 1)^{d\times1}\right)\). \(\delta_i\) ensure that treatment assignment is based on observed features \(\mathbf{x}_i\). This way, \(\lambda\) controls the level of selection bias. \(\lambda = 0\) results in \(\text{Beta}(1, 1)\) or the uniform distribution, which implies random maintenance assignment. Higher values of \(\lambda\) imply more selection bias with \(\lambda=30\) resulting in a maintenance distribution similar to the observed distribution. An illustration of the observed distribution and generated distributions for different values of \(\lambda\) is shown in [\[fig:beta_distribution\]](#fig:beta_distribution){reference-type="ref" reference="fig:beta_distribution"}.
## Evaluation {#ssec:Evaluation}
Evaluation is done using three different metrics. First, we evaluate the ability of the machine learning model to accurately predict a contract's potential outcomes. This is measured using the mean integrated square error (MISE): \[\text{MISE} = \frac{1}{n} \sum_{i=1}^n \int_0^m \left(y_i(t)-\hat{y}_i(t) \right)^2 \,\diff t .\] Second, we want to evaluate the accuracy of the prescribed maintenance frequencies. To this end, we consider a variant of the policy error (PE) that compares the prescribed maintenance frequency with the ideal level: \[\text{PE} = \frac{1}{n} \sum_{i=1}^n \left(t^*_i-\hat{t}^*_i \right)^2 .\] Third, we evaluate the prescribed maintenance frequency in terms of costs using the policy cost ratio (PCR) that compares the costs of the estimated optimal maintenance frequency with the ideal level: \[\text{PCR} = \frac{1}{n} \sum_{i=1}^n \frac{c_i(\hat{t}_i^*)}{c_i(t_i^*)} .\] For all metrics, a lower value indicates better performance with \(0\) being the optimal value for MISE and PE and \(1\) for PCR. Our proposed maintenancy policy uses SCIGAN to learn the individual treatment effects (ITE) and will be referred to as SCIGAN--ITE. We benchmark this policy to two other policies (see [\[tab:methodologies\]](#tab:methodologies){reference-type="ref" reference="tab:methodologies"}). First, a policy based on a neural network (MLP) that learns \(o_i\) and \(f_i\) given \(\mathbf{x}_i\) and \(t_i\) in a completely supervised manner without adjusting for selection bias (MLP--ITE). This allows us to assess whether there is a benefit of adjusting for selection bias. Second, the average policy (SCIGAN--ATE) sets a single optimal \(t^*\) for all contracts based on the average (instead of the individual) maintenance effect. This allows to validate the benefit of an individualized policy tailored towards each specific machine.
## Empirical results {#ssec:Empirical_results}
In this section, we present the results of the semi-synthetic experiments based on more than 4,000 maintenance contracts, as put forward in [\[ssec:Data,ssec:Semi_synthetic_data,ssec:Evaluation\]](#ssec:Data,ssec:Semi_synthetic_data,ssec:Evaluation){reference-type="ref" reference="ssec:Data,ssec:Semi_synthetic_data,ssec:Evaluation"}. The goal is to answer two research questions. First, does an individualized approach outperform a general approach? Second, does a causal, prescriptive approach outperform a supervised, predictive approach? We aim to answer these for the observed maintenance frequency ([5.4.1](#sssec:observed_results){reference-type="ref" reference="sssec:observed_results"}) and assess the different policies' sensitivity to varying levels of selection bias ([5.4.2](#sssec:selection_bias_results){reference-type="ref" reference="sssec:selection_bias_results"}).
### Results for the observed PM frequencies {#sssec:observed_results}
We present the results for the different methodologies given the maintenance frequency \(t_i\) that was observed in practice in [\[tab:results\]](#tab:results){reference-type="ref" reference="tab:results"} and [\[fig:decisions_comparison\]](#fig:decisions_comparison){reference-type="ref" reference="fig:decisions_comparison"}. For both failures and overhauls, SCIGAN more accurately predicts the potential outcomes in terms of MISE compared to MLP, the supervised approach. Moreover, the individualized, prescriptive approach (SCIGAN--ITE) most accurately prescribes the optimal PM frequency compared to the supervised (MLP--ITE) and non-individualized approach in terms of policy error. Finally, SCIGAN--ITE also results in lower costs compared to both MLP--ITE and SCIGAN--ATE. The improved performance of SCIGAN--ITE compared to MLP--ITE illustrates the importance of adjusting for selection bias when learning from observational data. Moreover, the relatively worse performance of the average approach, SCIGAN--ATE, indicates the benefit of an individualized, machine-dependent policy for imperfect maintenance that takes into account machine characteristics.
:::
### Results for different levels of selection bias {#sssec:selection_bias_results}
We compare performance for the SCIGAN--ITE and MLP--ITE for different levels of selection bias in terms of \(\lambda\) (see [\[eq:sim_selection_bias_lambda\]](#eq:sim_selection_bias_lambda){reference-type="ref" reference="eq:sim_selection_bias_lambda"}). SCIGAN achieves good predictive performance in terms of MISE for the entire range of operating conditions, ranging from randomized PM assignments (\(\lambda = 0\)) to realistic levels of selection bias (\(\lambda = 30\)). Conversely, the MLP, a completely supervised approach that does not adjust for selection bias, accurately predicts the potential outcomes when preventive maintenance is randomized (\(\lambda=0\)), but results in bad predictions for high levels of \(\lambda\). Similarly, SCIGAN is robust towards higher levels of \(\lambda\) in terms of decision-making, illustrated by stable values for PE and PCR across different levels of selection bias, whereas MLP results in less accurate and more costly decisions as bias increases.
# Conclusion {#sec:Conclusion}
This work proposes a novel prescriptive maintenance policy that accounts for imperfect maintenance effects by learning a machine-dependent maintenance effect conditional on the machine's characteristics from observational data. This is achieved by relying on state-of-the-art machine learning methodologies for causal inference. The benefit of our approach is that, unlike existing approaches, our methodology does need strong assumptions regarding the maintenance effect, but is instead able to learn it from observational data using flexible machine learning models. We validate our approach with semi-synthetic experiments using real-life data on more than 4,000 full-service maintenance contracts. We find that our proposed approach outperformed both a supervised approach and non-individualized approach in terms of both accuracy and cost of the prescribed preventive maintenance schedules. Moreover, our work highlights the importance of dealing with selection bias when learning from observational data. These findings show that our proposed approach offers a powerful and flexible policy for individualized maintenance. Causal inference requires strong assumptions, as does the proposed methodology proposed in this work. The first, overlap, implies overlap between distributions of machine's receiving different levels of maintenance. Overlap can be tested and characterized from data. Moreover, recent work has looked at characterizing uncertainty in regions where overlap is violated. The second assumption, unconfoundedness, is untestable in practice. It can however be assessed by people with domain-knowledge that are in charge of making maintenance decisions. The relevant question is whether all relevant variables regarding past maintenance decisions are included in the data. If there are unobserved confounders, adequately adjusting for selection bias might not be possible, which would result in biased estimates of the potential outcomes. Recent work has suggested the possibility of sensitivity analyses to assess the effect of hidden confounders. Finally, quantifying ignorance regarding the potential outcomes due to possible violations of these assumptions has been proposed. In terms of future work, it would be valuable to consider different types of maintenance interventions in terms of intensities and costs. Similarly, it would be useful to include more complex costs in this framework, such as stochastic costs or costs that need to be predicted from maintenance or machine characteristics. Moreover, it would be interesting to incorporate more flexible timing of maintenance interventions in our methodology and consider sequences of different maintenance interventions. Sequences of treatments have also received attention in the literature on causal inference. Finally, it would be interesting to look at ways of more closely integrating the predictive model in the decision-making step, e.g. by using predict-and-optimize or cost-sensitive approaches. | {'timestamp': '2022-06-06T02:14:23', 'yymm': '2206', 'arxiv_id': '2206.01562', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01562'} |
# Introduction
The task of constructing (quasi-)de Sitter configurations in string theory is an important topic of study both for its cosmological applications as well as for the general understanding of the structure of string theory and quantum gravity. The feasibility of de Sitter vacua in string theory remains a point of contention, with various proposals and objections, but no complete and explicit constructions to date. At least part of the challenge in constructing de Sitter space in string theory can be attributed to the fact that String theory is intrinsically supersymmetric, while de Sitter space is a positive energy configuration and therefore intrinsically non-supersymmetric. This means that any construction of de Sitter space within string theory must involve some form of supersymmetry breaking and various questions of stability and control of corrections must be approached with great care. This is, of course, also true of other non-supersymmetric scenarios within string theory. One possible approach to non-supersymmetric scenarios in string theory is to work directly with a system where supersymmetry is broken already at the string scale, however this gives up much of the computational control that supersymmetry offers and much remains to be understood in such systems. A more common approach is to have the supersymmetry breaking occur at lower energies where a lower-dimensional effective theory description is available. The form of the effective theory is then restricted by supersymmetry and the supersymmetry breaking itself can be understood in terms of specific string theory ingredients of the compactification. The possible sources of supersymmetry breaking in 4-d supergravity theories can come in the form of gauging global symmetries, superpotentials or explicit matter sectors with non-linear supersymmetry realizations. Of these options, the first is available in minimal as well as extended supergravity, and can be realized through internal fluxes in type II Calabi-Yau string compactifications. Superpotentials are of course only present in \(\mathcal{N}=1\) supersymmetric theories and also trace their origins to internal fluxes as well as possible non-perturbative effects. Non-linear realizations of supersymmetry, in the form of constrained superfields, occur on the worldvolumes of anti-branes and are a staple ingredient in most proposed stable de Sitter vacuum constructions. In this contribution, we will explore some obstacles to constructing controlled de Sitter states using these ingredients. \(\mathcal{N}=1\) superpotentials are the least constrained by any general physical principles, so we will not discuss them here. In section [2](#sec2){reference-type="ref" reference="sec2"}, we will present constraints that arise from combining the strict relationship between gauging in \(\mathcal{N}=2\) supergravity to the scalar potential with the magnetic Weak Gravity Conjecture, originally studied in. The results reveal an incompatibility between the Hubble scale of de Sitter critical points with charged but massless gravitini and the UV cutoff dictated by the conjecture. In section [3](#sec3){reference-type="ref" reference="sec3"} we will turn our attention to the dynamics of the non-linear supersymmetry sector described by constrained superfields, and summarize the approach and results of, where evidence of a previously unnoticed instability toward the formation of goldstino condensates was found. This instability potentially greatly alters the expected properties of string theory configurations involving anti-branes and, depending on its eventual endpoint, calls for a re-examination of many existing de Sitter constructions.
# Extended Supersymmetry and the Weak Gravity Conjecture {#sec2}
The scalar potential in supergravity theories with extended supersymmetry arises entirely from the gauging. The values of the energy at critical points of the scalar potential is therefore related to the values of the gauge couplings. On the other hand, general quantum gravity considerations, stemming primarily from the dynamics of charged black holes, produce some of the most robust constraints on effective theories that admit a quantum gravity UV completion. One of the most established such constraint is the Weak Gravity Conjecture (WGC), which can be obtained from the requirement that extremal black holes should be able to decay. This implies the existence of particles with \(U(1)\) charge greater than their mass (the electric Weak Gravity Conjecture), but also implies that the cutoff of any consistent effective theory with an unbroken \(U(1)\) gauge symmetry with coupling \(g\) is bounded by \[\Lambda_{UV} \leq g q M_p\] for every non-vanishing charge \(q\). The latter constraint is called the magnetic Weak Gravity Conjecture (mWGC). In it was further argued that the mWGC can be used to constrain de Sitter critical points and quasi-de Sitter configurations, by further demanding that the Hubble scale lies parametrically below the mWGC cutoff. \[H \ll \Lambda_{UV}\] This criterion can be motivated by the observation that, since powers of the Hubble scale govern the typical magnitude of higher curvature terms in the action, \(H/\Lambda_{UV}\) appears as the EFT expansion parameter. Furthermore, the presence of thermal fluctuations in de Sitter space, whose size is similarly given by \(H\), demanding that these fluctuations do not push the system outside the EFT regime of validity similarly requires parametric separation between the Hubble scale and the mWGC cutoff. The above argument, combined with the connection between the scalar potential and gauging in extended supergravity, opens the door for some strict constraints on de Sitter critical points in these theories. Indeed in it was noticed that many of the de Sitter critical points that one can obtain in models containing only vector multiplets have a Hubble scale on the order of the mWGC cutoff and therefore suffer a breakdown of their EFT description. It was further noted that these examples involved a vanishing gravitino mass matrix, and so it was conjectured that massless gravitini could serve as a signal for the breakdown of effective field theory descriptions. This idea was further elaborated in connecting massless gravitini with the swampland distance conjecture. For a more detailed discussion of the connections between de Sitter space, the mWGC and other swampland conjectures, we refer to the contribution to these proceedings by Niccolò Cribiori and references therein. Finally, in, the analysis of was extended to models with both vector and hypermultiplets and a rigorous proof was given that any de Sitter critical point in \(\mathcal{N}=2\) supergravity with charged massless gravitini will have an energy on the order of or above the cutoff dictated by the mWGC. In the rest of this section we will review the general proof of the main result as well as give two examples one of which illustrates the result in action while the other provides an example of de Sitter critical points that evade exclusion on mWGC grounds, due to the lack of an unbroken \(U(1)\) and a non-vanishing gravitino mass matrix. Both examples also illustrate some subtleties related to possible gauge group enhancements at certain points in field space.
## The general result
Here we prove that de Sitter critical point in \(\mathcal{N}=2\) supergravity with charged massless gravitini do not have parametric separation between the Hubble scale and the mWGC cutoff. Recall that the gravitino mass and charge matrices can be written as \[\label{gravMQ} \begin{aligned} S_{ij} &= i P^x_\Lambda L^\Lambda (\sigma_x)_i^k \epsilon_{jk} \\ (Q_A)_i^{\ j} &= \frac{1}{2} \mathcal{E}^\Lambda_A \big( P^0_\Lambda \delta_i^{\ j} + P^x_\Lambda (\sigma^x)_i^{\ j} \big) \end{aligned}\] where \(L^\Lambda\) denotes the choice of covariantly holomorphic section that specifies the scalar manifold for the vector multiplets and \(P_\Lambda^{0,x}\) denote the prepotentials, which are computed from the killing vectors that specify the scalar manifold isometries gauged by the available vector fields. The "gauge vielbein\" \(\mathcal{E}^\Lambda_A\) defines the canonically normalized charges and is related to the gauge-kinetic matrix \(\mathcal{I}\) by \[\mathcal{E}_A^\Lambda \mathcal{E}_B^\Sigma \delta^{AB} =-\mathcal{I}^{-1|\Lambda \Sigma}\] This means that the eigenvalues of \(Q_A\) are the physical gravitino charges, which are to be used when applying the mWGC. Meanwhile, the scalar potential is a sum of three terms \[\mathcal{V} = \mathcal{V}_1 + \mathcal{V}_2 + \mathcal{V}_3\] with \[\label{potterms} \begin{aligned} \mathcal{V}_1 &= g_{I\bar{J}} k^I_\Lambda k^{\bar{J}}_\Sigma \bar{L}^\Lambda L^\Sigma \\ \mathcal{V}_2 &= 4 h_{uv} k^u_\Lambda k^v_\Sigma L^\Lambda \bar{L}^\Sigma \\ \mathcal{V}_3 &= (U^{\Lambda \Sigma}-3 L^\Lambda \bar{L}^\Sigma ) P^x_\Lambda P^x_\Sigma \end{aligned}\] where \(g_{IJ}\) is the metric on the scalar manifold for the vector multiplets, \(h_{uv}\) is the metric on the scalar manifold of the hypermultiplets and \(U^{\Lambda \Sigma} = g^{I\bar{J}} \nabla_I L^\Lambda \nabla_{\bar{J}} \bar{L}^\Sigma =-\frac{1}{2} (\mathcal{I}^{-1})^{\Lambda \Sigma}-\bar{L}^\Lambda L^\Sigma\). A vanishing mass matrix \(S_ij\) requires \(P^x_\Lambda L^\Lambda = 0\). This simplifies the expression for the scalar potential to \[\mathcal{V} =-\frac{1}{2}\mathcal{I}^{-1|\Lambda \Sigma} \big( P^0_\Lambda P^0_\Sigma + P^x_\Lambda P^x_\Sigma \big) + 4 h_{uv} k^u_\Lambda k^v_\Sigma L^\Lambda \bar{L}^\Sigma\] where in the first term we used the relations between the killing vectors \(k_\Lambda^I\), their corresponding prepotentials \(P_\Lambda^0\), and the gauge-kinetic matrix to rewrite \(\mathcal{V}_1\). \(\mathcal{V}_3\) itself simplifies to a similar looking form, but involving the prepotentials associated to the hypermultiplet scalar manifold. The last term is simply \(\mathcal{V}_2\) and is positive definite. Since we are interested in finding a lower bound on the scalar potential we can drop this term and write \[\mathcal{V} \geq \frac{1}{2} \delta^{AB} \big( P^0_A P^0_B + P^x_A P^x_B \big)= \frac{1}{2} \delta^{AB} \big(\delta_i^j P_A^0 + \sigma^x \ _i^j P_A^x \big) \big(\delta_i^j P_B^0 + \sigma^x \ _i^j P_B^x \big)\] where we traded the gauge indices \(\Lambda, \Sigma\) for their "flat\" counterparts \(A,B\) using \(\mathcal{E}_\Lambda^A\) that we introduced earlier. In the second equality, we re-expressed the sum over the \(0,x\) in terms of the traces of products of Pauli matrices (along with the identity). We recognize the terms in the parentheses as the expressions for the gravitino charge matrix given in [\[gravMQ\]](#gravMQ){reference-type="eqref" reference="gravMQ"}. In particular, we can choose \(\mathcal{E}_\Lambda^A\) such that the \(U(1)\) charge that we intend to use for the mWGC lies along the \(A=1\) direction and we can thus write the lower bound \[\mathcal{V} \geq \frac{1}{2} \text{Tr} (Q_1 Q_1) = q_1^2 + q_2^2\] where \(q_1\) and \(q_2\) are the physical charges of the individual gravitini, which themselves provide an upper bound on the UV cutoff from the mWGC \(q_1^2+q_2^2\geq \Lambda_{UV}^2\). The final conclusion is that we have \[\mathcal{V} \geq q_1^2 + q_2^2 \geq \Lambda_{UV}^2 \implies H \geq \Lambda_{UV}/\sqrt{3}\] Thus the Hubble scale associated to the value of the potential is bounded below by an energy of order the mWGC cutoff, invalidating the EFT description of such critical points. [^1] A couple of remarks are in order here. First, the proof assumes that the full mass matrix vanishes, i.e. that both gravitini are massless. This assumption completely removes the negative contributions to the potential, allowing us to place the lower bounds in the way that we do. Second, the proof does *not* assume that we are precisely at a critical point of the potential, except for the last step where the Hubble scale is determined from the potential. This means that the conclusion should also be valid for quasi-de Sitter backgrounds, as long as the notion of a cosmological horizon and a corresponding Hubble scale makes sense. We will see both these considerations at play in the second example presented in the next subsection. Finally an important note is that our result is contained within the "Festina-Lente\" bound, which places a lower bound on the masses of all charged particles in a de Sitter background. This bound can be derived from considering the evaporation process of large black holes in de Sitter space and its connection to the Weak Gravity Conjecture has also been discussed. It is a non-trivial cross check that at least a sub-statement of this bound can also be derived purely on mWGC grounds.
## Examples and Caveats
As an illustrative example of our result at work, as well as of its caveats and possible extensions we will present two very similar models, which have the same matter content, but differ in the gauging. A distinguishing feature of the first example is a flat direction that tunes the gravitino mass, interpolating between regions that respect or violate the mWGC. The second example has several critical points, which exhibit either non-abelian enhancement of the gauge group, or its complete breaking, both of which are obstacles to applying the mWGC. Additional examples with interesting properties can be found in, including fully stable de Sitter critical points, whose EFT validity are however ruled out by the mWGC. Earlier works on finding de Sitter critical points in extended supergravity include The models we are going to consider will have the following scalar manifolds \[\mathcal{M}_{SK} = \frac{\text{SU(1,3)}}{\text{SU(3)}\times \text{U(1)}} \, \qquad \mathcal{M}_{QK} = \frac{\text{SO(4,2)}}{\text{SO(4)}\times \text{SO(2)}} \,,\] with holomorphic section \[Z = \begin{pmatrix} 1 \\ z^I \\-i/2 \\ i z^I / 2 \end{pmatrix} \, \qquad I = 1,2,3 \.\] There is an \(SO(3)\) isometry of \(\mathcal{M}_{SK}\), which rotates the \(z^I\) into each other, given by the killing vectors \[\kappa_1^I = \begin{pmatrix} 0 \\ z_3 \\-z_2 \end{pmatrix} \, \quad \kappa_2^I = \begin{pmatrix}-z_3 \\ 0 \\ z_1 \end{pmatrix} \, \quad \kappa_3^I = \begin{pmatrix} z_2 \\-z_1 \\ 0 \end{pmatrix}\] while \(\mathcal{M}_{QK}\) has an \(SO(3)\) isometry generated by \(T_{\underline{12}}, T_{\underline{13}}, T_{\underline{23}}\), an \(O(1,1)\) generated by \(T_{\underline{46}}\) and a \(U(1)\) isometry generated by \(T_{\underline{56}}\), where \[(T_{\underline{ab}})_{\underline{c}}{}^{\underline{d}} = \eta_{\underline{c}[\underline{a}} \delta_{\underline{b}]}^{\underline{d}}\] are the \(SO(4,2)\) generators used in the construction of \(\mathcal{M}_{QK}\). The \(\mathcal{M}_{SK}\) isometries are gauged by the three vector multiplets as \[k_\Lambda^I = e_1\, \big( 0 \, \ \kappa_1^I \, \ \kappa_2^I \, \ \kappa_3^I \big) \.\] and the SO(3)\(\times\)O(1,1) isometry of \(\mathcal{M}_{QK}\) are gauged as \[\begin{aligned} k_\Lambda^u = \big( e_0 k_{T_{\underline{46}}}^u \, \ e_1 k_{T_{\underline{12}}}^u \, \ e_1 k_{T_{\underline{13}}}^u \, \ e_1 k_{T_{\underline{23}}}^u \big). \end{aligned}\] i.e. the \(SO(3)\) isometry is still gauged by the three vector multiplets, with the same charge \(e_1\) as the \(\mathcal{M}_{SK}\) isometries, while the \(O(1,1)\) is gauged by the graviphoton with charge \(e_0\). The corresponding prepotentials given by \[\label{prepot} \begin{aligned} P_0^x = \begin{pmatrix} 0 \\ 0 \\ 0 \end{pmatrix} \, \quad P_1^x = \begin{pmatrix} e_1 \\ 0 \\ 0 \end{pmatrix} \, \quad P_2^x = \begin{pmatrix} 0 \\ e_1 \\ 0 \end{pmatrix} \, \quad P_3^x = \begin{pmatrix} 0 \\ 0 \\ e_1 \end{pmatrix} \, \end{aligned}\] The resulting scalar potential has a critical point at the center of field space with \[\begin{aligned} {\cal V} = 2\, e_0^2 + 3\, e_1^2 \, \end{aligned}\] and scalar masses related to the value of the potential as \[\begin{aligned}
\begin{split} m^2_{(mult.)} &= \left( 0_{(1)} \, \ 2(e_0^2-e_1^2)_{(3)} \, \ 4 e^2_{0 \ (1)} \, \ 4 e^2_{1 \ (2)} \right. \, \\ & \quad x_{1 \ {(1)}} \, \ x_{2 \ {(1)}} \, \ x_{3 \ {(1)}} \\[2mm] & \quad e_1^2 + \sqrt{4 e_0^4-4 e_0^2 e_1^2 + 9 e_1^4}_{(2)} \, \\ & \left. \ e_1^2-\sqrt{4 e_0^4-4 e_0^2 e_1^2 + 9 e_1^4}_{(2)} \right) \times {\cal V} \, \end{split} \end{aligned}\] where \(x_1,x_2,x_3\) are the solutions to \[x^3+6e_1 x^2 + (4 e_0^2 e_1^2-4e_0^4)x-(16e_0^4e_1^2-16e_0^2e_1^4+32e_1^6)=0\] Note that the critical point is unstable and its mass spectrum marginally respects the de Sitter conjecture. If we further choose \(e_0=e_1\) the scalar potential develops a flat direction along the imaginary directions of each of the \(z^I\). Moving along any one of these flat directions will break the \(SO(3)\) isometry to a \(U(1)\), with the corresponding gravitino charge given by \[\label{gravCharge1} q_{U(1)}=\pm\sqrt{\frac{1+z^2}{1-z^2}}\] where \(z\) is the magnitude of the imaginary part of the scalar that parametrizes our chosen flat direction. This \(U(1)\) physical charge, and therefore the mWGC cutoff, interpolates between \(1\) and infinity as \(z\) grows from \(0\) to \(1\). Meanwhile, the gravitino mass along this direction is given by \[S_{ij} = \begin{pmatrix} e_1\frac{z}{\sqrt{1-z^2}} & 0 \\ 0 & e_1 \frac{ z}{\sqrt{1-z^2}} \end{pmatrix}\] which vanishes as \(z\to 0\) and diverges as \(z \to 1\). This flat direction thus offers a perfect opportunity to illustrate the onset of the mWGC violation as we approach the massless point. Indeed, computing the mWGC cutoff in units of the Hubble scale as a function of \(z\) we can see that the well-controlled regime appears precisely as \(z\to 1\), where the gravitini are heavy, while the massless gravitino limit has the Hubble scale on the same order as the mWGC cutoff (see Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}). Finally we also note that at \(z=0\) the unbroken gauge symmetry gets enhanced to \(SO(3)\), which is non-abelian and so the mWGC does not directly apply. However, since points along the flat directions are excluded by the mWGC arbitrarily close to the central critical point, we may conclude by continuity that the central point is similarly excluded. As our second example, we may choose instead to gauge the \(U(1)\) isometry on \(\mathcal{M}_{QK}\) instead of the \(O(1,1)\), i.e. choose \[\begin{aligned} k_\Lambda^u = \big( e_0 k_{T_{\underline{56}}}^u \, \ e_1 k_{T_{\underline{12}}}^u \, \ e_1 k_{T_{\underline{13}}}^u \, \ e_1 k_{T_{\underline{23}}}^u \big). \end{aligned}\] with the rest of the model exactly as before. In particular the prepotentials are still given by [\[prepot\]](#prepot){reference-type="eqref" reference="prepot"}. With this choice of gauging, the central critical point still exists and has energy \[\mathcal{V}=3 e_1^2\] with the mass spectrum \[m^2_{(mult.)}=\left(-\frac{2}{3}_{(6)} \, \ \frac{4}{3}r^2_{(2)} \, \ \frac{4}{3}(r^2+1)_{(6)}\right) \times {\cal V}\] The vanishing of \(P^x_0\) implies that the gravitini are uncharged under the separate \(U(1)\) isometry. They do have charges \(\pm \frac{1}{\sqrt{2}} e_1\) under the three generators of the \(SO(3)\) isometry. However this isometry is non-abelian, so it isn't clear that the mWGC constraint can be applied. Furthermore, we no longer have the flat direction that allows us to break it to a single \(U(1)\). A possible way around this problem is to consider time-dependent trajectories, which have as initial boundary conditions points near, but not quite at, the central critical point, in such a way that the \(SO(3)\) symmetry is broken down to \(U(1)\) as in the previous example. In this case, the charge under this \(U(1)\) will still be approximately equal to \(\pm \frac{1}{\sqrt{2}} e_1\) and the energy, at least for some short time before the runaway due to the tachyonic directions takes over, will be above the mWGC cutoff, so we expect the EFT description around these trajectories to break down. By considering a limit of such runaway trajectories such that their initial conditions approach the critical point, we might wish to again conclude by continuity that the critical point itself also suffers a breakdown of its EFT description. An important caveat to the above argument is that since the masses of the tachyons are comparable to the energy itself and therefore does not satisfy slow-roll conditions, it isn't clear that the mWGC criterion applies to such trajectories in the first place, since one can not meaningfully assign a Hubble radius to this solution. This question deserves further investigation. Finally, a second interesting feature of this model is the presence of a second family of critical points located at \[\text{Re}\, z_I = \pm\frac{1}{2} \, \ \text{Im}\, z_J = \pm\frac{1}{2} \, I\neq J \.\] This critical point completely breaks the \(SO(3)\) gauge symmetry, and the gravitino \(U(1)\) charge continutes to vanish. The scalar mass spectrum is \[\begin{aligned} m^2_{(mult.)}&=\left(0_{(3)},-1_{(2)}, 8_{(1)},2+4r^2-2r_{(2)},\beta^2+\beta_{(2)}, \beta^2-\beta_{(2)}, 2+2r+4r^2_{(2)}\right) \times {\cal V} \, \end{aligned}\] and thus again has tachyons satisfying the de Sitter criterion. This critical point can therefore not be excluded by this criterion, nor by the mWGC due to a lack of any unbroken gauge group. The reason this is not a counterexample to our general theorem is that one of the gravitini becomes massive. Indeed the mass matrix evaluated at this critical point is \[S_{ij} = \begin{pmatrix} \sqrt{2}\, e_1 & 0 \\ 0 & 0 \end{pmatrix} \,\] illustrating that insufficiency of having a single massless gravitino to exclude a critical point on mWGC grounds.
# Spontaneous Supersymmetry Breaking and Goldstino Condensation {#sec3}
In the previous section, we considered constraints on (quasi-)de Sitter configurations that arise from the magnetic Weak Gravity Conjecture and the connection between the scalar potential and the gauging in theories with extended supersymmetry. In this section we turn our attention to a more recent result concerning de Sitter constructions that can be formulated in \(\mathcal{N}=1\) supersymmetry models involving nilpotent chiral multiplet. This nilpotent superfield represents the spontaneous supersymmetry breaking sector that appears, for example, on the worldvolume of anti-branes and are ubiquitous in de Sitter constructions such as. What we will show in this section is that the dynamics of this supersymmetry breaking sector, as described by the Volkov-Akulov model, include an instability toward the formation of composite states of the goldstino. Depending on the endpoint of this instability, which we currently can not determine, this formation of goldstino composite states can potentially have devastating effects on existing de Sitter constructions that involve anti-branes uplifts. Even in the case of a benign endpoint to the instability, the properties of the resulting configuration are likely to differ from what is commonly assumed and deserves further examination. For the sake of brevity we restrict ourselves to the main ideas and an outline of the calculations involved in deriving this result. All the necessary additional details can be found in.
## Volkov-Akulov with Lagrange multipliers
Our starting point is the Volkov-Akulov model, which is a theory of a single goldstone fermion with Lagrangian \[\begin{aligned} {\cal L} =-f^2 + i G \sigma^m \partial_m \overline G-\frac{1}{4 f^2} \overline G^2 \partial^2 G^2-\frac{1}{16 f^6} G^2 \overline G^2 \partial^2 G^2 \partial^2 \overline G^2 \end{aligned}\] The higher derivative terms are required by supersymmetry, which is realized non-linearly, with the supersymmetry transformation taking the form \(\delta_\epsilon G =-\sqrt{2} f \epsilon +...\). This model appears, at least as a sub-sector, in the low-energy description of theories with spontaneous supersymmetry breaking. The supersymmetry breaking scale is given by \(\sqrt{f}\). The model can alternatively be written in superspace by means of a nilpotent chiral superfield \[\begin{aligned} X = \phi + \sqrt{2} \theta G + \theta^2 F \, \ X^2 = 0 \implies \phi = \frac{G^2}{2F} \end{aligned}\] with Lagrangian given by \[\begin{aligned} L&=\int d^4 \theta K(X, \overline{X}) + \bigg( \int d^2 \theta W(X) + c.c \bigg) \\ &= \int d^4 \theta |X|^2 + \left(\int d^2\theta f X + \text{c.c.} \right) \end{aligned}\] or equivalently, one can impose the nilpotency condition via a second non-dynamical chiral superfield \(T = \tau + \sqrt{2} \theta \lambda + \theta^2 B\) that acts as a lagrange multipler \[\begin{aligned} L &= \int d^4 \theta |X|^2 + \left(\int d^2\theta f X + T X^2 + \text{c.c.} \right) \end{aligned}\] Despite appearing to have additional fields, the scalar fields \(\phi, \tau\) as well as the fermion \(\lambda\) are all non-dynamical and can be written as an expansion in goldstino multilinears as \[\begin{aligned} \langle \phi \rangle \ \sim \ \Big{\langle} \frac{G^2}{f} \Big{\rangle} + \dots \, \quad \langle \tau \rangle \ \sim \ \Big{\langle} \frac{\partial^2 \overline G^2 }{f^2} \Big{\rangle} + \dots \end{aligned}\] and \(\lambda\) can be determined by acting with a supersymmetry transformation on \(\tau\). Thus the goldstino is still the unique independent degree of freedom in the theory.
## Detecting Composite States with Exact RG
The presence of goldstino self-interactions presents the possibility that besides the perturbative goldstino states, composite states might also appear in the theory, represented precisely by expressions such as those for \(\phi\) and \(\tau\) above. One way to determine whether this happens is to allow the theory to evolve along the Wilsonian RG flow and see whether the non-dynamical field \(\tau\) acquires a correct-sign kinetic term, similar to Nambu--Jona-Lasinio and composite Higgs models. In this case, not only would a kinetic term indicate that \(\tau\) itself represents dynamical states, but also the nilpotency condition on \(X\) would get relaxed and so \(\phi\) would also enter the theory as an independent degree of freedom. The behavior of these new composite states will then be governed by the resulting scalar potential for \(\phi\) and \(\tau\). Since the \(T\) superfield starts out with a vanishing kinetic term, small variations of the Wilsonian cutoff will produce a small kinetic term, which is equivalent to strong coupling. We can therefore not rely on any perturbative calculation in the couplings and must turn to the formalism of the Exact Renormalization Group. The main idea behind this formalism is that given a Wilsonian effective action \(L[\Phi; \mu]\) with cutoff \(\mu\), the partition function \[\begin{aligned} \mathcal{Z}[\Phi] &= \int \mathcal{D}\Phi \ e^{L_{\text{prop.}}[\Phi; \mu] + L_{\text{int.}}[\Phi; \mu] } \\ L_{\text{prop.}}[\Phi; \mu] &= \int \frac{d^4 k}{(2\pi)^4} \Phi^A(-k) C_{AB}^{-1}(k) \Phi^B(k) \\ L_{\text{int.}}[\Phi; \mu] &= \int \sum_\lambda g_\lambda(\mu) \prod_{\{ A\}_\lambda} \bigg( \frac{d^4 k_A}{(2\pi)^4} k_A^{n_{A,\lambda}}\Phi^A(k_A) \bigg) \delta(\sum_{ \{ A\}_\lambda } k_A) \end{aligned}\] should be independent of \(\mu\). Here we separated the action into a regularized propagator piece \(L_{\text{prop.}}\), which needs not include the full kinetic term, and an interaction piece \(L_{\text{int.}}\), which can still have term quadratic in the field. All fields, couplings and momenta are taken here to be rendered dimensionless by rescaling with appropriate powers of \(\mu\). The requirement of invariance of the partition function leads to the condition \[\label{ERG} \begin{aligned} \dot{L}_{\text{int.}} &\equiv-\mu \partial_\mu L_{\text{int.} } \\ &= \int \frac{d^4 k}{(2\pi)^4} \tilde{C}^{AB}(k) \bigg(\frac{\delta^2 L_{\text{int.}} }{\delta \Phi^A(-k) \delta \Phi^B (k) } + \frac{\delta L_{\text{int.}} }{\delta \Phi^A(-k) } \frac{\delta L_{\text{int.}} }{ \delta \Phi^B (k) } \bigg) \end{aligned}\] where \(\tilde{C}^{AB}\) is related to \(C^{AB}\) in a specified way that depends on the spin of the field. It is worth emphasizing that this flow equation does not make use of any small-coupling approximation. It does however produce an infinite system of equations for all the higher derivative couplings that will generically be generated by the RG flow. Solving these equations thus requires an appropriate choice of truncation of these equations, typically based on a derivative expansion. For our case, we resort to a truncation of the equations that we dub the Supersymmetric Local Potential Approximation (SLPA). In this approximation we ignore all contributions *to and from* any terms that can not be expressed in terms of a Kähler or superpotential. In superspace, this amounts to ignoring all terms that are higher-order in superderivatives. The SLPA has the benefit of preserving supersymmetry (assuming a supersymmetric regulator function) and keeping track of the kinetic terms of the scalar fields, which is what we are interested in. This is in contrast to the more common Local Potential Approximation (LPA) which would simply truncate all derivative interactions. The truncation of higher order terms means that we should not trust the solutions to the remaining equations for large changes in \(\mu\). This can be an issue when the qualitative features one is after do not appear immediately in the RG flow, such as the tachyonic behavior in composite Higgs models. As we will see, the features describing here will not suffer from this problem, and will be immediately visible even for a small decrease in \(\mu\). In order to apply the ERG formalism to the Volkov--Akulov model, we start with the following Kähler and superpotentials \[\label{kahsupot} \begin{aligned} K &= \alpha |X|^2 + \beta |T|^2 + g \, |T|^2 |X|^2 + \frac14 q \, |X|^4 \\ W &= f X + \frac12 T X^2 \end{aligned}\] and separate \(K\) into a (regularized) propagator and interaction part \[\begin{aligned} K_{\text{prop.}} &= c^{-1} |X|^2 + c^{-1} |T|^2 \\ K_{\text{int.}} &= (\alpha-1) |X|^2 + (\beta-1) |T|^2 + \gamma \mu^{-2} |T|^2 |X|^2 + \frac14 \zeta \mu^{-2} |X|^4 \end{aligned}\] where we assume the regulator function can be expanded as \(c(p^2/\mu^2) = 1+ \sum c_n \ p^{2n}/\mu^{2n}\). This action contains all the terms that will be generated by the RG flow within the SLPA. Since the action is manifestly supersymmetric, we can read off the flow of the couplings \(\alpha, \beta, \gamma\) and \(\zeta\) from any of the terms that contain them. A convenient choice are the quadratic terms in the auxiliary fields \(F\) and \(B\). The calculation is laborious but straightforward and can be found in full detail in. Here we simply quote the resulting flow equations \[\begin{aligned} \dot \zeta &=-2 \zeta-2 c_1 \,, \quad \dot \gamma =-2 \gamma-2 c_1 \, \\ \dot \beta &=-2 N \gamma \,, \quad \dot \alpha =-2 N (\gamma + \zeta) \, \end{aligned}\] The flow of the 4-point couplings \(\zeta\) and \(\gamma\) is generated by the second term in [\[ERG\]](#ERG){reference-type="eqref" reference="ERG"} from the \(TX^2\) term in the superpotential. Once these terms are generated, the first term in [\[ERG\]](#ERG){reference-type="eqref" reference="ERG"} generates corrections to the kinetic terms from these 4-point couplings.
## Goldstino condensation in Volkov-Akulov and Beyond
We now need to solve the Volkov--Akulov ERG equations with boundary conditions \[\alpha\Big{|}_{t = 0} = 1 \,, \quad \beta\Big{|}_{t = 0} = 0 \,, \quad \gamma\Big{|}_{t = 0} = 0 \,, \quad \zeta\Big{|}_{t = 0} = 0 \,\] with \(t = \log(\Lambda/\mu)\), where \(\Lambda \leq \sqrt{f}\) is a matching UV scale where we might imagine all other degrees of freedom are integrated out. \(N\) is the result of a momentum loop integral that appears in the evaluation of the first term in [\[ERG\]](#ERG){reference-type="eqref" reference="ERG"} and \(c_1\) is the first coefficient in the expansion of the regulator function. A typical monotonic regulator will have \(c_1<0\) and \(N<0\). The solution to this system of ODE's is \[\label{ERGsoln} \begin{aligned} \zeta &=-c_1 \left( 1-e^{-2t} \right) \,, \quad \gamma =-c_1 \left( 1-e^{-2t} \right) \, ,\\ \alpha &= 1-2 c_1 N + 4 c_1 N \left(t + \frac12 e^{-2 t} \right) \,, \\ \beta &=-c_1 N + 2 c_1 N \left(t + \frac12 e^{-2 t} \right) > 0 \. \, \end{aligned}\] The \(T\) field, which started its life as a Lagrange multiplier, acquires a positive kinetic term, indicating that the composite states of the goldstino that it represents enter the effective theory as independent degrees of freedom. Furthermore, we can evaluate the mass spectrum around the original Volkov--Akulov point at \(X=T=0\) and we find that the masses take the form \[\begin{aligned} m_{\pm}^2 =-\tilde{f}^2\left[ \left(\tilde{\gamma} + 4 \tilde{\zeta}\right) \pm \sqrt{\frac{16 \ \tilde{g}^2}{\tilde{f}^2} + \left(\tilde{\gamma}-4 \tilde{\zeta}\right)^2} \ \right] \. \end{aligned}\] where \(\tilde{f}, \tilde{\gamma}, \tilde{\zeta}, \tilde{g}\) are rescaled couplings obtained after canonically normalizing the kinetic terms and are all positive. It isn't hard to see that at least one of the masses is always negative and thus at least one of the scalars will develop an expectation value, representing the formation of a goldstino condensate. The typical behavior of the scalar potential for small \(t\) is depicted in the left side of Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. For larger \(t\) contributions beyond the SLPA will alter the flow from that given by [\[ERGsoln\]](#ERGsoln){reference-type="eqref" reference="ERGsoln"}. In particular additional couplings will enter the Kähler potential and affect the shape of the potential away from the central point. This may ultimately affect the endpoint of the instability and its physical interpretation. The central instability, however, is expected to survive, since both the generation of the kinetic term for \(T\) and the tachyonic behavior at the origin is due to the \(TX^2\) term in the superpotential. To remove the central instability, the effect of the higher-derivative contributions in the ERG flow would have to give corrections to the Kähler potential that compete with the effect of the superpotential term, which would indicate a breakdown of the EFT description altogether. While this is not ruled out, it would require an even bigger revision of what is commonly assumed about the IR behavior of the Volkov--Akulov model. In either case, an analysis beyond the SLPA is called for. So far we have only considered the pure Volkov--Akulov model without additional fields and in the rigid limit. Additional fields are unlikely to change the qualitative features found above. This is, again, due to the fact that the \(TX^2\) term in the superpotential is responsible for both the kinetic term of \(T\) and for the presence of the central tachyon. At the UV matching scale, \(T\) only appears in that single superpotential term and can not couple to any other fields in any larger model. The SLPA calculation for the kinetic term will therefor be unaffected. Beyond the SLPA, additional fields may have an effect and considerations of the previous paragraph apply. Once again, analysis to higher orders remains a necessary future direction. As far as extending this analysis to supergravity, a major obstacle is the absence of a supersymmetric regulator, which could be used to maintain a manifestly supersymmetric RG flow. This remains an open problem. That said, the natural expectation is that corrections to the rigid limit results will be suppressed by powers of \(M_p\) and so the main qualitative features of a dynamic \(T\) and a tachyonic critical point should remain. As a naive first check, one can simply embed the RG-evolved V--A model into the KKLT scenario by adding the RG-evolved Kähler and superpotentials [\[kahsupot\]](#kahsupot){reference-type="eqref" reference="kahsupot"} evaluated at small \(t\) to the pre-uplift KKLT Kähler and superpotentials. The details of this calculation are also described in and the result is depicted in the right side of Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. The main result is that the dS critical point shifts in the \(X,T\)-plane and also develops a tachyonic instability toward goldstino condensation. Of course one can look at more complicated models where the goldstino sector couples to other matter fields in more sophisticated ways, such as, but the instability of the original critical point toward goldstino condensation is expected to remain for the reasons outlined above. Finally, we note that since the Volkov--Akulov model is generally understood to be contained in the low-energy description of the worldvolume dynamics of anti-branes, the instability we find should be present in the simple case of the \(\overline{D3}/O3\) system in 10 dimensions. In this context, a string theoretic analysis of the goldstino composite states and their dynamics could give insights into the eventual fate of the instability that would be difficult to access using effective field theory methods.
# Conclusion
In this contribution, we have explored potential obstacles to constructing de Sitter spacetimes in supersymmetric effective theories as explored in. We have shown that in \(\mathcal{N}=2\) supergavity, (quasi-)de Sitter configurations with massless charged gravitini necessarily have a Hubble scale on the order of, or above, the UV cutoff dictated by the magnetic Weak Gravity Conjecture, thus invalidating their EFT description. This constraint excludes many of the known de Sitter critical points in \(\mathcal{N}=2\) supergravity, including all known stable ones. Possible future directions in this area include clarifying the limits of applicability of the mWGC criterion to time-dependent backgrounds, in particular those that pass near critical points with non-abelian enhancement of the gauge group. Investigations of constraints coming from the charges of other particles, particularly at points where the gravitini are uncharged. These two lines of analysis could help decide the fate of the critical points encountered in the second example of section [2](#sec2){reference-type="ref" reference="sec2"}. We have then turned our attention to models where supersymmetry breaking is achieved through an explicit goldstino sector where supersymmetry is non-linearly realized. We have shown, using exact renormalization group techniques, that the pure Volkov--Akulov model has an instability toward goldstino condensation and argued that additional matter or supergravity couplings are unlikely to remove the instability, but may affect its endpoint. In either case, this new effect invites us to re-examine the properties of string models involving anti-branes. Future directions in this area include moving beyond the SLPA to determine the possible endpoint of the instability as well as analyzing the composite states in the context of the \(\overline{D3}/O3\) system using worldsheet or string field theory methods. | {'timestamp': '2022-06-06T02:15:29', 'yymm': '2206', 'arxiv_id': '2206.01603', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01603'} |
# Introduction
Network geometry is one of the most important features of real-world networks , as it explains frequently observed network properties, including scale-invariance, high clustering, and overlapping community structures. The concept of geometry is intimately connected to the idea of similarity in networks, that two similar vertices are likely to connect. Typical examples are social networks, where people sharing the same interests or living in the same region are more likely to interact, or the world wide web, where webpages sharing similar contents are often linked. The similarity of vertices can be modeled through geometry in a random graph. More specifically, two individuals can be modeled as vertices in a geometric space, and their similarity is then associated with a small distance between them. This class of models is often called geometric network models. In the past decades, several geometric models have been introduced, aiming to model and understand real network properties . The simplest example of geometric models are the random geometric graphs, where an underlying Euclidean space is fixed, and nodes within a given radius are connected. However, random geometric graphs fail to describe the typical scale-free nature of real-world networks, where many nodes of low degree coexist with a few nodes of extremely large degrees. Instead, the hyperbolic space has been found to be extremely useful for modeling real-world networks , as it explains the heavy-tailed degree distributions that many networks posses, their efficient routing and their community structures . Interestingly, the hyperbolic model is equivalent to the renown \(\mathbb{S}^1\) model, where nodes are uniformly distributed on the circle, and a soft constraint is assigned to the degrees . Recently, a very intuitive network model called geometric inhomogeneous random graph has been proposed as generalization of hyperbolic random graphs . GIRGs combine the essential real-world properties of the hyperbolic random graph with simpler mathematical formulations that make them mathematically more tractable than hyperbolic random graphs. In these models each vertex is positioned in a geometric space, and has a weight representing its genuine ability to attract connections. Then, the probability of a connection between two vertices in the GIRG is an increasing function of the weight of each vertex, and a decreasing function of a distance between them. Although in many cases network geometry is a natural assumption, networks may not be directly endowed with a metric space, and in principle they could also not posses any kind of geometry. Determining whether a given network has an underlying latent hyperbolic geometric space has been highlighted as a major open problem . Indeed, usually we can only observe the network connections, and not the network features which can cause geometry. The main question is then: can we assess whether these connections were formed by some underlying geometry? It has often been assumed that high clustering, or a large amount of triangles, implies the presence of geometry . Intuitively, in a geometric space, the triangle inequality ensures that two neighbors of a vertex are likely to be close to each other, and therefore likely to be connected. In fact, triangles are a powerful statistic to distinguish geometric and non-geometric networks, when assuming that the former can be endowed with an Euclidean space . In this paper however, we show that pure triangle counts, as well as the average clustering coefficient, have limitations in detecting geometry induced by hyperbolic spaces. In fact, network models without any source of geometry may contain the same amount of triangles as hyperbolic random graphs, and may have arbitrary large average clustering coefficients. This is essentially due to the trade-off between weights and geometry in the GIRG representation of the hyperbolic model: connections can either be formed between high-weight nodes, or between close--by nodes. When the degree distribution has a particularly heavy tail, this trade-off favors triangles formed by high weighted vertices. Therefore, the underlying geometry cannot always be detected by triangle counts or clustering coefficients. In this work we develop a novel and powerful statistic called *weighted triangles* to detect hyperbolic geometry in networks, where triangles are weighted based on how likely they are to be caused by geometry. Triangles that carry low evidence for geometry are discounted, so that our statistic is able to differentiate hyperbolic geometry from non-geometric networks in a regime where standard triangle counts or clustering coefficients are not . We also provide experimental evidence that weighted triangles can successfully detect a geometric nature of real-world networks
# Models
In this paper, our benchmark model is the Geometric Inhomogeneous random graph (GIRG) . This is a powerful, but tractable, random graph model that generalizes the hyperbolic random graph defined by Krioukov et al. in . In the GIRG model, each vertex \(i\in \{1,2,\ldots,n\}\) is equipped with a weight \(h_i\) and a uniformly sampled position \(x_i\) on a \(d\)-dimensional torus \([0,1]^d\) endowed with the infinity norm. Weights are sampled independently from the Pareto distribution, with density \[\label{eq:paretolaw} \rho(h)=K_1 h^{-\tau},\] for any \(h > h_0 > 0\), where \(K_1\) is the normalizing constant and \(\tau\in(2,3)\). Two nodes \(i\) and \(j\), are then connected independently with probability \[\label{eq:edgeprobGIRG} p(h_i,h_j,x_i, x_j)=K_2 \min\left(\frac{h_ih_j}{\mu n} ||x_i-x_j||^{-d},1\right)^{\gamma},\] for some \(\gamma>1\) and where \(\mu\) denotes the average weight. The constant \(K_2\) is a correction factor, which ensures that the expected degree of any vertex \(i\) is proportional to its weight \(h_i\) (see Appendix [\[app:expgedreeGIRG\]](#app:expgedreeGIRG){reference-type="ref" reference="app:expgedreeGIRG"}). In its original formulation, Bringmann et al. defined the GIRG as an entire class of models, where the connection probability [\[eq:edgeprobGIRG\]](#eq:edgeprobGIRG){reference-type="eqref" reference="eq:edgeprobGIRG"} is only required to hold asymptotically. Here we focus on a specific element of this class, to have a practical and tractable comparison with non-geometric models. An important reason to study the GIRG is that for \(d=1,\gamma=\infty\) it is asymptotically equivalent to the hyperbolic random graph, as shown in. Thus, the asymptotic results on the GIRG shown here are valid for the hyperbolic random graph as well. To determine when geometry can be detected, we compare the GIRG to the inhomogeneous random graph (IRG). This model differs from the GIRG only in the sense that it is non-geometric. Specifically, the vertices now possess only weights \(h_i\), sampled from the Pareto distribution as in [\[eq:paretolaw\]](#eq:paretolaw){reference-type="eqref" reference="eq:paretolaw"}, and are connected with probability \[\label{eq:edgeprobIRG} p(h_i,h_j)=\min\left(\frac{h_i h_j}{\mu n},1\right).\] The two models have the same weight, and degree distribution, and differ only in their connection probability through the presence of geometry. We will therefore show that under classical triangle-based statistics, these models cannot be distinguished, making it impossible to distinguish hyperbolic geometry from simple scale-free networks.
# Triangle counts
The most intuitive statistic to detect geometry is the triangle count in the network, \(\triangle\). Indeed, one would expect a high number of triangles in geometric networks due to the triangle inequality. The number of triangles in the IRG scales as \(n^{3(3-\tau)/2}\) . In the GIRG on the other hand, the triangle count undergoes a phase transition: when \(\tau < 7/3\) it scales as \(n^{3(3-\tau)/2}\), and when \(\tau > 7/3\) it scales as \(n\) . Then, for small values of \(\tau\), we cannot distinguish geometry from the triangle counts scaling, as in both models the number of triangles is proportional to \(n^{3(3-\tau)/2}\). In Figures [\[fig:pure_triangles_models_large\]](#fig:pure_triangles_models_large){reference-type="ref" reference="fig:pure_triangles_models_large"} and [\[fig:pure_triangles_models_small\]](#fig:pure_triangles_models_small){reference-type="ref" reference="fig:pure_triangles_models_small"} we clearly see that when \(\tau = 2.1\), the number of triangles in geometric and non-geometric models grow with the same scaling. Theoretically, even when \(\tau\) is small, it could still be possible to distinguish geometric/non-geometric inhomogeneous models through triangle counts. Indeed, if the number of triangles is \(\triangle = A n^{3(3-\tau)/2} (1 +o(1))\), we may be able to establish the correct underlying model, by identifying the leading order term \(A\). To this end, we can use precise asymptotic results from our earlier work :
- In the IRG, \[\frac{\triangle}{n^{3(3-\tau)/2}} \stackrel{\P}{\longrightarrow} A_{\text{IRG}}.\]
- In the GIRG, when \(\tau < 7/3\), \[\frac{\triangle}{n^{3(3-\tau)/2}} \stackrel{\P}{\longrightarrow} A_{\text{GIRG}}.\]
Here, \(A_{\text{IRG}}\) and \(A_{\text{GIRG}}\) are explicit constants that depend on the parameters of the two models. If these are all known, then the two models could be distinguished using the difference between \(A_{\text{IRG}}\) and \(A_{\text{GIRG}}\). However, in realistic situations, one observes a network only through its connections, while the possible underlying geometry of the network and the vertex weights are usually not known. Therefore, in practice, we cannot estimate the parameters of the underlying models, and thus it is impossible to establish whether the triangle count is close to \(A_{\text{IRG}} n^{3(3-\tau)/2}\) or \(A_{\text{GIRG}} n^{3(3-\tau)/2}\). We conclude that the number of triangles is not a satisfactory statistic for inferring the hyperbolic geometry.
# Average clustering coefficient
Another common network statistic to measure the tendency of the nodes to form triangles, is the clustering coefficient, which has also been ascribed to indicate a network geometry . The average clustering coefficient captures the probability that two randomly chosen neighbors of a randomly chosen vertex are connected as well. Equivalently, it is the average fraction of realized triangles out of all possible triangles that include a randomly chosen vertex: \[\overline{C} = \frac{1}{n} \sum_{i \in V} C_i = \frac{1}{n} \sum_{i \in V} \frac{2 \triangle_i}{d_i(d_i-1)},\] where \(\triangle_i\) is the number of triangles containing vertex \(i\), and \(d_i\) is the degree of vertex \(i\). Typically, if the value of \(\overline{C}\) does not vanish in \(n\), then this is considered to be evidence for geometry. The average clustering in GIRGs is \(\Omega(1)\), that is, non-vanishing as \(n\) increases . In particular, in hyperbolic random graphs, \(\overline{C}\) converges in probability to a positive constant . For the IRG on the other hand, \(\overline{C}\) decays asymptotically, as \(n^{\tau-2}\ln(n)\) . Even though \(\overline{C}\) vanishes for large \(n\), the decay of \(\overline{C}\) is extremely slow when \(\tau \approx 2\), as we can observe in Figure [\[fig:average_clustering_models_large\]](#fig:average_clustering_models_large){reference-type="ref" reference="fig:average_clustering_models_large"}. Thus, the average clustering coefficient is impractical to detect geometry when \(\tau\) is small. Even more importantly, in non-geometric models with a different heavy-tailed degree distribution, \(\overline{C}\) can even be constant. For example, consider an inhomogeneous random graph with weights \[h_i=\begin{cases} 2, & \text{with probability }1-1/(\sqrt{\mu n}),\\ \sqrt{\mu n}, & \text{with probability } 1/(\sqrt{\mu n}). \end{cases}\] We show in Appendix [\[app:2-type_model\]](#app:2-type_model){reference-type="ref" reference="app:2-type_model"} that \(\overline{C}=\Theta(1)\) in this IRG, even though the model does not contain any source of geometry. Hence, \(\overline{C}\) too, is not a good statistic to distinguish between geometric and non-geometric networks.
# Weighted triangles
We will now describe our proposed statistic that, contrary to standard clustering-based statistics, is able to detect hyperbolic geometry. This statistic is again triangle-based so that it has the same computational complexity as other clustering-based statistics. The difference between our statistic and the average clustering coefficient or triangle count, is that all triangles are weighted based on their evidence for geometry. Indeed, in a hyperbolic space, triangles can be formed because of popularity (high-degree vertices), or similarity (close--by vertices). Thus, we weigh each triangle, so that triangles which carry low evidence for geometry have low weight, and triangles which carry high evidence for being formed due to geometry have high weight. More precisely, we define \[W := \sum_{\substack{i,j,k \in V \\ i<j<k}} \frac{1}{d_i d_j d_k}\mathbbm{1}_{\{(i,j,k) = \triangle\}},\] where \(\mathbbm{1}_{\{(i,j,k) = \triangle\}}\) is the indicator function of the event that the vertices \(i,j,k\) form a triangle in the network. The intuition behind the weights \((d_id_jd_k)^{-1}\) is that, in non-geometric power-law models, a typical triangle is formed between vertices of high degrees. Thus, when we *discount* the triangles formed by high-degree vertices, \(W\) remains small. On the other hand, in the GIRG, there is a large number of triangles formed by close--by vertices with small degrees. Therefore, \(W\) is significantly greater for GIRGs than for IRGs. Indeed, as we show in Appendix [\[app:concentrationW\]](#app:concentrationW){reference-type="ref" reference="app:concentrationW"}, \[W = \begin{cases} O(1), & \text{ w.h.p. in IRG},\\ \Omega(n), & \text{ w.h.p. in GIRG}. \end{cases}\] Thus, weighted triangles make a very powerful statistic to infer hyperbolic geometry: in the absence of geometry, the statistic remains bounded, while it scales at least linearly under a geometric model. Moreover, this different behaviour is independent of \(\tau\). This is clearly seen in Figures [\[fig:weighted_triangles_models_large\]](#fig:weighted_triangles_models_large){reference-type="ref" reference="fig:weighted_triangles_models_large"} and [\[fig:weighted_triangles_models_small\]](#fig:weighted_triangles_models_small){reference-type="ref" reference="fig:weighted_triangles_models_small"}, where \(W\) shows different scaling in non-geometric/geometric models, regardless of \(\tau\). It is interesting to observe that the average clustering coefficient can be seen as a specific way of weighing triangles as well: \[\label{eq:nCapprox} \begin{split} &n\overline{C} = \sum_{\substack{i,j,k \in V \\ i<j<k}} \left[\frac{1}{\binom{d_i}{2}} + \frac{1}{\binom{d_j}{2}} + \frac{1}{\binom{d_k}{2}}\right] \mathbbm{1}_{\{i,j,k = \triangle\}}\\ &\approx n \overline{C}_h: = \sum_{\substack{i,j,k \in V \\ i<j<k}} 2 \left( \frac{1}{h_i^2} + \frac{1}{h_j^2} + \frac{1}{h_k^2} \right) \mathbbm{1}_{\{i,j,k = \triangle\}}. \end{split}\] On the other hand, \[\label{eq:Wapprox} W \approx W_h := \sum_{\substack{i,j,k \in V \\ i<j<k}} \frac{1}{h_i h_j h_k} \mathbbm{1}_{\{i,j,k = \triangle\}}.\] By , triangles containing high-weighted vertices are likely to appear in scale-free models regardless of the geometry of the system. Thus, an effective statistic for geometry detection should discount triangles formed by high-degree vertices. From [\[eq:nCapprox\]](#eq:nCapprox){reference-type="eqref" reference="eq:nCapprox"} and [\[eq:Wapprox\]](#eq:Wapprox){reference-type="eqref" reference="eq:Wapprox"} we see that such discounting occurs in both \(n\overline{C}_h\) and \(W_h\). However, in \(W_h\), the factor \(1/(h_i h_j h_k)\) is small as soon as *at least one* of the three vertices has high weight. On the contrary, in \(n\overline{C}_h\), the factor \(1/h_i^2 + 1/h_j^2 + 1/h_k^2\) is small if and only if *all three vertices* have high weights. Therefore, \(W\) discounts triangles formed by high-weight vertices more effectively than \(n \overline{C}\) does. Indeed, Appendix [\[app:2-type_model\]](#app:2-type_model){reference-type="ref" reference="app:2-type_model"} shows that the average clustering coefficient may also be constant for non-geometric networks, while weighted triangles are able to detect the non-geometric nature of the underlying network. This makes weighted triangles \(W\) a better geometry detector than the average clustering coefficient.
# Real-world networks
We will now investigate the performance of our statistic on real-world data. As our statistic was designed to infer hyperbolic geometry, we focus only on networks with a heavy-tailed degree distribution. Furthermore, we investigate data sets that contain multiple snapshots of different sizes of the same network, to compare the theoretical scaling of \(W\) with the scaling found in the data. These networks are described in detail in Table [1](#tab:realnetworks){reference-type="ref" reference="tab:realnetworks"}.
:::
Figure [\[fig:real_networks\]](#fig:real_networks){reference-type="ref" reference="fig:real_networks"} shows the behavior of triangle counts \(\triangle\), the average clustering coefficient \(\overline{C}\) and weighted triangles \(W\). In all three data sets, \(W\) behaves roughly linearly in \(n\), as our theoretical analysis predicts for geometric networks. The clustering coefficient as well as the pure triangle counts on the other hand, has more erratic behavior that is difficult to interpret. We will now discuss the results for each network in some more detail. \(W\) grows linearly for the ArXiv collaboration networks (Figure [\[fig:real_networks\]](#fig:real_networks){reference-type="ref" reference="fig:real_networks"}a): this suggests that geometry is present in the network. This corresponds to an intuition that ArXiv collaborations have a natural geometric structure: researchers working in the same group, region, or country are more likely to write a paper together. Moreover, a collaboration between two researchers is more likely if they work on similar topics. The average clustering coefficient \(\overline{C}\) is rather small for the CAIDA AS networks (Figure [\[fig:real_networks\]](#fig:real_networks){reference-type="ref" reference="fig:real_networks"}b), and it seems to be decreasing, as the network size increases. On the other hand, the weighted triangles show a linear growth, revealing a source of geometry, that, in fact, is consistent with the domain knowledge about this network. Indeed, interactions between autonomous systems occur through peering or transit. An important part of peer selection is the following: internet service providers (ISPs) have a set of *points-of-presence*, that are the locations where they have routers/servers and related equipment/personnel across the world. An ISP can peer with another ISP only at locations where both have a point-of-presence. Such location-dependent connectivity rules naturally suggest the presence of an implicit geometry in the network formation. Lastly, the Gnutella peer-to-peer networks (Figure [\[fig:real_networks\]](#fig:real_networks){reference-type="ref" reference="fig:real_networks"}c) show an extremely small average clustering coefficient, failing to detect geometry. On the contrary, the weighted triangles reveal geometry, as \(W\) increases linearly with the network size. Indeed, it is reasonable to expect a geometric structure in this network because peer-to-peer file transfers are more likely between users who share similar interests. Then, we could construct an hidden metric space, by embedding users into this space according to their interests. Overall, the plots in Figure [\[fig:real_networks\]](#fig:real_networks){reference-type="ref" reference="fig:real_networks"} suggest that weighted triangles can be an effective statistic to detect geometry in real-world networks.
# Conclusion and discussion
In this paper we have analyzed in detail different methods to detect hyperbolic geometry in networks, when only the connections of the network are known. For theoretical purposes, we used the IRG and GIRG as benchmark models, on which we established analytical results. In both random graph models, the probability of connection depends on the popularity features of individuals vertices, expressed through their weights, so that the empirical degree distribution is power law with parameter \(\tau\). However, GIRG also models similarity of individuals, by locating the graph vertices randomly in a \(d\)-dimensional torus endowed with a metric. First, we have established that triangle counts and the average clustering coefficient have serious limitations in detecting the presence of geometry when the degree distribution has a particular heavy tail (when \(\tau<7/3\)). Using numerical simulations we confirmed these limitations, by comparing the number of triangles and the average clustering coefficient in different models. Next, we have introduced a new statistic, weighted triangles \(W\), that effectively discounts triangles formed by high degree vertices. The underlying intuition is that such triangles carry low evidence for geometry because they are common in power law networks with or without an underlying geometric space. Our main result states that \(W\) is a suitable statistic for detecting geometry, because, asymptotically in the network size, \(W\) remains bounded in IRGs, while it grows linearly in GIRGs. Numerical experiments on the real-world data show a remarkable agreement with this analysis, and confirm the high potential of the \(W\) statistic for uncovering a hidden network geometry. For practical applications, it would be interesting to provide some rule of thumb, how large \(W\) should be to signal geometry. In Appendix [\[app:W_IRG\]](#app:W_IRG){reference-type="ref" reference="app:W_IRG"} we obtain that \(\Exp{W} = O(1)\), in IRGs. Then from Markov's inequality it follows that for sufficiently large \(n\), \[\Prob{W \geq n^\alpha} \leq \frac{1}{ n^{\alpha-\varepsilon}} \to 0\] for any \(0<\varepsilon<\alpha<1\). Furthermore, in GIRGs, we have that \(\Exp{W} = \Omega(n)\), while \(\operatorname{Var}(W)=O(n)\) (see Appendices [\[app:W_GIRG\]](#app:W_GIRG){reference-type="ref" reference="app:W_GIRG"}-[\[app:varianceW\]](#app:varianceW){reference-type="ref" reference="app:varianceW"}). Hence, from Chebyshev's inequality, we get \[\Prob{W \leq \sqrt{n}} \to 0\] in the large network limit. This suggests \(\sqrt{n}\) as a possible threshold for \(W\) to detect geometry. However, if \(n\) is not sufficiently large, this asymptotic approach could fail. Indeed, even though \(W = \Omega(n)\) in GIRGs, the leading order term can be very small. For example, in Figure [\[fig:real_networks\]](#fig:real_networks){reference-type="ref" reference="fig:real_networks"}c, for \(n = 50000\), the estimated \(W\) is still smaller than 2. In fact, a smaller cutoff of for example \(n^{0.1}\) seems to work better. These empirical observations require further investigations. Based on our results that \(W\) is an asymptotically powerful statistic, designing statistical tests for finite \(n\) is a natural next step in the quest of detecting network geometry. | {'timestamp': '2022-06-06T02:13:53', 'yymm': '2206', 'arxiv_id': '2206.01553', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01553'} |
# Introduction and presentation of the results
This paper is a contribution to the study of hereditary classes of finite graphs. We classify these classes according to their proper subclasses. With this idea, our simplest classes are those who contain finitely many proper subclasses, hence these classes are finite. At the next level, there are the classes who contain infinitely many proper subclasses, but every proper subclass contains only finitely many. That is such classes are infinite but proper subclasses are finite. It is a simple exercise based on Ramsey's theorem that there are only two such classes: the class of finite cliques and the class of their complements. Pursuing this idea further, we would like to attach a rank to each class, preferably an ordinal. If we do this, it turns out that a class has a rank if and only if the set of its proper subclasses ordered by set inclusion is well founded. This latter condition amounts to the class being well-quasi-ordered (this follows from Higman's characterization of well-quasi-orders ). This puts forward the importance of well-quasi-ordered hereditary classes. A basic construction of well-quasi-ordered hereditary classes of finite graphs and more generally of finite structures goes as follows: chose a finite hereditary class of finite binary structures and take its closure under lexicographical sums over elements of the class. The fact that this latter class is well-quasi-ordered is a consequence of a theorem of Higman . An important property of such a class is that it contains only finitely many prime structures (see Definition [\[def:module\]](#def:module){reference-type="ref" reference="def:module"}). A concrete example of such a class is the class of finite cographs (prime structures in this class have cardinality at most two). A natural question then arises: under what conditions a class that contains infinitely many primes is well-quasi-ordered? Among hereditary classes which contain infinitely many prime members, we show that there are minimal ones with respect to set inclusion (Theorem [\[thm:minimalprime\]](#thm:minimalprime){reference-type="ref" reference="thm:minimalprime"}). Furthermore, we show that the minimal ones are well-quasi-ordered ages (Theorem [\[minimal\]](#minimal){reference-type="ref" reference="minimal"}). We obtain some general results that we are able to refine in some special cases like graphs, ordered sets, and bichains. We give a complete description of minimal prime ages of graphs (Theorem [\[thm:charact-minimal-prime-ages\]](#thm:charact-minimal-prime-ages){reference-type="ref" reference="thm:charact-minimal-prime-ages"}), of posets, and bichains (Corollary [\[cor:thm:charact-minimal-prime-ages-bichains\]](#cor:thm:charact-minimal-prime-ages-bichains){reference-type="ref" reference="cor:thm:charact-minimal-prime-ages-bichains"}). It turns out that there are \(2^{\aleph_0}\) such ages (Corollary [\[thm:minprimeages\]](#thm:minprimeages){reference-type="ref" reference="thm:minprimeages"}). Eleven of these ages are almost multichainable; they remain well-quasi-ordered when labels in a well-quasi-ordering are added, five being exhaustible. Among the remaining ones, countably many remain well-quasi-ordered when one label is added and these have finitely many bounds (except for the age of the infinite path and its complement). The others have infinitely many bounds (Theorem [\[thm:bound-uniform\]](#thm:bound-uniform){reference-type="ref" reference="thm:bound-uniform"}). Except for six examples, members of these ages we characterize are permutation graphs. In fact, every age which is not among the eleven ones is the age of a graph associated to a uniformly recurrent word on the integers (this is a consequence of Theorems [\[thm:permutation-graph\]](#thm:permutation-graph){reference-type="ref" reference="thm:permutation-graph"}, [\[thm:recurrent-word\]](#thm:recurrent-word){reference-type="ref" reference="thm:recurrent-word"} and [\[thm:uniformly-ages\]](#thm:uniformly-ages){reference-type="ref" reference="thm:uniformly-ages"}). This result supports the conjecture that if a hereditary class of finite graphs does not remain well-quasi-ordered by adding labels in a well-quasi ordered set to these graphs, then it is not well-quasi-ordered if we add just two constants to each of these graphs. Our description of minimal prime classes uses a description of minimal prime graphs [@pouzet-zaguia2009] and previous work by Sobrani and the authors on properties of uniformly recurrent words and the associated graphs. The completeness of our description is based on classification results of Chudnovsky, Kim, Oum and Seymour and Malliaris and Terry [@malliaris].
# Organisation of the paper
In section [3](#sec:preq){reference-type="ref" reference="sec:preq"} we present some prerequisites on graphs, posets and words. In section [\[section:min-hered-class\]](#section:min-hered-class){reference-type="ref" reference="section:min-hered-class"} we consider binary relational structures with a finite signature, we give the definition of a minimal prime hereditary class of binary structures and prove their existence. Section [\[section:min-hered-class\]](#section:min-hered-class){reference-type="ref" reference="section:min-hered-class"} contains also the proof of Theorem [\[minimal\]](#minimal){reference-type="ref" reference="minimal"} (see subsection [\[subsection:infinitely-prime\]](#subsection:infinitely-prime){reference-type="ref" reference="subsection:infinitely-prime"}) and a proof of Theorem [\[thm:main1\]](#thm:main1){reference-type="ref" reference="thm:main1"} (see subsection [\[sec:proof-thm:main1\]](#sec:proof-thm:main1){reference-type="ref" reference="sec:proof-thm:main1"}). In Section [\[section:minimalprimesgraphs\]](#section:minimalprimesgraphs){reference-type="ref" reference="section:minimalprimesgraphs"} we start with the classification results of Chudnovsky, Kim, Oum and Seymour and Malliaris and Terry [@malliaris]. Then, we present our main results on minimal prime ages. In Section [\[section:bornes\]](#section:bornes){reference-type="ref" reference="section:bornes"} we look at the number of bounds of our minimal prime ages. In section [\[sec:proof-thm:permutation-graph\]](#sec:proof-thm:permutation-graph){reference-type="ref" reference="sec:proof-thm:permutation-graph"} we provide a proof of Theorem [\[thm:permutation-graph\]](#thm:permutation-graph){reference-type="ref" reference="thm:permutation-graph"} and a characterization of order types of realizers of transitive orientations of \(0\)-\(1\) graphs. In section [\[sec:modules\]](#sec:modules){reference-type="ref" reference="sec:modules"} we characterize the modules of a \(0\)-\(1\) graph. We prove among other things, that if \(G_\mu\) is not prime, then \(\mu\) contains large factors of \(0\)'s or \(1\)'s. Section [\[section:embeddings\]](#section:embeddings){reference-type="ref" reference="section:embeddings"} is devoted to the study of the relation between embeddings of \(0\)-\(1\) words and their corresponding graphs. Results obtained in this section will be used in the proof of Theorem [\[thm:recurrent-word\]](#thm:recurrent-word){reference-type="ref" reference="thm:recurrent-word"}. In section [\[sec:proof-thm:recurrent-word\]](#sec:proof-thm:recurrent-word){reference-type="ref" reference="sec:proof-thm:recurrent-word"} we give a proof of Theorem [\[thm:recurrent-word\]](#thm:recurrent-word){reference-type="ref" reference="thm:recurrent-word"}. Theorem [\[thm:uniformly-ages\]](#thm:uniformly-ages){reference-type="ref" reference="thm:uniformly-ages"} is proved in section [\[sec:proof-thm:uniformly-ages\]](#sec:proof-thm:uniformly-ages){reference-type="ref" reference="sec:proof-thm:uniformly-ages"}. In section [\[sec:proof-thm:bound-uniform\]](#sec:proof-thm:bound-uniform){reference-type="ref" reference="sec:proof-thm:bound-uniform"} we investigate bounds of \(0\)-\(1\) graphs and give a proof of Theorems [\[thm:bound-uniform\]](#thm:bound-uniform){reference-type="ref" reference="thm:bound-uniform"}.
# Prerequisites {#sec:preq}
## Graphs, posets and relations
This paper is mostly about graphs and posets. Sometimes, we will need to consider binary relational structures, that is ordered pairs \(R:=(V,(\rho_{i})_{i\in I})\) where each \(\rho_i\) is a binary relation or a unary relation on \(V\). The set \(V\), sometimes denoted by \(V(R)\), is the *domain* or *base* of \(R\). The sequence \(s:= (n_i)_{i\in I}\) of arity \(n_i\) of \(\rho_i\) is the *signature* of \(R\) (this terminology is justified since we may identify a unary relation on \(V\), that is a subset \(U\) of \(V\), with the binary relation made of pairs \((u,u)\) such that \(u\in U\)). We denote by \(\Omega_s\) the collection of finite structures of signature \(s\). In the sequel we will suppose the signature finite, i.e. \(I\) finite. For example, we will consider *bichains*, i.e., relational structures \(R:= (V, (\leq', \leq''))\) made of a set \(V\) and two linear orders \(\leq'\) and \(\leq''\) on \(V\). The framework of our study is the theory of relations as developed by Fraı̈ssé and subsequent investigators. At the core is the notion of embeddability, a quasi-order between relational structures. We recall that a relational structure \(R\) is *embeddable* in a relational structure \(R'\), and we set \(R\leq R'\), if \(R\) is isomorphic to an induced substructure of \(R'\). Several important notions in the study of these structures, like hereditary classes, ages, bounds, derive from this quasi-order. For example, a class \(\mathcal C\) of relational structures, of signature \(s\), is *hereditary* if it contains every relational structure that embeds into a member of \(\mathcal C\). The *age* of a relational structure \(R\) is the class \(\age(R)\) of all finite relational structures, considered up to isomorphy, which embed into \(R\). This is an *ideal* of \(\Omega_s\) that is a nonempty, hereditary and *up-directed* class \(\mathcal{C}\) (any pair of members of \(\mathcal C\) are embeddable in some element of \(\mathcal C\)). A characterization of ages was given by Fraı̈ssé (see chapter 10 of ). Namely, a class \(\mathcal C\) of finite relational structures is the age of some relational structure if and only if \(\mathcal C\) is an ideal of \(\Omega_s\). We recall that a *bound* of a hereditary class \(\mathcal C\) of finite relational structures (e.g. graphs, ordered sets) is any relational structure \(R\not \in \mathcal C\) such that every proper induced substructure of \(R\) belongs to \(\mathcal C\). For a wealth of information on these notions see.
### Graphs
Unless otherwise stated, the graphs we consider are undirected, simple and have no loops. That is, a *graph* is a pair \(G:=(V, E)\), where \(E\) is a subset of \([V]^2\), the set of \(2\)-element subsets of \(V\). Elements of \(V\) are the *vertices* of \(G\) and elements of \(E\) its *edges*. The *complement* of \(G\) is the graph \(\overline{G}\) whose vertex set is \(V\) and edge set \({\overline { E}}:=[V]^2\setminus E\). If \(A\) is a subset of \(V\), the pair \(G_{\restriction A}:=(A, E\cap [A]^2)\) is the *graph induced by \(G\) on \(A\)*. A *path* is a graph \(\mathrm P\) such that there exists a one-to-one map \(f\) from the set \(V(\mathrm P)\) of its vertices into an interval \(I\) of the chain \({\mathbb Z}\) of integers in such a way that \(\{u,v\}\) belongs to \(E(\mathrm P)\), the set of edges of \(\mathrm P\), if and only if \(|f(u)-f(v)|=1\) for every \(u,v\in V(\mathrm P)\). If \(I=\{1,\ldots,n\}\), then we denote that path by \(\mathrm P_n\); its *length* is \(n-1\) (so, if \(n=2\), \(\mathrm P_2\) is made of a single edge, whereas if \(n=1\), \(\mathrm P_1\) is a single vertex.
### Posets {#subsubsection:posets}
Throughout, \(P :=(V, \leq)\) denotes an ordered set (poset), that is a set \(V\) equipped with a binary relation \(\leq\) on \(V\) which is reflexive, antisymmetric and transitive. We say that two elements \(x,y\in V\) are *comparable* if \(x\leq y\) or \(y\leq x\), otherwise, we say they are *incomparable*. The *dual* of \(P\) denoted \(P^{*}\) is the order defined on \(V\) as follows: if \(x,y\in V\), then \(x\leq y\) in \(P^{*}\) if and only if \(y\leq x\) in \(P\). According to Szpilrajn, every order \(\leq\) on a set \(V\) has a *linear extension*, that is a linear (or total) order \(\preceq\) on the \(V\) such that \(x\preceq y\) whenever \(x\leq y\), for all \(x,y\in V\). Let \(P:=(V,\leq)\) be a poset. A *realizer* of \(P\) is a family \(\mathcal{L}\) of linear extensions of the order of \(P\) whose intersection is the order of \(P\). Observe that the set of all linear extensions of \(P\) is a realizer of \(P\). The *dimension* of \(P\), denoted \(dim(P)\), is the least cardinal \(d\) for which there exists a realizer of cardinality \(d\). It follows from the Compactness Theorem of First Order Logic that an order is intersection of at most \(n\) linear orders (\(n\in {\mathbb N}\)) if and only if every finite restriction of the order has this property. Hence, the class of posets with dimension at most \(n\) is determined by a set of finite obstructions, each obstruction is a poset \(Q\) of dimension \(n+1\) such that the deletion of any vertex of \(Q\) leaves a poset of dimension \(n\); such a poset is said *critical*. For \(n\geq 2\) there are infinitely many critical posets of dimension \(n+1\). For \(n=2\), critical posets of dimension three (and hence finite comparability graphs of critical posets of dimension three) were characterized by Kelly. Beyond, the task is considered as hopeless.
### Comparability and incomparability graphs
The *comparability graph*, respectively the *incomparability graph*, of a poset \(P:=(V,\leq)\) is the graph, denoted by \(\comp(P)\), respectively \(\ainc(P)\), with vertex set \(V\) and edges the pairs \(\{u,v\}\) of comparable distinct vertices (that is, either \(u< v\) or \(v<u\)) respectively incomparable vertices. A graph \(G:= (V, E)\) is a *comparability graph* if the edge set is the set of comparabilities of some order on \(V\). From the Compactness Theorem of First Order Logic, it follows that a graph is a comparability graph if and only if every finite induced subgraph is a comparability graph. Hence, the class of comparability graphs is determined by a set of finite obstructions. The complete list of minimal obstructions was determined by Gallai (see for an English translation). The list can also be found in Figures 4(a) and 4(b).
### Permutation graphs
A graph \(G:= (V, E)\) is a *permutation graph* if there is a linear order \(\leq\) on \(V\) and a permutation \(\sigma\) of \(V\) such that the edges of \(G\) are the pairs \(\{x, y\}\in [V]^2\) which are reversed by \(\sigma\). Denoting by \(\leq_{\sigma}\) the set of oriented pairs \((x, y)\) such that \(\sigma(x) \leq \sigma (y)\), the graph is the comparability graph of the poset whose order is the intersection of \(\leq\) and the dual of \(\leq_{\sigma}\). Hence, a permutation graph is the comparability graph of an order intersection of two linear orders, that is the comparability graph of an order of dimension at most two. The converse holds if the graph is finite. As it is well known, a finite graph \(G\) is a permutation graph if and only if \(G\) and \(\overline G\) are comparability graphs; in particular, a finite graph is a permutation graph if and only if its complement is a permutation graph. The comparability graph of an infinite order which is intersection of two linear orders is not necessarily a permutation graph. A one way infinite path is a permutation graph, but the complement of this infinite path is not a permutation graph. There are examples of infinite posets which are intersection of two linear orders and whose comparability and incomparability graphs are not permutation graphs. For an example see Figure [\[fig:omega-z\]](#fig:omega-z){reference-type="ref" reference="fig:omega-z"}. However, via the Compactness Theorem of First Order Logic, an infinite graph is the comparability graph of a poset intersection of two linear orders if an only if each finite induced subgraph is a permutation graph (sometimes these graphs are called permutation graphs, while there is no possible permutation involved). For more about permutation graphs, see ,.
### Initial segment, ideal
An *initial segment* of a poset \(P:= (V, \leq)\) is any subset \(I\) of \(V\) such that \(x\in V\), \(y\in I\) and \(x\leq y\) imply \(x\in I\). An *ideal* is any nonempty initial segment \(J\) of \(P\) which is up-directed (that is \(x, y\in J\) implies \(x,y\leq z\) for some \(z\in J\)). If \(X\) is a subset of \(V\), the set \(\downarrow X:=\{y\in V: y\leq x \; \text{for some}\; x\in X\}\) is the least initial segment containing \(X\), we say that it is *generated* by \(X\). If \(X\) is a singleton, say \(X=\{x\}\), we denote by \(\downarrow x\), instead of \(\downarrow X\), this initial segment and say that it is *principal*. We denote by \(\mathbf {I}(P)\), resp. \(\mathbf {Id}(P)\), the set of initial segments, respectively ideals, of \(P\), ordered by set inclusion.
## Well-quasi-order {#section:w.q.o.-thm2-thm6}
We present the notion of well-quasi-order and introduce the notion of better-quasi-order; we refer to. A poset is *well-founded* if every nonempty subset has some minimal element. Such a poset has a *level decomposition* \((P_{\alpha})_{\alpha< h(P)}\) indexed by ordinal numbers. Level \(P_{\alpha}\) is the set of minimal elements of \(P\setminus \bigcup \{P_{\beta}: \beta< \alpha\}\) and \(h(P)\), the *height* of \(P\), is the least ordinal \(\alpha\) such that \(P_{\alpha}= \emptyset\). The poset is *level-finite* if each level \(P_{\alpha}\) is finite. A quasi-ordered-set (quoset) \(Q\) is *well-quasi-ordered* (w.q.o.), if every infinite sequence of elements of \(Q\) contains an infinite increasing subsequence. If \(Q\) is an ordered set, this amounts to say that every nonempty subset of \(Q\) contains finitely many minimal elements (this number being non zero). Equivalently, \(Q\) is w.q.o. if and only if it contains no infinite descending chain and no infinite antichain.
### Better-quasi-order {#subsubsection:bqo}
Proofs that some classes of countable structures are w.q.o. under embeddability may require a strengthening of that notion, e.g; the notion of *better-quasi-order* (b.q.o) (see Subsection [\[subsection:finitelymanyprimes\]](#subsection:finitelymanyprimes){reference-type="ref" reference="subsection:finitelymanyprimes"}). We just recall that b.q.o.'s are w.q.o.s. As for w.q.o.'s, finite sets and well-ordered sets are b.q.o.'s, finite unions, finite products, subsets and images of b.q.o.s by order preserving maps are b.q.o.'s. (see for more). Nash-Williams 1965 p.700, asserted that \"one is inclined to conjecture that most w.q.o. sets which arise in a reasonably 'natural' manner are likely to be b.q.o.\" It is not known if the answer is positive for hereditary classes of finite graphs. The first classes to consider are probably those which are minimal prime. Due to the description of these classes, the answer is positive. | {'timestamp': '2022-06-06T02:13:59', 'yymm': '2206', 'arxiv_id': '2206.01557', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01557'} |
# Introduction
We describe a real-world application that identifies semantically similar questions in English language business telephone calls between a customer and a sales or support agent. Identifying questions with similar meanings but different formulations provides insight into customer conversations that can be used for business analytics or for training agents on how to respond better to customer needs. Pre-trained contextualized embeddings such as those that can be extracted from BERT and GPT-2 have become essential building blocks for natural language processing (NLP) solutions. Embeddings extracted from these models can be used directly in unsupervised NLP applications, such as clustering and similarity comparison. Starting with pre-trained models reduces the need for expensive training and is particularly important in real-world settings, where high-quality data annotation is a major challenge. At our company we offer transcription of calls via automatic speech recognition and natural language understanding services for business conversations that enable product features such as sentiment analysis, identification of action items, and note-taking. Questions play a key role in natural conversation. It would be of great value to users if we could group questions asked in the calls that are semantically similar, giving our clients deeper insight in what types of questions their customers are asking. It is worth noting here that questions that arise naturally in the course of human-to-human conversation are quite different from the reading comprehension questions in public question answering datasets such as Stanford Question Answering Dataset (SQuAD), making this problem all the more challenging. We provide a detailed comparison in section 4.1. In this work, we demonstrate that the representations of question sentences only occupy a narrow cone in the output embedding space, a property known as *anisotropy*. This property makes it difficult to directly use the embeddings as input to semantic comparison functions. Then, we show how Sentence-BERT, a fine-tuned model based on BERT can overcome this problem and improve the performance in our applied task. Additionally, we propose an exemplar-based semantic matching approach to efficiently find sentences that are semantically similar to a specific topic.
# Background
## Pre-trained Contextualized Embeddings
There have been various pre-trained contextualized models based on deep neural networks in recent years. They include:
- BERT: one of the most popular large pre-trained models. Pooling over the last hidden state of BERT can be used as the sentence representation in many applications. Additionally, BERT prepends a *CLS* token (short for "classification") to the start of each input sentence, which is trained by the Next Sentence Prediction task to represent sentence-level classification.
- GPT-2: A large pre-trained language model that shows superior results in many NLP tasks. We use the pooling of the last hidden state as the representation of the sentence.
- Sentence-BERT (SBERT): A modification of pre-trained BERT that is fine-tuned on Natural Language Inference (NLI) and Semantic Textual Similarity (STS) datasets. SBERT achieves better performance than BERT in semantic textual similarity benchmarks.
## Industry Applications
At our company, we provide VoIP (Voice over Internet Protocol) telephony for businesses, with products for sales and support call centres. Calls are optionally recorded and transcribed using an in-house Automatic Speech Recognition (ASR) system. In addition to transcription, we offer NLP analysis of our in-house transcripts, e.g., identifying and extracting questions asked by customers in calls. We group extracted questions into semantically related sets. Questions within a set may vary in their syntactic structure but have similar or equivalent meaning. For instance, people can ask questions in various forms when inquiring about pricing, such as "What does it cost?", "How much is it?" or "What's the price of that?" Grouping related questions can help managers better coach their agents and prepare answers for commonly asked questions. Since hand-labelling our call transcripts would be expensive and time-consuming, we leveraged the rich semantic and syntactic information encoded in contextualized embeddings to build a semantic matching pipeline that requires no supervision or data labelling.
# Methods
## Sentence Representation
We experiment with several contextualized embedding models, using these models to construct sentence representations for our question dataset. For each model, we also test different methods for extracting embeddings as the sentence representations. The models and representations in our experiments include:
- BERT base: we compare using the *CLS* token versus average pooling of the last hidden layer, with an embedding size of 768, noted as "BERT-base-cls" and "BERT-base-mean" respectively.
- GPT-2: average pooling of last hidden state with a size of 768, referred to as "GPT2-mean".
- SBERT base fine-tuned on the NLI and STS datasets: we compare using the *CLS* token and average pooling of the last hidden layer with an embedding size of 768. These are referred to as "SBERT-base-cls" and "SBERT-base-mean" respectively.
## Exemplar-based Semantic Matching
We propose an efficient semantic matching formula that leverages the power of sentence representations from pre-trained models, and uses them directly to find semantically similar sentences for predefined semantic catgeories. Based on discussions with customers, we identified frequently occurring topics that provide business insight. For each predefined question topic, we manually identify a few exemplar sentences exhibiting a variety of lexical and syntactic formulations to use as the references for the semantic matching. Table 1 gives selected exemplars for two business topics, "Pricing" and "Contact Information". We compare two formulae for measuring semantic similarity given the sentence embeddings. The first formula is an unweighted similarity score of any query against the topic defined as: \[\label{eq:1} score = \frac{\sum_{i=1}^{N} \cos (q, s_i)}{N}\] \(N\) is the number of exemplar sentences, \(q\) denotes the embedding representation of the query sentence, while \(s_i\) is the embedding representation of the \(i^{th}\) exemplar sentence. Cosine similarity is used as the measurement of similarity as it is the most commonly used directional metric in vector space. We pair the query sentence with each exemplar question and compute the cosine similarity score, then average over the number of exemplars. We compare the unweighted score to a weighted similarity score defined as: \[\label{eq:2} score = \frac{\sum_{i=1}^{N} p_i}{N}\] where if \(\cos (q, s_i) < threshold\): \[p_i=\cos (q, s_i)\] else: \[p_i=w\cdot N\] When measuring a match between a given new question with exemplars, the unweighted score treats all exemplar questions equally, whereas the weighted score emphasizes matches to specific exemplars. The \(threshold\) is a manually chosen parameter based on the \(90^{th}\) percentile of the cosine similarity score distribution over all pairs in our semantic matching candidate pool. The parameter \(w\) is a manually selected constant (typically in the range \(w \epsilon [3, 5]\) in our application) that biases towards highly similar pairs.
# Experiment
## Dataset
Our data are customer support and sales call transcripts. A customer support call is typically an interaction in which a customer calls into a support centre with a problem or question relating to a particular product or service offered by the company, and then a customer service representative of the company tries to understand and solve the problem to the customer's satisfaction. A sales call is typically an interaction in which a sales representative calls a client or potential client to discuss a potential sale of a product, or subscription to a product, either to negotiate pricing or to discuss the client's product-specific needs (or some combination of both). From these transcripts we automatically identified and extracted customer-side questions. Note that these questions are part of a real, human conversation, and thus are mostly short in length, and use context-dependent devices such as pronouns. For instance, under the topic of "Pricing", typical questions include examples like "How much does it cost?" or "What's the price for monthly subscription?" In addition, as our transcripts are produced by our ASR engine directly from speech, some sentences may contain transcription errors. The questions that occur naturally in human conversation are very different from the questions found in public NLP data sets such as SQuAD. The SQuAD questions test reading comprehension. The questions can be interpreted without prior discourse context. Table 2 shows a comparison between SQuAD and our naturally occurring questions found in our data. The SQuAD questions are longer compared to our data. However, the percentage of stop words in our data indicates that questions asked in human conversations tend to use more stop words than those in SQuAD.
## Similarity Score Distribution
To better understand the distribution of sentence embeddings produced by each contextualized model and how they perform when directly used in similarity evaluation, we conducted the following experiments to test the similarity score distribution for each model: we sampled 1,000 question sentences from our dataset, and calculated the cosine similarity for all pairs of embeddings. Figure 1 shows the calculated pairwise question cosine similarity distributions for different contextualized models. The similarity scores from BERT and GPT-2 follow a narrow distribution, i.e. any two questions have high cosine similarity, whether they are semantically related or not. Use of the BERT *CLS* token or average pooling of the last hidden layer as the sentence representations does not greatly affect the pairwise similarity spread. In contrast, the SBERT scores exhibit a much wider distribution. Recall that SBERT was fine-tuned on natural language inference tasks, which is clearly beneficial for the task of determining the semantic similarity of questions. refers to the phenomenon of very strong cosine similarity between vector representations as *anisotropy*. The author shows that anisotropy is found in word-level representations across all layers in many contextualized models including BERT and GPT-2. Our experiment further shows that it also applies to sentence-level representations in BERT and GPT-2. Anisotropy can be problematic when using embeddings directly for sentence similarity comparison tasks, as all sentences are similar in some senses. As shown in Figure 2, we tested the pairwise cosine similarity of four different question sentences. Only \(t1\) and \(t2\) should have relatively high score as they are semantically close, but BERT and GPT-2 produced similar high scores for all pairs that cannot be distinguished easily. While SBERT successfully differentiated pairs according to their semantic relevance.
## Evaluation
### Selecting Exemplars
In this experiment, we seeded a question topic "Pricing" for the semantic matching. Following the procedures described in section 3.2, we selected ten questions related to "Pricing" as our exemplars. As a practical matter, ten exemplars is a reasonable number to ask a domain specialist to provide for a new topic. These exemplar questions are all semantically similar but differ in syntactic formulation, such as "How much does it cost?" and "What's the price for the monthly subscription?". We use embeddings from the testing models as sentence representations and conduct the semantic matching on 10,000 sampled question sentences from our dataset described in section 4.1.
### Metrics
Section 3.2 proposes two score computation methods for semantic matching, we can therefore rank the samples by the score using either approach. Since our dataset is unlabelled, our internal raters manually evaluated the precision at certain cutoff points after ranking: top 50, top 100 and top 200 samples with highest similarity score. This metric is noted as **Prec\@50** to describe the precision at top 50 samples in below experiments. Because we do not have labels for our data, we are not able to calculate recall. Instead we calculate the "hit rate" defined as: \[\label{eq:3} HR = \frac{number\;of\;predicted\;positive\;samples}{number\;of\;all\;samples}\] To compare the hit rate of different approaches, we set a target precision score and calculate the hit rate at that precision level. **HR\@Prec=0.9** denotes the hit rate of the level when the precision is 0.9
### Performance Comparison
In Table 3, we evaluate unweighted semantic matching using ([\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"}) for different embedding models. It is clear that SBERT, a BERT variant that was fine-tuned on inferences tasks, performs better at measuring the semantic similarity of sentences. Also, for BERT and SBERT, using mean as sentence representation outperforms *CLS* token in this task. However, sentence representations from GPT-2 performs poorly thus are not a good fit for semantic comparison. In discussions with business users, we have observed that when reading reports, false positives are highly distracting. We therefore require a precision score of 0.9 or higher. As Table 4 shows, the weighted score using ([\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"}) has slightly higher precision at all cutoff points, it also provides a 23% relative improvement in hit rate over the unweighted score. This results in a higher density of true positives in the analytics reports that collate the semantically similar questions.
# Early-Stage Deployment
We have implemented a data visualization dashboard (Figure 3) based on the best semantic matching approach described above (SBERT-mean with weighted similarity score). The dashboard groups semantically similar questions for topics of interest to managers such as "Pricing" and "Contact Information". The dashboard is currently available for users enrolled in an early adopter program. The dashboard helps the call center businesses collect and group questions asked by the customers. More importantly, it provides managers with information that they could use to coach agents (e.g. by drilling down to review the occurrence of specific questions in transcribed conversations) so that they can better prepare answers for the commonly asked questions in calls.
# Conclusion
We showed that when evaluating using a directional similarity metric, sentence-level representations produced by out-of-the-box contextualized embedding models such as BERT and GPT-2 are anisotropic, which makes them unsuitable for semantic similarity comparison tasks such as grouping semantically similar questions found in natural, human-to-human conversations. However, SBERT, which fine-tunes BERT model on NLI and STS datasets mitigates the anisotropy issue, thereby improving the performance of the semantic comparisons. Combining the best sentence embedding representation with the weighted scoring functions yields more precise groupings of semantically similar sets of the naturally-occurring questions in our dataset. The business solution that we have described only requires a few exemplar question sentences as seeds, to yield performance that meets our user needs. The solution does not require expensive data labelling to train a supervised classifier. We have implemented a visualization dashboard using this solution for our early enrolled users. | {'timestamp': '2022-06-06T02:14:53', 'yymm': '2206', 'arxiv_id': '2206.01585', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01585'} |
null | null |
Subsets and Splits