It is known that ECSs play an important role in the turbulent dynamics. The simplest solutions are fixed points that were calculated by Nagata (Reference Nagata1990<\/a>) and Gibson, Halcrow\u00a0& Cvitanovi\u0107 (Reference Gibson, Halcrow and Cvitanovi\u01072009<\/a>) in plane Couette flow. Travelling waves were discovered in pipe flow by Faisst\u00a0& Eckhardt (Reference Faisst and Eckhardt2003<\/a>) and Wedin\u00a0& Kerswell (Reference Wedin and Kerswell2004<\/a>), and in plane Couette flow by Gibson et\u00a0al. (Reference Gibson, Halcrow and Cvitanovi\u01072009<\/a>). Periodic orbits and relative periodic orbits found by Viswanath (Reference Viswanath2007<\/a>), Willis, Cvitanovi\u0107\u00a0& Avila (Reference Willis, Cvitanovi\u0107 and Avila2013<\/a>) and others have been demonstrated to be embedded in the turbulent state (Budanur et\u00a0al. Reference Budanur, Short, Farazmand, Willis and Cvitanovi\u01072017<\/a>). Furthermore, complex behaviour tends to develop through bifurcations of simpler ECSs. In particular, a period doubling cascade (Moore et\u00a0al. Reference Moore, Toomre, Knobloch and Weiss1983<\/a>), or the bifurcation of invariant tori (Ruelle\u00a0& Takens Reference Ruelle and Takens1971<\/a>; Gollub\u00a0& Swinney Reference Gollub and Swinney1975<\/a>) or intermittent behaviour (Pomeau\u00a0& Manneville Reference Pomeau and Manneville1980<\/a>) precedes the appearance of chaos (Eckmann Reference Eckmann1981<\/a>).<\/p>\n For these reasons, great effort has been directed towards identifying unstable ECSs in direct numerical simulations (DNS) (Kawahara\u00a0& Kida Reference Kawahara and Kida2001<\/a>; Graham\u00a0& Floryan Reference Graham and Floryan2021<\/a>; Page et\u00a0al. Reference Page, Norgaard, Brenner and Kerswell2024<\/a>) and in experiments (Suri et\u00a0al. Reference Suri, Pallantla, Schatz and Grigoriev2019<\/a>).<\/p>\n Developing predictive dynamical models is even more challenging. A frequently used, equation-driven method of model reduction employs Galerkin projection to a low-dimensional linear subspace. However, the optimal selection of modes spanning that subspace is often non-trivial, hence a large number of modes might be necessary to represent the turbulent statistics with sufficient accuracy (Cavalieri\u00a0& Nogueira Reference Cavalieri and Nogueira2022<\/a>). In addition, correspondence between trajectories of the full and reduced models is not guaranteed, since the low-dimensional subspace is generally not invariant.<\/p>\n Among data-driven approaches, linear methods such as dynamic mode decomposition (Schmid Reference Schmid2010<\/a>) are necessarily inapplicable due to the nonlinearisable nature of turbulent flows, as shown by Page\u00a0& Kerswell (Reference Page and Kerswell2019<\/a>). Recently, advances in machine learning have yielded a promising alternative. Specifically, the deep-learning-based DManD method (Linot\u00a0& Graham Reference Linot and Graham2020<\/a>) captures turbulent statistics and can also represent ECSs not enforced in their training (Linot\u00a0& Graham Reference Linot and Graham2023<\/a>). However, deep learning methods generally suffer from a need for large amounts of training data, a time-consuming training process, and an a priori unclear choice of hyperparameters.<\/p>\n In addition to neural ordinary differential equation (ODE) based methods such as DManD, alternative machine learning approaches have been developed to model the dynamics in a latent space. A notable example is reservoir computing, and echo state networks (ESNs), in particular. These networks are universal approximators (Jaeger Reference Jaeger2001<\/a>; Ahmed, Tennie\u00a0& Magri Reference Ahmed, Tennie and Magri2025<\/a>), and the associated optimisation problem is quadratic and requires no iterative training. Reservoir computing has shown success in modelling spatio-temporal chaotic systems (Pathak et\u00a0al. Reference Pathak, Lu, Hunt, Girvan and Ott2017<\/a>, Reference Pathak, Hunt, Girvan, Lu and Ott2018<\/a>). Recently, it has also been coupled with dimension reduction by Racca, Doan\u00a0& Magri (Reference Racca, Doan and Magri2023<\/a>), who introduced the convolutional autoencoder-echo state network (CAE-ESN). Their approach has successfully modelled Kolmogorov flow, the flow in a minimal channel, and other chaotic partial differential equations (PDEs), accurately reproducing statistical quantities and Lyapunov spectra (Margazoglou\u00a0& Magri Reference Margazoglou and Magri2023<\/a>; \u00d6zalp\u00a0& Magri Reference \u00d6zalp and Magri2025<\/a>).<\/p>\n In contrast, reducing a shear flow to a low-dimensional, attracting and structurally stable invariant manifold in its phase space offers a mathematically exact and robust construction of a reduced-order model. The recently introduced theory of spectral submanifolds (SSMs) (Haller Reference Haller2025<\/a>) targets low-dimensional attracting invariant manifolds emanating from stationary states, such as ECSs. Specifically, primary SSMs are defined as the smoothest invariant manifolds tangent to a selected spectral subspace of the linearised dynamics at the stationary state. Under mild non-resonance conditions, the existence and uniqueness of primary SSMs for attracting fixed points can be established (Haller Reference Haller2025<\/a>).<\/p>\n More recently, Haller et\u00a0al. (Reference Haller, Kasz\u00e1s, Liu and Ax\u00e5s2023<\/a>) pointed out the existence of entire families of (secondary) SSMs having a lower degree of smoothness (fractional SSMs) or perturbing from mixed-mode spectral subspaces containing both unstable and stable modes (mixed-mode SSMs). Mixed-mode SSMs, in particular, can represent unstable but recurrent dynamics, and have been shown to support chaotic dynamics (Liu, Ax\u00e5s\u00a0& Haller Reference Liu, Ax\u00e5s and Haller2024<\/a>; Xu et\u00a0al. Reference Xu, Kasz\u00e1s, Cenedese, Berti, Coletti and Haller2024<\/a>), and to aid the parametrisation of the basin boundary (Kasz\u00e1s\u00a0& Haller Reference Kasz\u00e1s and Haller2024<\/a>).<\/p>\n Apart from providing verifiable mathematical conditions for its applicability, SSM-based model reduction has the advantage over usual manifold learning approaches of providing the dimension of the underlying SSM a priori. This is in contrast to e.g. autoencoder-based methods that require careful optimisation of the latent dimension as a hyperparameter.<\/p>\n These features enable SSM-based model reduction to capture even chaotic attractors from data, as demonstrated by Liu et\u00a0al. (Reference Liu, Ax\u00e5s and Haller2024<\/a>). However, no data-driven reduced-order models have been constructed for fluid flows with coexisting non-trivial attractors. In this paper, we fill this gap in reduced flow modelling by constructing the slowest three-dimensional (3-D) mixed-mode SSM of the edge state in the plane Couette flow studied by Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>).<\/p>\n We show that this SSM-reduced model captures both the chaotic attractor and the laminar state of the flow with the SSM acting as an inertial manifold (Foias, Sell\u00a0& Temam Reference Foias, Sell and Temam1988<\/a>; Liu et\u00a0al. Reference Liu, Ax\u00e5s and Haller2024<\/a>). Our approach, therefore, extends the results of Kasz\u00e1s et\u00a0al. (Reference Kasz\u00e1s, Cenedese and Haller2022<\/a>), who obtained an SSM-reduced model for plane Couette flow with only periodic attractors, and also justifies the empirical low-dimensional model of Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>).<\/p>\n 2. Set-up<\/p>\n We focus on the chaotic dynamics observed in the plane Couette flow, which is the incompressible flow in a channel between two infinite plates moving in opposite directions with velocity <\/p>\n $\\pm U$<\/p>\n . The channel is defined as the domain <\/p>\n $\\varOmega = \\{(x,y,z) \\in \\mathbb{R}^3 : [0, L_x] \\times [-h, h] \\times [0, L_z] \\}$<\/p>\n , in which the velocity field <\/p>\n $\\boldsymbol{u} = [u, v, w](x,y,z,t)$<\/p>\n and the pressure <\/p>\n $p$<\/p>\n satisfy the Navier\u2013Stokes equations<\/p>\n (2.1)<\/p>\n \\begin{equation} \\frac {\\partial \\boldsymbol{u}}{\\partial t} + \\boldsymbol{u}\\boldsymbol{\\cdot }\\boldsymbol{\\nabla }\\boldsymbol{u} = -\\boldsymbol{\\nabla }p + \\nu\\, \\Delta \\boldsymbol{u}, \\quad \\boldsymbol{\\nabla }\\boldsymbol{\\cdot }\\boldsymbol{u} = 0, \\end{equation}<\/p>\n where <\/p>\n $\\nu$<\/p>\n is the kinematic viscosity. The main parameter of the problem is the Reynolds number, defined as <\/p>\n ${Re}=Uh\/\\nu$<\/p>\n , where <\/p>\n $2h$<\/p>\n is the distance between the moving walls. This also sets the relevant time unit as <\/p>\n $h\/U$<\/p>\n .<\/p>\n We work with a streamwise- and spanwise-periodic computational domain corresponding to the minimal flow unit studied by Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>) and Kreilos, Eckhardt\u00a0& Schneider (Reference Kreilos, Eckhardt and Schneider2014<\/a>), and fix <\/p>\n $h=1$<\/p>\n , <\/p>\n $L_x=2\\pi$<\/p>\n and <\/p>\n $L_z=\\pi$<\/p>\n . This domain is also comparable to those used by Nagata (Reference Nagata1990<\/a>), Page\u00a0& Kerswell (Reference Page and Kerswell2019<\/a>), Linot\u00a0& Graham (Reference Linot and Graham2023<\/a>) and Kasz\u00e1s et\u00a0al. (Reference Kasz\u00e1s, Cenedese and Haller2022<\/a>).<\/p>\n We simulate the flow using the open source Channelflow library (Gibson et\u00a0al. Reference Gibson2019<\/a>), which employs a pseudo-spectral discretisation using Fourier modes in the streamwise and spanwise directions, and Chebyshev modes in the wall-normal direction. For a direct comparison with earlier studies, we fix the number of streamwise and spanwise modes as <\/p>\n $N_x=N_z = 32$<\/p>\n , and the number of wall-normal modes as <\/p>\n $N_y=33$<\/p>\n . The boundary and incompressibility conditions are enforced using the influence-matrix method and tau correction (Kleiser\u00a0& Schumann Reference Kleiser and Schumann1980<\/a>). The time stepping algorithm implemented in Channelflow uses semi-implicit backward differentiation (Gibson Reference Gibson2014<\/a>; Gibson et\u00a0al. Reference Gibson2019<\/a>), and generates the forward time flow map of the discretised PDE (2.1<\/a>). This time evolution can be implicitly interpreted as a trajectory of a finite-dimensional dynamical system (Gibson, Halcrow\u00a0& Cvitanovi\u0107 Reference Gibson, Halcrow and Cvitanovi\u01072008<\/a>)<\/p>\n (2.2)<\/p>\n \\begin{equation} \\dot {\\boldsymbol{x}}=\\boldsymbol{f}(\\boldsymbol{x}), \\quad \\boldsymbol{x}\\in \\mathbb{R}^N, \\quad \\boldsymbol{f}\\in \\mathcal{C}^\\infty , \\end{equation}<\/p>\n with <\/p>\n $\\boldsymbol{x}$<\/p>\n denoting the collection of velocity values in physical space returned by Channelflow, and the total number of degrees of freedom is <\/p>\n $N\\approx O(10^5)$<\/p>\n . The laminar flow, which is a stable fixed point of (2.2<\/a>), is expressed as <\/p>\n $u=y,\\ v=w=0$<\/p>\n . In the following, we focus on the dynamical system (2.2<\/a>).<\/p>\n As shown by Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>) and Kreilos et\u00a0al. (Reference Kreilos, Eckhardt and Schneider2014<\/a>), in this computational cell, the Nagata upper and lower branch fixed points (Nagata Reference Nagata1990<\/a>; Gibson et\u00a0al. Reference Gibson, Halcrow and Cvitanovi\u01072009<\/a>) of (2.2<\/a>) appear in a saddle-node bifurcation at <\/p>\n ${Re}=163.8$<\/p>\n . For higher Reynolds numbers, the upper branch fixed point (UB) undergoes a Hopf bifurcation, followed by a period doubling cascade, eventually leading to chaotic dynamics.<\/p>\n The lower branch fixed point remains an edge state for a wide range of Reynolds numbers, and its codimension-one stable manifold forms the edge of chaos (Wang et\u00a0al. Reference Wang, Gibson and Waleffe2007<\/a>; Schneider et\u00a0al. Reference Schneider, Gibson, Lagha, De Lillo and Eckhardt2008<\/a>). The chaotic attractor, which can be traced back to the upper branch fixed point, undergoes a boundary crisis (Grebogi, Ott\u00a0& Yorke Reference Grebogi, Ott and Yorke1983<\/a>) and disappears at <\/p>\n ${Re}=188.51$<\/p>\n . After this point, the transient chaotic behaviour associated with turbulence is generated by a chaotic saddle (Lai\u00a0& T\u00e9l Reference Lai and T\u00e9l2011<\/a>) and hence has a finite lifetime (Kreilos et\u00a0al. Reference Kreilos, Eckhardt and Schneider2014<\/a>).<\/p>\n We focus on a Reynolds number value before the boundary crisis, <\/p>\n ${Re}= 187.8$<\/p>\n , where a genuine chaotic attractor coexists with the stable laminar state. Furthermore, we also restrict the flow to the symmetry invariant subspace of the Nagata equilibria by solving (2.1<\/a>) with the shift\u2013reflect symmetries imposed.<\/p>\n 2.1. Spectral submanifolds<\/p>\n We aim to construct a simple reduced model that faithfully represents the bistability of (2.2<\/a>). To this end, we seek a low-dimensional invariant manifold containing the edge state, labelled <\/p>\n $\\boldsymbol{x}_{\\textit{LB}}$<\/p>\n . The linearised dynamics around the edge state are governed by the Jacobian <\/p>\n $\\boldsymbol{Df}({\\boldsymbol{x}_{\\textit{LB}}})$<\/p>\n , whose eigenvalues are <\/p>\n $\\lambda _1,\\ldots ,\\lambda _N\\in \\mathbb{C}$<\/p>\n .<\/p>\n Let us assume that the eigenvalues satisfy the non-resonance conditions<\/p>\n (2.3)<\/p>\n \\begin{equation} \\sum _{j=1}^N m_{\\!j}\\lambda _{\\!j} \\neq \\lambda _i,\\quad m_{\\!j}\\in \\mathbb{N}, \\end{equation}<\/p>\n for <\/p>\n $i=1,\\ldots ,N$<\/p>\n and <\/p>\n $\\sum _{j=1}^Nm_{\\!j}\\geq 2$<\/p>\n . The conditions (2.3<\/a>) guarantee the applicability of the linearisation theorem of Sternberg (Reference Sternberg1958<\/a>) for class <\/p>\n $\\mathcal{C}^\\infty$<\/p>\n dynamical systems, such as (2.2<\/a>). This theorem states the existence of a smooth transformation mapping (2.2<\/a>) to its linearisation around the edge state.<\/p>\n As shown by Haller et\u00a0al. (Reference Haller, Kasz\u00e1s, Liu and Ax\u00e5s2023<\/a>), this implies the existence of a family of SSMs tangent to any given spectral subspace of the linearised dynamics. Precisely one of these SSMs, the primary SSM, is <\/p>\n $\\mathcal{C}^\\infty$<\/p>\n smooth, whereas all others have reduced differentiability. We target the slowest family of SSMs, tangent to the slowest spectral subspace spanned by eigenvectors of the linearisation with eigenvalues closest to the imaginary axis. This is a normally attracting slow manifold if the remaining eigenvalues of <\/p>\n $\\boldsymbol{Df}(\\boldsymbol{x}_{\\textit{LB}})$<\/p>\n have negative real parts.<\/p>\n The non-resonance conditions (2.3<\/a>) are violated only if a resonance occurs among both the real and imaginary parts of the eigenvalues, i.e. the frequencies and the decay rates, simultaneously. As discussed by Haller et\u00a0al. (Reference Haller, Kasz\u00e1s, Liu and Ax\u00e5s2023<\/a>), such an exact resonance among generic complex numbers is highly unlikely in a typical finite-dimensional system, such as (2.2<\/a>). In addition, 1 : 1 resonances appearing as repeated eigenvalues enforced by physical symmetries do not violate (2.3<\/a>). Therefore, we expect that a unique, <\/p>\n $\\mathcal{C}^\\infty$<\/p>\n , mixed-mode SSM of the edge state exists. Alternatively, we can also invoke the results of Buza (Reference Buza2024<\/a>), who showed that a <\/p>\n $\\mathcal{C}^1$<\/p>\n -smooth pseudo-unstable manifold exists for the Navier\u2013Stokes equations (2.1<\/a>).<\/p>\n Liu et\u00a0al. (Reference Liu, Ax\u00e5s and Haller2024<\/a>) and Xu et\u00a0al. (Reference Xu, Kasz\u00e1s, Cenedese, Berti, Coletti and Haller2024<\/a>) showed that slow mixed-mode SSMs tangent to both stable and unstable linear modes can contain the chaotic attractor of a dynamical system. Therefore, they often serve as inertial manifolds (Foias et\u00a0al. Reference Foias, Sell and Temam1988<\/a>).<\/p>\n Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>) constructed a one-dimensional (1-D) Poincar\u00e9 map by tracking the maxima of the kinetic energy signal on the chaotic attractor. Although such a construction only decreases the dimension of the system (2.2<\/a>) by one, this 1-D map already reveals periodic orbits in the attractor. These findings indicate that the attractor is likely low-dimensional.<\/p>\n 3. Results<\/p>\n \nFigure 1<\/a>(a,b) show the edge state with its spectrum. The unstable manifold is 1-D, and is tangent to the eigenvector <\/p>\n $\\boldsymbol{v}_1$<\/p>\n . The next slowest eigenvalues are stable and form a complex conjugate pair, with the eigenvectors <\/p>\n $\\boldsymbol{v}_2$<\/p>\n , <\/p>\n $\\boldsymbol{v}_3 = \\bar {\\boldsymbol{v}}_2$<\/p>\n . The linear span of the slowest eigenvectors, corresponding to an unstable real eigenvalue and a pair of stable complex eigenvalues, is the slowest mixed-mode spectral subspace <\/p>\n $E=\\text{span}(\\boldsymbol{v}_1, \\text{Re}\\ \\boldsymbol{v}_2, \\text{Im}\\ \\boldsymbol{v}_2)$<\/p>\n . These eigenvectors are shown in figure\u00a01<\/a>(c). The slow SSM, <\/p>\n $\\mathcal{W}(E)$<\/p>\n , is therefore also 3-D and is constructed as a graph over the subspace spanned by these eigenvectors.<\/p>\n We follow the SSMLearn algorithm of Cenedese et\u00a0al. (Reference Cenedese, Ax\u00e5s, B\u00e4uerlein, Avila and Haller2022<\/a>) to approximate the mixed-mode SSM, <\/p>\n $\\mathcal{W}(E)$<\/p>\n , from data. Our training trajectories are initialised near the edge state, and are first attracted to the slow SSM before converging to either the laminar state or the chaotic attractor. The DNS solver Channelflow also computes the eigenvectors <\/p>\n $\\boldsymbol{v}_{1,2,3}$<\/p>\n using the Arnoldi method, which we use to enforce the exact tangency between <\/p>\n $E$<\/p>\n and <\/p>\n $\\mathcal{W}(E)$<\/p>\n , i.e. we parametrise the SSM over the subspace spanned by <\/p>\n $\\boldsymbol{v}_{1}$<\/p>\n , <\/p>\n $\\text{Re}\\ \\boldsymbol{v}_2$<\/p>\n and <\/p>\n $\\text{Im}\\ \\boldsymbol{v}_2$<\/p>\n .<\/p>\n The coordinate chart returning the reduced coordinates <\/p>\n $\\boldsymbol{\\eta }$<\/p>\n is a projection to the eigenmodes<\/p>\n (3.1)<\/p>\n \\begin{equation} \\boldsymbol{\\eta }=\\boldsymbol{V}^{\\rm T}\\boldsymbol{x}, \\end{equation}<\/p>\n where <\/p>\n $\\boldsymbol{V}$<\/p>\n is the matrix containing <\/p>\n $\\boldsymbol{v}_{1}$<\/p>\n , <\/p>\n $\\text{Re}\\ \\boldsymbol{v}_2$<\/p>\n and <\/p>\n $\\text{Im}\\ \\boldsymbol{v}_2$<\/p>\n . In addition, the eigenvectors allow us to ensure that the training trajectories lie close to the SSM. Specifically, close to the fixed point, the SSM <\/p>\n $\\mathcal{W}(E)$<\/p>\n is well approximated by the spectral subspace <\/p>\n $E$<\/p>\n . We initialise the trajectories as small perturbations of the edge state along the spectral subspace, which results in initial velocity fields of the form<\/p>\n (3.2)<\/p>\n \\begin{equation} \\boldsymbol{x}=\\boldsymbol{x}_{\\textit{LB}}+\\alpha _1\\boldsymbol{v}_1 + \\alpha _2\\ \\text{Re}\\ \\boldsymbol{v}_2 + \\alpha _3\\ \\text{Im}\\ \\boldsymbol{v}_2, \\end{equation}<\/p>\n with small coefficients <\/p>\n $\\alpha _{1,2,3}$<\/p>\n . The sign of <\/p>\n $\\alpha _1$<\/p>\n decides on which side of the edge of chaos the initial condition lies. Since we prepare initial conditions to lie close to the SSM, and we consider that the tangent space of the SSM is known a priori, our approach is not strictly data-driven, but data-assisted, using the terminology of Cenedese et\u00a0al. (Reference Cenedese, Marconi, Haller and Jain2025<\/a>).<\/p>\n We advect 12 initial conditions (3.2<\/a>) using Channelflow up to time <\/p>\n $T_{max}=2000$<\/p>\n , sampling the trajectories at integer multiples of the dimensionless time unit <\/p>\n $\\Delta t=1$<\/p>\n . For comparison, we note that the characteristic time scale of the instability of the edge state is approximately <\/p>\n $1\/ \\lambda _1=21$<\/p>\n , as seen in figure\u00a01<\/a>(c). In total, this results in <\/p>\n $N_d=24\\,000$<\/p>\n data points in the training set. The kinetic energy <\/p>\n $E=({1}\/{2})\\,\\|\\boldsymbol{u}\\|^2$<\/p>\n of these trajectories, averaged over the domain <\/p>\n $\\varOmega$<\/p>\n , is shown in figure\u00a02<\/a>(a). The reduced coordinates of the training trajectories, given by (3.1<\/a>), are shown in figure\u00a02<\/a>(c). On the chaotic side of the edge, we see saddle-spiral type dynamics, reminiscent of a R\u00f6ssler-type attractor (R\u00f6ssler Reference R\u00f6ssler1976<\/a>), while on the other side, the dynamics are essentially 1-D, leading to rapid laminarisation.<\/p>\n Figure 2. (a) The average kinetic energy along training trajectories. (b) Correlation dimension estimation based on (3.3<\/a>), in the full phase space and the reduced phase space. The corresponding power-law fits are shown in blue and orange, respectively. (c) The reduced coordinates of the same trajectories as they converge to the chaotic attractor. The inset shows the laminarising trajectories as well. This figure is also available as a Jupyter notebook (https:\/\/www.cambridge.org\/S0022112025107957\/JFM-Notebooks\/files\/figure2\/figure2.ipynb<\/a>).<\/p>\n To find further evidence of the low-dimensionality of the chaotic attractor, we estimate its correlation dimension using the training trajectories. As defined by Grassberger\u00a0& Procaccia (Reference Grassberger and Procaccia1983<\/a>), the correlation dimension is the scaling exponent <\/p>\n $\\gamma$<\/p>\n in the relation<\/p>\n (3.3)<\/p>\n \\begin{equation} C(\\varepsilon ) \\sim \\varepsilon ^{\\gamma }\\quad \\text{as } \\varepsilon \\to 0, \\end{equation}<\/p>\n where <\/p>\n $C(\\varepsilon )$<\/p>\n is the correlation sum, i.e. the number of pairs of points in the attractor that are separated by a distance less than <\/p>\n $\\varepsilon \\gt 0$<\/p>\n .<\/p>\n We computed the correlation sum <\/p>\n $C(\\varepsilon )$<\/p>\n in the full, <\/p>\n $N$<\/p>\n -dimensional phase space and in the reduced, 3-D phase space. The results are shown in figure\u00a02<\/a>(b). A linear fit on a logarithmic scale returns a correlation dimension of approximately <\/p>\n $\\gamma =1.8$<\/p>\n in both cases. Therefore, the chaotic attractor is indeed low-dimensional, and no topological information is lost by restricting the dynamics to the SSM.<\/p>\n We note that in principle, the chaotic attractor might not be smoothly embedded in three dimensions. This would be manifested by self-intersections of the trajectories in the reduced space. In this case, however, we could embed it in an SSM of the same edge state that is higher than <\/p>\n $2\\gamma$<\/p>\n -dimensional, as required by the Whitney (Reference Whitney1944<\/a>) embedding theorem. The reduced trajectories in figure\u00a02<\/a>(c) show no self-intersection, therefore we restrict the dynamics to the lowest dimensional slow SSM, <\/p>\n $\\mathcal{W}(E)$<\/p>\n .<\/p>\n More generally, the SSM <\/p>\n $\\mathcal{W}(E)$<\/p>\n may develop a fold over its tangent space <\/p>\n $E$<\/p>\n , hence the parametrisation, constructed as a graph over <\/p>\n $E$<\/p>\n , may become singular. This would also be signalled by self-intersections of trajectories in the reduced space. Such examples are not common, but a notable one is discussed by Cenedese et\u00a0al. (Reference Cenedese, Ax\u00e5s, B\u00e4uerlein, Avila and Haller2022<\/a>), who showed that the two-dimensional (2-D) SSM in the flow behind a cylinder folded over its tangent space. As in that case, reparametrising the manifold over a different set of observables generally solves this issue.<\/p>\n 3.2. The reduced dynamics on the SSM<\/p>\n After identifying the geometry of the SSM, <\/p>\n $\\mathcal{W}(E)$<\/p>\n , we approximate the reduced dynamics. We model it as a discrete dynamical system given by the 3-D iterated map<\/p>\n (3.6)<\/p>\n \\begin{equation} \\boldsymbol{\\eta }_{n+1}=\\boldsymbol{F}(\\boldsymbol{\\eta }_n), \\quad \\boldsymbol{F}:\\mathbb{R}^3\\longrightarrow \\mathbb{R}^3, \\end{equation}<\/p>\n where <\/p>\n $\\boldsymbol{\\eta }_n$<\/p>\n is the reduced state at time step <\/p>\n $n$<\/p>\n .<\/p>\n Although the flow of (2.2<\/a>) is fundamentally continuous in time, we use a sampling time <\/p>\n $\\Delta t=1$<\/p>\n to obtain the reduced map (3.6<\/a>).<\/p>\n Figure 3. (a) Model predictions of test trajectories. (b) Power spectral densities computed from a chaotic kinetic energy signal in the full model and the SSM-reduced model. (c) A subset of the edge of chaos, constructed as the boundary of the basins of attraction in the reduced model, shown with the training trajectories. Trajectories initialised on either side of the edge of chaos are also indicated as black lines. This figure is also available as a Jupyter notebook (https:\/\/www.cambridge.org\/S0022112025107957\/JFM-Notebooks\/files\/figure3\/figure3.ipynb<\/a>).<\/p>\n A polynomial approximation of the reduced dynamics is insufficient for capturing the chaotic dynamics, as reported by Liu et\u00a0al. (Reference Liu, Ax\u00e5s and Haller2024<\/a>) and Xu et\u00a0al. (Reference Xu, Kasz\u00e1s, Cenedese, Berti, Coletti and Haller2024<\/a>). Instead, we use the SSMLearn algorithm with an alternative interpolation method. In particular, we approximate the dynamics as a linear combination of radial basis functions (RBFs) (Buhmann Reference Buhmann2003<\/a>). Interpolation using RBFs is a standard function approximation method, often used to interpolate unstructured data. In addition, computing the interpolants does not require iterative training.<\/p>\n The map (3.6<\/a>) is approximated in the form<\/p>\n (3.7)<\/p>\n \\begin{equation} \\boldsymbol{F}(\\boldsymbol{\\eta })=\\sum _{i=1}^{N_d}\\boldsymbol{c}_i k(\\|\\boldsymbol{\\eta }_i-\\boldsymbol{\\eta }\\|), \\end{equation}<\/p>\n where <\/p>\n $\\boldsymbol{\\eta }_i$<\/p>\n , <\/p>\n $i=1,\\ldots ,N_d$<\/p>\n , are the training data points, and <\/p>\n $k:\\mathbb{R}_{\\geq 0}\\longrightarrow \\mathbb{R}$<\/p>\n is a radial function, where <\/p>\n $\\mathbb{R}_{\\geq 0}$<\/p>\n denotes the non-negative reals. The coefficients <\/p>\n $\\boldsymbol{c}_i\\in \\mathbb{R}^3$<\/p>\n are determined by linear regression, minimising the squared error, analogously to (3.5<\/a>), between the training data and the RBF approximation. We use the multiquadric kernel <\/p>\n $k(r)=\\sqrt {r^2 + \\epsilon ^2}$<\/p>\n , which is a popular choice for RBF approximation (Fasshauer Reference Fasshauer2007<\/a>); it is also a default option in the implementation that we follow, available from the scipy library (Virtanen et\u00a0al. Reference Virtanen2020<\/a>). We select the scale parameter as <\/p>\n $\\epsilon =10^{-8}$<\/p>\n .<\/p>\n Since the RBF approximation interpolates the training trajectories, we need to use the reduced dynamics (3.7<\/a>) to predict the time evolution of previously unseen initial conditions, i.e. test trajectories, to show that we avoid overfitting. We show two such predictions in figure\u00a03<\/a>(a). The initial conditions are close to the edge state, but on opposite sides of the edge of chaos. The laminarising trajectory is accurately predicted over the whole time interval. As expected, predictions of the chaotic trajectory are accurate only on shorter time scales. This is due to the sensitivity of chaotic trajectories to initial conditions.<\/p>\n We can also compute relevant statistics on the chaotic attractor. In figure\u00a03<\/a>(b), we show the broad power spectral density obtained from the Fourier transform of the kinetic energy signal, characteristic of chaotic dynamics. A comparable time series of the full model yields a similar spectrum, with inaccuracies visible only towards high frequencies.<\/p>\n We further investigate the chaotic dynamics on the SSM by constructing the basins of the two coexisting attractors. We fill the domain of the reduced phase space shown in figure\u00a03<\/a>(c) with a total of <\/p>\n $10^6$<\/p>\n initial conditions, and iterate them forwards under the reduced dynamics (3.7<\/a>) for <\/p>\n $T = 1000$<\/p>\n steps. We then record the initial conditions of the laminarising trajectories. This yields a characteristic function of the basin of attraction of the laminar state, with the basin of the chaotic attractor obtained as its complement.<\/p>\n Instead of visualising the characteristic function directly in the reduced phase space, we construct the boundary between the two basins. We define this basin boundary as the level set of the characteristic function at value <\/p>\n $0.5$<\/p>\n . The largest connected component of this level set, approximating the stable manifold of the edge state, is shown in figure\u00a03<\/a>(c). This set is a 2-D intersection of the edge of chaos with the SSM <\/p>\n $\\mathcal{W}(E)$<\/p>\n .<\/p>\n We note, however, that since all our training trajectories have been initialised near the edge state, the basin boundary that we obtain is necessarily only a local approximation. Indeed, as shown by Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>), the basin of attraction of the chaotic attractor has a bubble-like shape, encircling the attractor. To accurately capture this in our reduced model, trajectories distributed along the entire basin boundary, also on both sides of the upper edge of chaos (Budanur et\u00a0al. Reference Budanur, Marensi, Willis and Hof2020<\/a>), would need to be included in the training. This would increase the size of the required training set, and since our aim is not to map out the global shape of the basin boundary, we do not pursue this feature further.<\/p>\n Figure 4. (a) Average rate of separation between nearby trajectories in the SSM-based model and in the DNS. (b) Relative reconstruction error of the autoencoders of the DManD models for various latent space dimensions <\/p>\n $d_h$<\/p>\n . (c,d) Average rates of separation of nearby trajectories in the best and worst DManD models, as defined in the text. This figure is also available as a Jupyter notebook (https:\/\/www.cambridge.org\/S0022112025107957\/JFM-Notebooks\/files\/figure4\/figure4.ipynb<\/a>).<\/p>\n The reduced dynamics (3.7<\/a>) can also be used to compute the Lyapunov exponents (Oseledets Reference Oseledets1968<\/a>) of the chaotic attractor. The Lyapunov exponents are defined as the average exponential growth rates of perturbations along the chaotic attractor. Here, we compute the leading exponent, i.e. the most positive Lyapunov exponent, by tracking the average growth rate of perturbations in an ensemble of trajectories initialised close to an arbitrary point on the attractor (Cvitanovi\u0107 et\u00a0al. Reference Cvitanovi\u0107, Artuso, Mainieri, Tanner and Vattay2016<\/a>). Denoting the distance between two initially close trajectories as <\/p>\n $\\delta (t)$<\/p>\n , the distance grows exponentially as <\/p>\n $\\delta (t)\\sim {\\rm e}^{\\varLambda t}$<\/p>\n , where <\/p>\n $\\varLambda$<\/p>\n is the leading Lyapunov exponent.<\/p>\n This computation can be carried out for the full system (Nastac et\u00a0al. Reference Nastac, Labahn, Magri and Ihme2017<\/a>), as well as for the SSM-reduced dynamics. Figure\u00a04<\/a>(a) shows the ensemble-averaged divergence of trajectories. By fitting a line to its initial, exponential trend, we find that the largest Lyapunov exponent is <\/p>\n $\\varLambda_{\\textit{SSM}}= 0.0150\\pm 0.0002$<\/p>\n . This overestimates the leading Lyapunov exponent of the full system, which is <\/p>\n $\\varLambda _{\\textit{DNS}}=0.0090 \\pm 0.0001$<\/p>\n . As a possible cause of this discrepancy, we note that our training set contains substantial data near the unstable edge state, whose eigenvalue <\/p>\n $\\lambda _1 = 0.047$<\/p>\n (see figure\u00a01<\/a> $\\varLambda _{\\textit{DNS}}$<\/p>\n . In principle, this bias could be reduced by including more data from trajectories on the attractor.<\/p>\n We have also employed the recent neural ODE-based data-driven modelling method of Linot\u00a0& Graham (Reference Linot and Graham2023<\/a>) with our training data. The DManD algorithm reduces the dynamics to an inertial manifold by an initial linear projection to 500 proper orthogonal decomposition (POD) modes, followed by the application of an autoencoder. This maps the data to a latent space of dimension <\/p>\n $d_h$<\/p>\n , parametrising the inertial manifold. The reduced dynamics is then modelled as a neural ODE trained on the dynamics of the latent variables. We used the code published by Linot\u00a0& Graham (Reference Linot and Graham2023<\/a>) to train DManD models with <\/p>\n $d_h = 3, \\ldots , 13$<\/p>\n . Due to the stochastic nature of the training algorithm, we repeated the training ten times for each latent dimension, as described by Linot\u00a0& Graham (Reference Linot and Graham2020<\/a>).<\/p>\n \nFigure\u00a04<\/a>(b) shows the autoencoder reconstruction error, which saturates at approximately <\/p>\n $d_h=5$<\/p>\n . Depending on the training process, there can be a considerable difference in the performance of models with otherwise identical hyperparameters. We compute the reconstruction error of all ten neural ODE models on a validation trajectory for each <\/p>\n $d_h$<\/p>\n . Based on this metric, we select the best and worst performing models.<\/p>\n To compare the DManD models to our SSM-based models, we first show <\/p>\n $d_h=3$<\/p>\n , which matches the dimension of our SSM. As figure\u00a04<\/a>(c) shows, many of the models that we trained produce non-chaotic behaviour, and the best model underestimates the Lyapunov exponent. On the other hand, models with a higher-dimensional latent space achieve a very good representation of the chaotic dynamics on the attractor. We show the estimated leading Lyapunov exponent for <\/p>\n $d_h=13$<\/p>\n in figure\u00a04<\/a>(d).<\/p>\n We also observed that even if they perform well on chaotic trajectories, higher-dimensional DManD models tend to mispredict trajectories converging to the laminar attractor. This suggests that the DManD models have difficulties representing the two coexisting attractors of our system, which is to be expected since the method was originally proposed to model systems with a single global chaotic attractor. Addressing these issues will require more training data to resolve the basin boundary accurately, or adjustments to the network architecture. We have included further details of the training and evaluation of the DManD models in the JFM Notebook accompanying figure\u00a04<\/a>.<\/p>\n 3.3. Poincar\u00e9 map<\/p>\n Once the reduced dynamics (3.7<\/a>) is obtained, we can reduce the model dimension even further by defining a Poincar\u00e9 map (Guckenheimer\u00a0& Holmes Reference Guckenheimer and Holmes1983<\/a>). Since the sampling time <\/p>\n $\\Delta t=1$<\/p>\n is small compared to other relevant time scales, we formally treat the trajectories of (3.6<\/a>) as if they were continuous in time by linearly interpolating between successive discrete time steps. This allows us to compute the Poincar\u00e9 map, which is defined for continuous-time dynamical systems.<\/p>\n The Poincar\u00e9 (or first-return) map <\/p>\n $\\boldsymbol{P}(\\boldsymbol{\\eta })$<\/p>\n is then defined by successive intersections of the trajectory with a plane in the reduced phase space. We select the plane to be <\/p>\n $\\eta _2=0$<\/p>\n , and record intersections with <\/p>\n $({\\eta }_2)_{n+1}\\gt 0\\gt ({\\eta }_2)_{n}$<\/p>\n . Figure\u00a04<\/a>(a) shows the intersection of the chaotic attractor with <\/p>\n $\\eta _2=0$<\/p>\n .<\/p>\n Using the reduced dynamics, the map <\/p>\n $\\boldsymbol{P}$<\/p>\n can be evaluated arbitrarily many times to obtain a high-resolution representation of the chaotic attractor. This is shown in figure\u00a04<\/a>(b), which features a total of <\/p>\n $10^6$<\/p>\n intersection points generated by <\/p>\n $\\boldsymbol{P}$<\/p>\n , revealing a much richer structure than the limited number of sample points obtained from the training trajectories. In particular, the SSM-reduced model revealed the small-scale fractal structure of the attractor in the inset of figure\u00a05<\/a>(b). Visually, the attractor resembles the attractor of the H\u00e9non map, as remarked also by Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>).<\/p>\n The Poincar\u00e9 section of the SSM, <\/p>\n $\\mathcal{W}(E)$<\/p>\n , denoted <\/p>\n $\\mathcal{W}(E)|_{\\eta _2=0}$<\/p>\n , can be computed by simply setting <\/p>\n $\\eta _2=0$<\/p>\n in the parametrisation (3.4<\/a>). In figure\u00a05<\/a>(c), we visualise the attractor of the Poincar\u00e9 map on the SSM, by computing the kinetic energy of the flow fields.<\/p>\n We can also use the low-dimensional Poincar\u00e9 map on the SSM to extract previously unobserved features of the chaotic dynamics. In particular, since the chaotic attractor is the closure of infinitely many unstable periodic orbits (Guckenheimer\u00a0& Holmes Reference Guckenheimer and Holmes1983<\/a>), we can use the Poincar\u00e9 map to find some of these underlying orbits.<\/p>\n Computing periodic orbits by approximating the Poincar\u00e9 map sampled using simulated trajectories has previously been applied to chaotic systems with low-dimensional attractors. Christiansen, Cvitanovic\u00a0& Putkaradze (Reference Christiansen, Cvitanovic and Putkaradze1997<\/a>) computed periodic orbits of the Kuramoto\u2013Shivashinsky equations (KSE) using a 1-D Poincar\u00e9 map. A similar approach was used by Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>), who defined a Poincar\u00e9 map by interpolation, using empirically computed successive maxima of the kinetic energy signal. More recently, Wang et\u00a0al. (Reference Wang, Ayats, Deguchi, Meseguer and Mellibovsky2025<\/a>) used the cylinder torques measured in Taylor\u2013Couette flow to define a 1-D map and prove its chaoticity. Also recently, Abadie et\u00a0al. (Reference Abadie, Beck, Parker and Schneider2025<\/a>) defined a Poincar\u00e9 map in a low-dimensional subspace spanned by the leading POD modes and in the latent space of an autoencoder to find periodic orbits of the KSE.<\/p>\n With our SSM-reduced model, the Poincar\u00e9 map is not approximated directly from full system simulations. Instead, it is a prediction of the reduced model, derived from the formally continuous time dynamics (3.6<\/a>). As figure\u00a05<\/a>(b) shows, the training data used to obtain the continuous time reduced model (3.6<\/a>) contain a much lower number of observed intersections with the Poincar\u00e9 section. Although they do not define a Poincar\u00e9 map, Linot\u00a0& Graham (Reference Linot and Graham2023<\/a>) also use the low-dimensional model to find periodic orbits of the full system.<\/p>\n Figure 5. (a) Black dots indicate the Poincar\u00e9 section of the chaotic attractor with the <\/p>\n $\\eta _2=0$<\/p>\n plane (grey). (b) Black dots indicate intersections computed based on the training trajectories. Red dots are iterations of the SSM-reduced Poincar\u00e9 map. (c) Poincar\u00e9 section of the SSM <\/p>\n $\\mathcal{W}(E)$<\/p>\n , containing the chaotic attractor (black) and periodic orbits (red). A period-3 orbit is highlighted in white, with the flow fields shown in (d\u2013f) using the same visualisation as in figure\u00a01<\/a>. The kinetic energy of this orbit is shown in (g). This figure is also available as a Jupyter notebook (https:\/\/www.cambridge.org\/S0022112025107957\/JFM-Notebooks\/files\/figure5\/figure5.ipynb<\/a>).<\/p>\n The Poincar\u00e9 map <\/p>\n $\\boldsymbol{P}$<\/p>\n bypasses the problem of finding an appropriate period, because periodic orbits appear as fixed points of iterates of <\/p>\n $\\boldsymbol{P}$<\/p>\n . We therefore search for solutions of the equation <\/p>\n $\\boldsymbol{P}^n(\\boldsymbol{\\eta })-\\boldsymbol{\\eta } = \\boldsymbol{0}$<\/p>\n , while systematically increasing <\/p>\n $n$<\/p>\n . We use a Newton\u2013Raphson-like method to find these roots.<\/p>\n We initialise the root-finding method with <\/p>\n $n=1,\\ldots ,5$<\/p>\n from random initial conditions distributed along the attractor. Upon successful identification of a periodic orbit, the initial condition is then provided as an initial guess to Channelflow\u2019s Newton\u2013Krylov solver (Viswanath Reference Viswanath2007<\/a>). Although we did not include any periodic orbits in the training data, we successfully found eight periodic ECSs of the full system using the initial guesses from the reduced model. One of these orbits is shown in figure 5<\/a>(d\u2013g).<\/p>\n We emphasise that our goal was not to compile an extensive library of periodic orbits of this particular system, but to demonstrate that an SSM-based reduced model already captures the essential chaotic dynamics. Indeed, the reduced dynamics enables the efficient sampling of the natural measure of the chaotic attractor to compute the probability distribution of quantities of interest, such as the spectrum of the kinetic energy in figure\u00a03<\/a>(b).<\/p>\n 4. Conclusions and discussion<\/p>\n We have applied the theory of spectral submanifolds (SSMs) to plane Couette flow in the parameter regime supporting a chaotic attractor. We restricted the flow to the slowest 3-D mixed-mode SSM of the edge state, which functions as an inertial manifold. Our very-low-dimensional reduced model is given by an iterated map interpolated using radial basis functions. Although contemporary machine learning methods, such as neural ODEs (Linot\u00a0& Graham Reference Linot and Graham2020<\/a>), could also be implemented to model the dynamics, we found that simpler interpolation methods already deliver excellent accuracy.<\/p>\n Compared to these machine learning methods, the classical function approximation used here has the advantage that it does not require any iterative training. This makes the models less data-hungry and faster to compute, and their performance is more predictable. This is due to the exact mathematical background supporting the existence and smoothness properties of the underlying SSMs. In addition, we found that the state-of-the-art neural-network-based model DManD (Linot\u00a0& Graham Reference Linot and Graham2023<\/a>) may have difficulties modelling the dynamics due to the coexisting chaotic and laminar attractors. In order for the neural ODE model to produce chaotic dynamics consistently, larger-dimensional inertial manifolds were required.<\/p>\n We have also defined a 2-D Poincar\u00e9 map associated with the SSM-reduced model to extract previously unseen characteristics of the dynamics. In particular, the Poincar\u00e9 map revealed the fractal structure of the chaotic attractor, and facilitated the search for periodic orbits. Although Kreilos\u00a0& Eckhardt (Reference Kreilos and Eckhardt2012<\/a>) inferred a similar 1-D map based on simulated time histories of the kinetic energy to find periodic orbits, our approach makes a rigorous connection between the dynamics in the full phase space and a 2-D Poincar\u00e9 section in the reduced phase space.<\/p>\n Our SSM-reduced model has been proven to generalise effectively and represent the chaotic attractor accurately. Therefore, we believe that constructing SSM models, attached to known ECSs, is a promising method for modelling other flows, too. In particular, they allow efficient sampling of the probability distributions of physical variables without relying on periodic orbit theory (Cvitanovi\u0107 et\u00a0al. Reference Cvitanovi\u0107, Artuso, Mainieri, Tanner and Vattay2016<\/a>).<\/p>\n Here, we have focused on plane Couette flow in a minimal flow unit at a relatively low Reynolds number. This allowed us to conclude the existence of an SSM serving as an inertial manifold, since a hyperbolic ECS close to the chaotic attractor, with an slow attracting mixed-mode SSM, was available. As a result, although the system exhibits temporal chaos, the observed spatial flow structures remain simple. Nevertheless, successful modelling of such simple, but physically relevant systems is an important step towards modelling flows with a higher degree of spatial complexity.<\/p>\n A logical future direction will be to pursue a similar model reduction approach in the same system at a higher Reynolds number, after the boundary crisis (Kreilos et\u00a0al. Reference Kreilos, Eckhardt and Schneider2014<\/a>). For example, another popular parameter choice is <\/p>\n $Re=400$<\/p>\n , where the lower branch fixed point persists with a 1-D unstable manifold. The flow is more complex at this Reynolds number, and the dimension of its chaotic set likely increases but stays below <\/p>\n $18$<\/p>\n , according to Linot\u00a0& Graham (Reference Linot and Graham2023<\/a>). The main difficulty in that case comes from the transient nature of the chaotic dynamics. Since this dynamics is supported on a chaotic saddle, trajectories may need to be carefully initialised near its stable manifold in order to explore its dynamics for sufficiently long times.<\/p>\n We finally note that systems exhibiting a supercritical transition to chaos and turbulence may also be reduced to a low-dimensional SSM. In that case, the model needs to be anchored to the unstable laminar state, which is often available in closed form. The appropriate slowest mixed-mode SSM of sufficiently high dimension can then be constructed in a way similar to that presented here. We expect an SSM-reduced model to be successful in those cases as well, at least for moderately chaotic flows. This is also supported by the recent results of Wang et\u00a0al. (Reference Wang, Ayats, Deguchi, Meseguer and Mellibovsky2025<\/a>), who constructed an empirical 1-D Poincar\u00e9 map for a low Reynolds number Taylor\u2013Couette flow and found unstable periodic orbits embedded in the attractor.<\/p>\n","protected":false},"excerpt":{"rendered":"1. Introduction Certain shear flows, such as pipe flow (Avila, Barkley\u00a0& Hof Reference Avila, Barkley and Hof2023) and…\n","protected":false},"author":2,"featured_media":291479,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[49],"tags":[199,79],"class_list":{"0":"post-291478","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-physics","8":"tag-physics","9":"tag-science"},"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/posts\/291478","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/comments?post=291478"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/posts\/291478\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/media\/291479"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/media?parent=291478"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/categories?post=291478"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/us\/wp-json\/wp\/v2\/tags?post=291478"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline2.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline3.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline4.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline5.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn1.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline6.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline7.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline8.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline9.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline10.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline11.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline12.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline13.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline14.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn2.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline15.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline16.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline17.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline18.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline19.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline20.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline21.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline22.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline23.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn3.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline24.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline25.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline26.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline27.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline28.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline29.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline30.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline31.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline32.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline33.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline34.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline35.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_fig1.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline44.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline45.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline46.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline47.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline48.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline49.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline50.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline51.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn4.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline52.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline53.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline54.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline55.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline56.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline57.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn5.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline58.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline59.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline60.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline61.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline62.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline63.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline64.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline65.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_fig2.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline66.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn6.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline67.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline68.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline69.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline70.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline71.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline72.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline73.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline74.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline75.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline76.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline86.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn9.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline87.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline88.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline89.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_fig3.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_eqn10.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline90.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline91.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline92.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline93.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline94.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline95.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline96.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline97.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline98.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline99.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline100.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_fig4.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline101.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline102.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline103.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline104.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline105.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline106.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline107.png\")
<\/p>\n
\nb) is considerably larger than <\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline108.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline109.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline110.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline111.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline112.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline113.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline114.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline115.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline116.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline117.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline118.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline119.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline120.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline121.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline122.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline123.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline124.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline125.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_fig5.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline126.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline127.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline128.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline129.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline130.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline131.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline132.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline133.png\")
<\/p>\n![](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/11\/urn:cambridge.org:id:binary:20251107144636768-0801:S0022112025107957:S0022112025107957_inline134.png\")
<\/p>\n