1 Introduction
Thermal convection in a gap between two parallel infinite walls maintained at different fixed temperatures, a system known as Rayleigh–Bénard convection, is a thermally driven non-equilibrium system that exhibits many different complex convection patterns (e.g. Cross & Greenside Reference Cross and Greenside2009). When inclining the walls against gravity, hot and cold fluid flows up and down the incline, respectively, creating a cubic laminar flow that breaks the isotropy of a horizontal layer and produces shear forces. This system is known as inclined layer convection (ILC). Inclined layer convection has three control parameters: the temperature difference between the walls, the Prandtl number
$Pr$ parametrising the diffusive properties of the fluid and the angle of inclination against gravity.
Recent experiments of ILC using compressed
$\text{CO}_{2}$ at a pressure of
$41.4$ bars and a mean temperature of
$27\,^{\circ }\text{C}$ yielding a Prandtl number of
$Pr=1.07$ have systematically varied the temperature difference and the inclination angle over a wide range, and report ten different spatio-temporal convection patterns (Daniels, Plapp & Bodenschatz Reference Daniels, Plapp and Bodenschatz2000). In these experiments, the flow domain has a lateral extent much larger than the gap height and thereby allows large-scale patterns to form. The observed convection patterns show spatio-temporally complex dynamics. This includes intermittent temporal bursting of spatially localized convection structures, observed both at small angles of inclination (Busse & Clever Reference Busse and Clever2000; Daniels et al. Reference Daniels, Plapp and Bodenschatz2000) as well as at large angles of inclination (Daniels, Wiener & Bodenschatz Reference Daniels, Wiener and Bodenschatz2003). Other examples include transient oblique patterns forming unsteady interfaces between spatial domains of differently oriented wavy roll patterns (Daniels & Bodenschatz Reference Daniels and Bodenschatz2002; Daniels et al. Reference Daniels, Brausch, Pesch and Bodenschatz2008), bimodal patterns, turbulent patterns like crawling rolls at intermediate inclinations and chaotically switching diamond panes (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000). These convection patterns have also been reproduced in direct numerical simulations of ILC (Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). How the large variety of patterns at different values of the control parameters emerges from the nonlinear equations describing the flow is however not completely understood.
Theoretical approaches towards explaining spatio-temporal convection patterns in ILC can be described as either an approach ‘close to thresholds’ or an approach ‘far above thresholds’. Approaches ‘close to thresholds’ include linear stability analysis and the construction of weakly nonlinear amplitude equations. At critical stability thresholds of the control parameters, flow states become unstable and give rise to new pattern motifs. Linear stability analysis of laminar ILC (Gershuni & Zhukhovitskii Reference Gershuni and Zhukhovitskii1969; Vest & Arpaci Reference Vest and Arpaci1969; Hart Reference Hart1971; Ruth, Hollands & Raithby Reference Ruth, Hollands and Raithby1980; Chen & Pearlstein Reference Chen and Pearlstein1989; Fujimura & Kelly Reference Fujimura and Kelly1993) identified two different types of primary instabilities. A buoyancy driven instability gives rise to straight convection rolls oriented along the base flow at small inclinations. A shear driven instability gives rise to straight convection rolls oriented transverse to the base flow at large inclinations. Secondary instabilities of finite amplitude straight convection rolls and subsequent tertiary instabilities at an increased temperature difference and certain angles of inclination have been investigated using Floquet analysis of two- and three-dimensional states (Clever & Busse Reference Clever and Busse1977; Busse & Clever Reference Busse and Clever1992; Clever & Busse Reference Clever and Busse1995; Busse & Clever Reference Busse, Clever and Riahi1996; Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). Such stability analysis can only explain the onset of convection patterns at, or very close to, the critical stability thresholds in control parameters.
Theoretical approaches to convection patterns ‘far above thresholds’ include the construction of finite amplitude states within a nonlinear analysis at control parameter values far above the critical stability thresholds. Finite amplitude states can be constructed by choosing a Galerkin projection for the governing equations of ILC, often motivated by pattern motifs and their symmetries as identified in a stability analysis at critical stability thresholds (Busse & Clever Reference Busse, Clever and Riahi1996; Golubitsky & Stewart Reference Golubitsky and Stewart2002). Galerkin approximations can then be evolved in time under the fully nonlinear governing equations until their amplitudes saturate at finite values with either steady or periodic time evolution (Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). Alternative to forward time integration, finite amplitudes of a Galerkin projection may also be calculated using a Newton–Raphson iteration giving access also to dynamically unstable finite amplitude states (Busse & Clever Reference Busse and Clever1992; Fujimura & Kelly Reference Fujimura and Kelly1993; Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). If Galerkin projections invoke a complete basis and fully resolve all spatial scales and modal interactions in the three-dimensional flow, exact finite amplitude states with steady or periodic time evolution can be found. These so-called invariant states are time-invariant exact solutions of the full nonlinear partial differential equations governing the flow. Depending on their temporal dynamics, invariant states are steady equilibrium states, travelling waves or periodic orbits, all of which capture particular structures in the flow. Invariant states can either be dynamically stable or dynamically unstable. In subcritical shear flows like pipe or Couette flows, the construction and analysis of unstable invariant states has lead to significant progress in understanding the complex dynamics of weakly turbulent flows by describing chaotic state space trajectories relative to invariant states (Kerswell (Reference Kerswell2005), Eckhardt et al. (Reference Eckhardt, Schneider, Hof and Westerweel2007), Kawahara, Uhlmann & van Veen (Reference Kawahara, Uhlmann and van Veen2012), and references therein).
In ILC only a few highly resolved three-dimensional invariant states had been constructed (Busse & Clever Reference Busse and Clever1992; Clever & Busse Reference Clever and Busse1995) before Reetz & Schneider (Reference Reetz and Schneider2020, referred to as Reference Reetz and SchneiderRS20 in the following) identified stable and unstable invariant states underlying various convection patterns at
$Pr=1.07$ observed in experiments (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000) and simulations (Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). These invariant states are found to transiently attract and repel the dynamics of ILC which are numerically simulated in minimal periodic domains. Minimal periodic domains accommodate only a single spatial period of a periodic convection pattern. Any invariant state computed in minimal periodic domains is also an invariant state in larger extended domains where the pattern of the state periodically repeats in space. To capture a specific pattern with an invariant state in a minimal periodic domain, the size of the domain must be chosen appropriately to match the wavelengths of the pattern. A suitable domain size for a specific pattern can be suggested by Floquet analysis, which determines the most unstable pattern wavelength of an instability. At the critical thresholds of instabilities, invariant states emerge in bifurcations and may continue as state branches far above critical thresholds. Thus, bifurcating state branches provide a connection between instabilities ‘at thresholds’ and invariant states ‘far above thresholds’.
In general, bifurcations are structural changes in a system’s state space across which the dynamics of the system changes qualitatively (Guckenheimer & Holmes Reference Guckenheimer and Holmes1983). Emerging stable invariant states that may continue ‘far above thresholds’ correspond to a supercritical, forward bifurcation leading to continuous changes in the dynamics. Subcritical bifurcations, however, create discontinuous changes in the dynamics allowing for sudden transitions from one state of the system to a very different state. Prominent and potentially harmful examples of such bifurcations, also called tipping points, have been identified in the earth’s climate system (e.g. Lenton et al. Reference Lenton, Held, Kriegler, Hall, Lucht, Rahmstorf and Joachim2008) or in combustion chambers (e.g Juniper & Sujith Reference Juniper and Sujith2018). In low-dimensional nonlinear model systems, like the three-dimensional Lorenz model for thermal convection (Lorenz Reference Lorenz1963), various types of bifurcations have been found and related to different routes to chaos (see Argyris, Faust & Haase (Reference Argyris, Faust and Haase1993), for a review). Thus, different types of bifurcations change the dynamics in different ways. Complex temporal dynamics may be observed where invariant states coexist at equal values of the control parameters (Reference Reetz and SchneiderRS20). Complex spatial dynamics, like spatial coexistence of different states in a non-conservative system such as ILC, suggests that the spatially coexisting states also coexist as individual states at equal values of the control parameters (Knobloch Reference Knobloch2015). Coexistence of invariant states at equal control parameter values is a consequence of bifurcations creating these invariant states. Thus, bifurcations creating invariant states that underlie observed convection patterns in ILC not only provide a parametric connection between invariant states and instabilities, but may also explain the state space structure underlying the spatio-temporally complex dynamics observed in spatially extended domains.
Computing bifurcation diagrams in nonlinear dynamical systems requires in practice to numerically continue branches of stable and unstable invariant states under changes in control parameters (see Dijkstra et al. (Reference Dijkstra, Wubs, Cliffe, Doedel, Hazel, Lucarini, Salinger, Phipps, Sanchez-Umbria and Schuttelaars2014), for a review). Numerically fully resolved invariant states in minimal periodic domains of ILC require a discretization between
${\sim}10^{4}$ and
${\sim}10^{6}$ degrees of freedom (Reference Reetz and SchneiderRS20), fewer than the earth’s climate system but much more than the Lorenz equations. Due to the numerically demanding size of the state space, not many prior studies have computed bifurcation diagrams in ILC. Using four degrees of freedom, Fujimura & Kelly (Reference Fujimura and Kelly1993) traced states of mixed longitudinal and transverse modes in almost vertical fluid layers. Using
${\sim}10^{3}$ degrees of freedom, Busse & Clever (Reference Busse and Clever1992) continued invariant states underlying three-dimensional wavy rolls at selected
$Pr$ and angles of inclinations, and Clever & Busse (Reference Clever and Busse1995) followed a sequence of supercritical bifurcations in vertical fluid layers. Bifurcation diagrams of two-dimensional invariant states have been computed in vertical convection (Mizushima & Tanaka Reference Mizushima and Tanaka2002a,Reference Mizushima and Tanakab) and horizontal convection (Waleffe, Boonkasame & Smith Reference Waleffe, Boonkasame and Smith2015), not addressing three-dimensional dynamics. Recent advances in matrix-free algorithms and computer hardware allow for the efficient construction and continuation of fully resolved three-dimensional invariant states in double-periodic domains with channel geometry (Viswanath Reference Viswanath2007; Gibson, Halcrow & Cvitanović Reference Gibson, Halcrow and Cvitanović2008). We use an extension to the existing numerical framework of the MPI-parallel code Channelflow 2.0 (Gibson et al. Reference Gibson, Reetz, Azimi, Ferraro, Kreilos, Schrobsdorff, Farano, Yesil, Schütz and Culpo2019) that also handles ILC (Reference Reetz and SchneiderRS20).
The aim of this paper is to systematically compute and describe bifurcations in ILC. These bifurcations explain the spatio-temporal complexity observed both experimentally and numerically. Using numerical continuation, we trace invariant states that have been constructed in Reference Reetz and SchneiderRS20 and that underlie the observed basic convection patterns at specific values of the control parameters. While Reference Reetz and SchneiderRS20 analyses dynamical connections between invariant states at those fixed specific values of the control parameters, the present article discusses bifurcations and connections between state branches when control parameters are varied. The analysis covers the same range of control parameters as recent experimental (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000) and theoretical work (Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016) at
$Pr=1.07$ and leads to an extensive network of bifurcating branches across values of the control parameters. To understand how temporal and spatio-temporal complexity arises in ILC, we specifically address the following three questions.
(Q1) Bifurcation types: Complex temporal dynamics between coexisting invariant states is a result of bifurcations creating the associated invariant states. Different bifurcation types change the dynamics in different ways. What types of bifurcations create invariant states underlying the observed convection patterns in ILC?
(Q2) Connection to instabilities: Floquet analysis characterizes instabilities at critical control parameter values. Results from such an analysis are valid close to the critical thresholds for small amplitude solutions. Do the fully nonlinear invariant states, found in Reference Reetz and SchneiderRS20 to underlie the observed convection patterns far from critical thresholds in ILC, bifurcate at the corresponding secondary instabilities reported from a Floquet analysis in Subramanian et al. (Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016)?
(Q3) Range of existence: Spatio-temporally complex dynamics suggests the existence of invariant states at the associated control parameter values. How do the bifurcation branches of invariant states in ILC continue across control parameter values and what are the limits of their existence?
The present article is structured in the following way. In § 2 we describe the numerical methods and outline the systematic bifurcation analysis. The results of the bifurcation analysis are stated in § 3. In five subsections, we report in detail on selected bifurcation diagrams explaining individual convection patterns. The results are discussed in response to Q1–Q3 in § 4.
2 Bifurcation analysis of invariant states
Before introducing the approach of the bifurcation analysis in § 2.3, we summarize the basic numerical concepts underlying direct numerical simulations of ILC (§ 2.1), and describe the invariant states that capture relevant convection patterns (§ 2.2). More details on the direct numerical simulations and identified invariant states are described elsewhere (Reference Reetz and SchneiderRS20).
2.1 Direct numerical simulation of inclined layer convection
Inclined layer convection is studied by numerically solving the non-dimensionalized Oberbeck–Boussinesq equations for the velocity
$\boldsymbol{U}$, temperature
${\mathcal{T}}$ and pressure
$p$ relative to the hydrostatic pressure
in numerical domains with
$x$,
$y$ and
$z$ indicating the streamwise, spanwise and wall-normal dimensions, respectively. The domains are bounded in
$z$ by two parallel walls at
$z=\pm 0.5$. In the streamwise dimension
$x$ and the spanwise dimension
$y$ periodic boundary conditions are imposed at
$x=[0,L_{x}]$ and
$y=[0,L_{y}]$, respectively. The walls are stationary with
$\boldsymbol{U}(z=\pm 0.5)=0$, have prescribed temperatures
${\mathcal{T}}(z=\pm 0.5)=\mp 0.5$, and are inclined against the gravitational unit vector
$\hat{\boldsymbol{g}}=-\text{sin}(\unicode[STIX]{x1D6FE})\boldsymbol{e}_{x}-\cos (\unicode[STIX]{x1D6FE})\boldsymbol{e}_{z}$ by inclination angle
$\unicode[STIX]{x1D6FE}$. With these boundary conditions, (2.1)–(2.3) admit the laminar solution
with arbitrary pressure constant
$\unicode[STIX]{x1D6F1}$. Equations (2.1)–(2.3) are non-dimensionalized by three characteristic scales of the system. We have chosen the temperature difference
$\unicode[STIX]{x0394}{\mathcal{T}}$ between the walls, the gap height
$H$ and the free-fall velocity
$U_{f}=(g\unicode[STIX]{x1D6FC}\unicode[STIX]{x0394}{\mathcal{T}}H)^{1/2}$ as characteristic scales. This non-dimensionalization defines the parameters
$\tilde{\unicode[STIX]{x1D708}}=(Pr/Ra)^{1/2}$ and
$\tilde{\unicode[STIX]{x1D705}}=(Pr\,Ra)^{-1/2}$ in terms of the Rayleigh number
$Ra=g\unicode[STIX]{x1D6FC}\unicode[STIX]{x0394}{\mathcal{T}}H^{3}/(\unicode[STIX]{x1D708}\unicode[STIX]{x1D705})$ and the Prandtl number
$Pr=\unicode[STIX]{x1D708}/\unicode[STIX]{x1D705}$. Here,
$\unicode[STIX]{x1D6FC}$ is the thermal expansion coefficient,
$\unicode[STIX]{x1D708}$ is the kinematic viscosity and
$\unicode[STIX]{x1D705}$ is thermal diffusivity. Thus, ILC has three control parameters,
$\unicode[STIX]{x1D6FE}$,
$Ra$ and
$Pr$, of which we fix
$Pr=1.07$, in line with previous studies using compressed
$\text{CO}_{2}$ as a working fluid (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000; Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016).
Time is measured in free-fall units
$H/U_{f}$ but will also be compared with other relevant time scales of ILC, like the heat diffusion time
$H^{2}/\unicode[STIX]{x1D705}$ and the laminar mean advection time
$L_{x}/\bar{U}_{0}$. The latter follows from the laminar velocity profile (2.4) integrated over the lower half of the domain where
$-0.5\leqslant z\leqslant 0$, that is,
$\bar{U}_{0}=\sin (\unicode[STIX]{x1D6FE})/384\tilde{\unicode[STIX]{x1D708}}$.
The pseudo-spectral code Channelflow 2.0 (Gibson et al. Reference Gibson, Reetz, Azimi, Ferraro, Kreilos, Schrobsdorff, Farano, Yesil, Schütz and Culpo2019) has been extended to solve (2.1)–(2.3) using Fourier–Chebychev–Fourier expansions with
$N$ spectral modes in space and a third-order implicit–explicit multistep algorithm to march forward in time (see Reference Reetz and SchneiderRS20; and references therein). Any time evolution computed with Channelflow-ILC represents a unique state vector trajectory
$\boldsymbol{x}(t)=[\boldsymbol{u},\unicode[STIX]{x1D703}](x,y,z,t)$ in a state space with
$N$ dimensions. This state space contains all solenoidal velocity fluctuations
$\boldsymbol{u}=\boldsymbol{U}-\boldsymbol{U}_{0}$ and temperature fluctuations
$\unicode[STIX]{x1D703}={\mathcal{T}}-{\mathcal{T}}_{0}$.
2.2 Computing invariant states
Invariant states are particular state vectors
$\boldsymbol{x}^{\ast }(t)$ representing roots of a recurrent map:
Here,
${\mathcal{F}}^{T}(\boldsymbol{x},\unicode[STIX]{x1D707})$ is the dynamical map integrating (2.1)–(2.3) from state
$\boldsymbol{x}$ over time period
$T$ at control parameter
$\unicode[STIX]{x1D707}\in [\unicode[STIX]{x1D6FE},Ra,Pr]$. The invariant state is either an equilibrium state if
$T$ is a free parameter, or a periodic orbit if
$T$ must match a specific period. Definition (2.7) includes a symmetry transformation
$\unicode[STIX]{x1D70E}\in S_{\mathit{ILC}}$. The symmetry group
$S_{\mathit{ILC}}=O(2)\times O(2)$, where ‘
$\times$’ is the direct product, is an equivariance of (2.1)–(2.3) in
$x$–
$y$-periodic domains. Here
$S_{\mathit{ILC}}$ is generated by a spanwise reflection
$\unicode[STIX]{x03C0}_{y}$, a combined reflection in the streamwise and the wall-normal direction
$\unicode[STIX]{x03C0}_{xz}$ and
$x$- and
$y$-translations
$\unicode[STIX]{x1D70F}(a_{x},a_{y})$ such that
with shift factors
$a_{x},a_{y}\in [0,1)$ scaling the spatial periods
$L_{x}$ and
$L_{y}$ of the periodic domain. All invariant states discussed here are invariant under transformations within subgroups of
$S_{\mathit{ILC}}=\langle \unicode[STIX]{x03C0}_{y},\unicode[STIX]{x03C0}_{xz},\unicode[STIX]{x1D70F}(a_{x},a_{y})\rangle$, where angle brackets
$\langle \rangle$ imply all products of elements given in the brackets. The specific coordinate transformations for reflection symmetries depend on the spatial phase of the flow structure relative to the origin. We choose the spatial phase such that three-dimensional inversion
$\unicode[STIX]{x03C0}_{xyz}=\unicode[STIX]{x03C0}_{y}\unicode[STIX]{x03C0}_{xz}$, where applicable to invariant states, applies with respect to the domain origin at
$(x,y,z)=(0,0,0)$.
If
$\unicode[STIX]{x1D70E}\neq 1$ in (2.7), the invariant state is a relative invariant state. A travelling wave state, where
$\unicode[STIX]{x1D70E}=\unicode[STIX]{x1D70F}(a_{x},a_{y})$ with specific shift factors
$a_{x}$ and
$a_{y}$, is a relative equilibrium state. A relative periodic orbit is either travelling, where
$\unicode[STIX]{x1D70E}=\unicode[STIX]{x1D70F}(a_{x},a_{y})$ must be applied after period
$T$, or is ‘pre-periodic’ if
$\unicode[STIX]{x1D70E}=1$ after a full period
$T$ but
$\unicode[STIX]{x1D70E}\neq 1$ after time interval
$T^{\prime }=T/n$ with
$n\in \mathbb{N}$.
Invariant states are computed by solving (2.7) with a Newton–Raphson iteration using matrix-free Krylov methods (Gibson et al. Reference Gibson, Reetz, Azimi, Ferraro, Kreilos, Schrobsdorff, Farano, Yesil, Schütz and Culpo2019). Practically, we stop iterations if
$\Vert {\mathcal{G}}(\boldsymbol{x},\unicode[STIX]{x1D707})\Vert _{2}<10^{-13}$ where
A residual of
${<}10^{-13}$ is sufficiently close to double machine precision to consider the iteration as fully converged.
Invariant states may be dynamically stable or unstable. The dynamical stability is characterized by the eigenvalues and eigenmodes of the linearized equations computed using Arnoldi iteration (Gibson et al. Reference Gibson, Reetz, Azimi, Ferraro, Kreilos, Schrobsdorff, Farano, Yesil, Schütz and Culpo2019), and depends on the specific symmetry subspace defined by size
$[L_{x},L_{y}]$ of the periodic domain and potentially imposed discrete symmetries
$\unicode[STIX]{x1D70E}\in S_{\mathit{ILC}}$. We impose
$\unicode[STIX]{x1D70E}$ on a state vector
$\boldsymbol{x}(t)$ using a projection
$(\boldsymbol{x}(t)+\unicode[STIX]{x1D70E}\boldsymbol{x}(t))/2$ which requires
$\unicode[STIX]{x1D70E}^{2}=1$. We will specify the considered symmetry subspace for each computation of the eigenvalue spectrum.
Previously, invariant states underlying observed convection patterns in ILC have been identified by combining direct numerical simulations in small periodic domains with Newton–Raphson iteration (Reference Reetz and SchneiderRS20). There, simulations from unstable laminar flow perturbed by small-amplitude noise lead to temporal transitions between seven invariant states. All of these seven invariant states are either linearly stable or posses few unstable eigendirections, depending on the symmetry subspace corresponding to the chosen periodic domain, control parameter values and potentially imposed discrete symmetries. As a consequence, the temporal dynamics is either asymptotically or transiently attracted to these invariant states. Moreover, the temporal dynamics is found to visit these invariant states in a specific sequential order. In figure 1 we summarize the observed transition sequences and visualize the flow structures of all seven invariant states. Which invariant state is visited by the dynamics depends on the values of the control parameters. Following Daniels et al. (Reference Daniels, Plapp and Bodenschatz2000) and Subramanian et al. (Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016), we explore the two-dimensional parameter space of inclination angle
$\unicode[STIX]{x1D6FE}$ and Rayleigh number
$Ra$ at fixed
$Pr=1.07$. When traversing this two-dimensional parameter space, the laminar base flow (
$B$) may lose stability in one of two different primary instabilities corresponding to two different transitions. At any Prandtl number, sufficiently increasing
$Ra$ yields a transition from
$B$ to either longitudinal rolls (
$LR$) for angles
$\unicode[STIX]{x1D6FE}<\unicode[STIX]{x1D6FE}_{c2}$, or from
$B$ to transverse rolls (
$TR$) for angles
$\unicode[STIX]{x1D6FE}>\unicode[STIX]{x1D6FE}_{c2}$. Only at a single point in this two-dimensional parameter space,
$[\unicode[STIX]{x1D6FE}_{c2},Ra_{c2}]$, do the two instabilities occur simultaneously. Such points are commonly referred to as codimension-2 points. At
$Pr=1.07$, we determine the codimension-2 point accurately at
$[\unicode[STIX]{x1D6FE}_{c2},Ra_{c2}]=[77.7567^{\circ },8053.1]$ when computing
$LR$ and
$TR$ in a domain with periodicity
$[L_{x},L_{y}]=[\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$, where
$[\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]=[2.2211,2.0162]$ and grid size
$[n_{x},n_{y},n_{z}]=[32,32,25]$. Wavelengths
$\unicode[STIX]{x1D706}_{x}$ and
$\unicode[STIX]{x1D706}_{y}$ are suggested by linear stability analysis (Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). As in Reference Reetz and SchneiderRS20, we fix the wavelengths
$[\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$ and the grid resolution
$[n_{x},n_{y},n_{z}]$ throughout this study and choose all computational domains as multiples of this minimal box. Subharmonic standing waves (
$SSW$) are computed in a domain with periodicity
$[L_{x},L_{y}]=[2\unicode[STIX]{x1D706}_{x},2\unicode[STIX]{x1D706}_{y}]$, wavy rolls (
$WR$) with
$[L_{x},L_{y}]=[2\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$, knots (
$KN$) with
$[L_{x},L_{y}]=[\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$ and transverse oscillations (
$TO$) with
$[L_{x},L_{y}]=[12\unicode[STIX]{x1D706}_{x},6\unicode[STIX]{x1D706}_{y}]$. The grid size is scaled accordingly. Choosing all domains as integer multiples of the same minimal box ensures commensurable wavelengths and thus allows for potential bifurcations between invariant states.
Invariant states found as sequentially visited states in temporal transitions at selected control parameter values (Reference Reetz and SchneiderRS20). Black arrows indicate the direction of temporal transitions, starting from laminar base flow
$B$, via longitudinal and transverse rolls (equilibria
$LR$ and
$TR$), to four different tertiary invariant states representing the convection patterns of subharmonic standing waves (periodic orbit
$SSW$), wavy rolls (equilibrium
$WR$), knots (equilibrium
$KN$) and transverse oscillations (periodic orbit
$TO$). While
$B$ is shown as the vector fields of total velocity
$\boldsymbol{U}$ at the box sides and three-dimensional contours of total temperature
${\mathcal{T}}$, all other states are shown in terms of velocity and temperature fluctuations,
$\boldsymbol{u}$ and
$\unicode[STIX]{x1D703}$, around
$B$. Iso-contours of temperature are at
$1/3[\min (\unicode[STIX]{x1D703}),\max (\unicode[STIX]{x1D703})]$. The states are computed at different inclinations of the domain (green gravity vectors):
$B$,
$LR$ and
$TR$ at
$\unicode[STIX]{x1D6FE}_{c2}$, and the four tertiary states at the control parameters marked in figure 2.

The approach of combining direct numerical simulations from an unstable laminar flow with Newton–Raphson iteration allows us to determine all of the above invariant states. However, this approach fails in the case of the pattern emerging from the skewed varicose instability,
$SV_{i}$, at
$\unicode[STIX]{x1D6FE}=0^{\circ }$ (Reference Reetz and SchneiderRS20). There, the dynamics does not asymptotically approach or transiently visit an invariant state underlying the pattern, suggesting that no associated invariant state exists above thresholds. Therefore, we search for the bifurcating branch below critical parameters of the skewed varicose instability by taking the following steps. The bifurcating eigenmode that destabilizes
$x$-aligned straight convection rolls at wavelength
$\unicode[STIX]{x1D706}_{y}$ in a domain of periodicity
$[L_{x},L_{y}]=[4\unicode[STIX]{x1D706}_{x},4\unicode[STIX]{x1D706}_{y}]$ is computed using Arnoldi iteration. Since ILC for
$\unicode[STIX]{x1D6FE}=0^{\circ }$, corresponding to the Rayleigh–Bénard system, has isotropic symmetry, there is no distinction between ‘longitudinal’ and ‘transverse’ rolls. Consequently, we indicate straight convection rolls by
$R_{\unicode[STIX]{x1D706}2}$, where the subscript indicates their approximate wavelength. Different finite amplitude perturbations of
$R_{\unicode[STIX]{x1D706}2}$ with the bifurcating eigenmode are integrated forward in time to generate a large set of initial states for brute-force Newton–Raphson iterations below critical threshold parameters of the instability. Using this approach we identified an unstable equilibrium state that underlies the skewed varicose pattern and is described in § 3.1. Consequently, invariant states in thermal convection cannot be assumed to generically exist above critical control parameter values, but may also be found below thresholds suggesting a backward bifurcation. Whether bifurcations are forward or backward in control parameters, is studied in the present bifurcation analysis.
2.3 Bifurcation analysis
The general approach of our bifurcation analysis is to compute bifurcation branches of invariant states in ILC and to characterize the resulting bifurcation diagrams. Branches of invariant states are computed using continuation methods to solve (2.7) under a changing value of a control parameter
$\unicode[STIX]{x1D707}$ (Dijkstra et al. Reference Dijkstra, Wubs, Cliffe, Doedel, Hazel, Lucarini, Salinger, Phipps, Sanchez-Umbria and Schuttelaars2014). There are two iterative schemes for numerical continuation implemented in Channelflow-ILC. The control parameter continuation uses quadratic extrapolation to predict a state vector
$\boldsymbol{x}$ for a specific value of
$\unicode[STIX]{x1D707}$ which is fixed in the following Newton–Raphson iteration. The pseudo-arclength continuation does not fix
$\unicode[STIX]{x1D707}$ but solves for
$\unicode[STIX]{x1D707}$ as an additional unknown entry in state vector
$\boldsymbol{x}$ under an additional arclength constraint. Depending on the shape of the continued state branch, one continuation scheme might outperform the other (Gibson et al. Reference Gibson, Reetz, Azimi, Ferraro, Kreilos, Schrobsdorff, Farano, Yesil, Schütz and Culpo2019). Continuation of periodic orbits with long periods may require a multi-shooting method to converge (Gibson et al. Reference Gibson, Reetz, Azimi, Ferraro, Kreilos, Schrobsdorff, Farano, Yesil, Schütz and Culpo2019). Where invariant states have discrete reflection symmetries
$\unicode[STIX]{x03C0}_{y}$ or
$\unicode[STIX]{x03C0}_{xz}$ (2.8)–(2.9), we impose reflections during numerical continuation because they fix the spatial phase of the flow relative to the
$x$- or
$y$-coordinates. If the spatial phase is free, states may translate under numerical continuation reducing the computational efficiency. Since both continuation schemes solve (2.7) and the algorithmic details do not change the resulting bifurcation diagrams, we use the better performing scheme for each branch.
Continuations of the invariant states cover a priori chosen sections across the considered parameter space at
$Pr=1.07$ covering
$0^{\circ }\leqslant \unicode[STIX]{x1D6FE}<120^{\circ }$ and
$0\leqslant \unicode[STIX]{x1D716}\leqslant 2$, as illustrated by thin grey lines in figure 2. The control parameter
$\unicode[STIX]{x1D716}=(Ra-Ra_{c}(\unicode[STIX]{x1D6FE}))/Ra_{c}(\unicode[STIX]{x1D6FE})$ indicates
$Ra$ normalized by a critical threshold function
$Ra_{c}(\unicode[STIX]{x1D6FE})$, which here approximates the true critical control parameters
$Ra_{c}^{\prime }(\unicode[STIX]{x1D6FE})$ of the primary instability in ILC (see figure 2a in Reference Reetz and SchneiderRS20). Thus, the primary instability defining the onset of convection is always at
$\unicode[STIX]{x1D716}\approx 0$, independent of the inclination angle. Critical thresholds of bifurcation points are denoted as
$\unicode[STIX]{x1D716}_{c}$. To continue invariant states in
$\unicode[STIX]{x1D6FE}$ at
$\unicode[STIX]{x1D716}=\text{const.}$, also
$Ra$ needs to be adjusted accounting for variations in
$Ra_{c}(\unicode[STIX]{x1D6FE})$. Since the true critical
$Ra_{c}^{\prime }(\unicode[STIX]{x1D6FE})$ cannot be expressed in closed-form, we define the function
to keep
$\unicode[STIX]{x1D716}=(Ra-Ra_{c}(\unicode[STIX]{x1D6FE}))/Ra_{c}(\unicode[STIX]{x1D6FE})=\text{const.}$ under
$\unicode[STIX]{x1D6FE}$ continuations. The definition of function
$Ra_{c}(\unicode[STIX]{x1D6FE})$ has three precise coefficients, namely the critical parameters for horizontal convection
$Ra_{c}^{\prime }(\unicode[STIX]{x1D6FE}=0^{\circ })=1707.76$ (Busse Reference Busse1978) and the codimension-2 point
$[\unicode[STIX]{x1D6FE}_{c2},Ra_{c2}^{\prime }]=[77.7567^{\circ },8053.1]$. Relation (2.12), already found by Gershuni & Zhukhovitskii (Reference Gershuni and Zhukhovitskii1969), is a geometric consequence of the linear laminar temperature profile. Polynomial (2.13) is a least-square-fit of the empirical critical thresholds for
$\unicode[STIX]{x1D6FE}_{c2}<\unicode[STIX]{x1D6FE}<120^{\circ }$ reported in Subramanian et al. (Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016), and is an approximation of the true
$Ra_{c}^{\prime }(\unicode[STIX]{x1D6FE})\approx Ra_{c}(\unicode[STIX]{x1D6FE})$. The purpose of defining
$Ra_{c}(\unicode[STIX]{x1D6FE})$ in (2.12)–(2.13) is not to most accurately capture the true
$Ra_{c}^{\prime }(\unicode[STIX]{x1D6FE})$, but to provide a closed-form function for converting values between
$Ra$ and
$\unicode[STIX]{x1D716}$. The conversion allows
$\unicode[STIX]{x1D6FE}$ continuations at
$\unicode[STIX]{x1D716}=\text{const.}$ and a comparison of the present results with other work reported in terms of a similar
$\unicode[STIX]{x1D716}^{\prime }$, based on the empirically determined primary instabilities.
The considered parameter space is spanned by inclination angle
$\unicode[STIX]{x1D6FE}$ and
$\unicode[STIX]{x1D716}=(Ra-Ra_{c})/Ra_{c}$ at
$Pr=1.07$. The parameters at which invariant states have been identified in Reference Reetz and SchneiderRS20 (figure 1) are marked by circles. The invariant state underlying the skewed varicose pattern (
${\mathcal{S}}{\mathcal{V}}$) emerging from the skewed varicose instability (
$SV_{i}$) is not described in Reference Reetz and SchneiderRS20 but identified in this work. Thick solid lines indicate critical thresholds in control parameters of the five secondary instabilities determined by Subramanian et al. (Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). Bifurcations between invariant states are computed along 15 sections across the parameter space (grey solid lines). The thin grey dotted line marks the inclination angle of the codimension-2 point at
$\unicode[STIX]{x1D6FE}_{c2}=77.7567^{\circ }$.

Linear stability analysis of invariant states is performed at selected points along continued branches. Under continuation, we consider invariant states in their minimal periodic domain capturing only one spatial period of the pattern. In order to compare the dynamical stability between different invariant states, Arnoldi iterations must be performed in identical symmetry subspaces. This implies using the same periodic domains and imposing the same discrete symmetries for all considered states. Wherever we compute the dynamical stability along selected bifurcation branches, we specifically choose and report the symmetry subspace for the full branch.
Summary of all bifurcation branches of invariant states continued in
$\unicode[STIX]{x1D716}$ at constant
$\unicode[STIX]{x1D6FE}\in \{0^{\circ },10^{\circ },20^{\circ },\ldots ,110^{\circ }\}$. Each branch is plotted in terms of
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ (3.1) and is labelled by the name of the invariant state to which the branch belongs. Insets enlarge or isolate features of the bifurcations diagrams. All panels are labelled by the angle of inclination and share the same axes. To avoid clutter
$TR$ is left out in panels
$\unicode[STIX]{x1D6FE}\in \{0^{\circ },10^{\circ },20^{\circ }\}$. In panel
$\unicode[STIX]{x1D6FE}=90^{\circ }$, the grey vertical line crosses the bifurcation branches where invariant states are shown in figure 15 and discussed in appendix B. The
$KN$ state at
$\unicode[STIX]{x1D6FE}=90^{\circ }$ connects to equilibrium state
$SL$ showing a subharmonic lambda pattern (figure 14), discussed in appendix A.

Summary of all bifurcation branches of invariant states continued in
$\unicode[STIX]{x1D6FE}$ at constant
$\unicode[STIX]{x1D716}\in \{0.1,0.5,1.5\}$. Each branch is plotted in terms of
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ (3.1) and is labelled by the name of the invariant state to which the branch belongs. All panels are labelled by the
$\unicode[STIX]{x1D716}$ value and share the same axes. Grey dotted lines mark the inclination angle of the codimension-2 point. Insets enlarge features of the bifurcations diagrams around the inclination angle of the codimension-2 point. In panel
$\unicode[STIX]{x1D716}=1.5$, the grey vertical line crosses the bifurcation branches where invariant states are shown in figure 15 and discussed in appendix B. The
$TO$ at
$\unicode[STIX]{x1D716}=0.1$ (inset) connect to equilibrium state
$TSV$ showing a transverse subharmonic varicose pattern (figure 14), discussed in appendix A.

Many bifurcation types of vector fields are known (e.g. Guckenheimer & Holmes Reference Guckenheimer and Holmes1983). The most common bifurcations we encounter in ILC are pitchfork bifurcations, Hopf bifurcations, saddle-node bifurcation and mutual annihilation of two periodic orbits, all of which are also well-known bifurcations in ordinary differential equations (e.g. Schaeffer & Cain Reference Schaeffer and Cain2016). The two latter types we simply refer to as ‘folds’. If bifurcations are not one of these four common types, we provide explicit references that discuss the bifurcation type in detail, as such discussions would be beyond the scope of the present work. When discussing symmetry-breaking bifurcations, the classification into supercritical/subcritical bifurcations refers to a ‘more stable’/‘less stable’ bifurcating branch in comparison to the stability of the coexisting parent branch (Tuckerman & Barkley Reference Tuckerman and Barkley1990). The orientation of symmetry-breaking bifurcations along a control parameter
$\unicode[STIX]{x1D707}$ is given specifically as
$\unicode[STIX]{x1D707}$-forward or
$\unicode[STIX]{x1D707}$-backward.
3 Results
We first provide an overview of the results from the bifurcation analysis. Considering twelve sections at constant
$\unicode[STIX]{x1D6FE}$ and three sections at
$\unicode[STIX]{x1D716}=\text{const.}$ (figure 2), we present 15 bifurcation diagrams here in figures 3 and 4. In addition to branches of the seven invariant states shown in figure 1, the bifurcation diagrams also contain branches of nine additional invariant states. We refer to these states as subharmonic travelling wave (
$STW$), skewed varicose state (
$SV$), ribbons (
$RB$), oblique rolls (
$OR$), oblique wavy rolls (
$OWR$), disconnected wavy rolls (
$DWR$), longitudinal subharmonic varicose state (
$LSV$), transverse subharmonic varicose state (
$TSV$) and subharmonic lambda plumes (
$SL$). For three out of these nine additional states their existence was suggested previously. The existence of
$SV$ and
$OR$ has been suggested by stability analysis (e.g. Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016) and
$TSV$ is similar, yet not identical, to an equilibrium discussed by Clever & Busse (Reference Clever and Busse1995). In table 1 we provide an index of all invariant states considered here. A complete systematic analysis of all branches in these diagrams is beyond the scope of this paper. Instead in this section, we first summarize the bifurcation diagrams and then focus on selected state branches covering the control parameter values where spatio-temporally complex convection patterns are observed and temporal dynamics between invariant states has been studied (Reference Reetz and SchneiderRS20).
Invariant states considered in the bifurcation analysis. The states are of one of three types: steady equilibrium states (EQ), travelling wave states (TW), pre-periodic orbits (PPO). Their spatial periodicity along the
$x$ and
$y$ dimension is given in units of
$\unicode[STIX]{x1D706}_{x}=2.2211$ and
$\unicode[STIX]{x1D706}_{y}=2.0162$.

We specifically discuss the branches that bifurcate from straight convection rolls via the five secondary instabilities that were identified by Subramanian et al. (Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). These are the skewed-varicose instability, longitudinal subharmonic oscillatory instability, wavy roll instability, knot instability and transverse oscillatory instability. Branches of equilibrium and travelling wave states are plotted in terms of the norm of the temperature fluctuations,
as a function of the bifurcation parameter. Periodic orbits are illustrated by a pair of branches indicating the minimum and maximum of
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ over one orbit period, at instances
$t_{\pm }$. Bifurcation branches are labelled inside the diagram with the name of the invariant state. We recommend reading each diagram panel by first identifying the branches of
$LR$ and/or
$TR$. In most cases,
$LR$ or
$TR$ have the largest
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ and tertiary branches bifurcate to lower
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$. See figure 3 for bifurcations while varying
$\unicode[STIX]{x1D716}$ and figure 4 for bifurcations while varying
$\unicode[STIX]{x1D6FE}$.
The
$\unicode[STIX]{x1D716}$-bifurcations at fixed
$\unicode[STIX]{x1D6FE}$, confirm the common observation that
$LR$ and
$TR$ always bifurcate in supercritical,
$\unicode[STIX]{x1D716}$-forward pitchfork bifurcations from the laminar base state. At
$\unicode[STIX]{x1D6FE}=0$, longitudinal and transverse rolls are related via symmetries, and we refer to both of them as
$R_{\unicode[STIX]{x1D706}}$, where the subscript indicates the wavelength of the roll pattern. At
$0<\unicode[STIX]{x1D6FE}\leqslant 20^{\circ }$, branches of
$LR$ and
$TR$ still are very close to each other. Only the
$LR$ branches, defining the onset of convection, are plotted to avoid clutter. At
$30^{\circ }\leqslant \unicode[STIX]{x1D6FE}\leqslant 40^{\circ }$,
$TR$ bifurcates outside the considered interval of
$0\leqslant \unicode[STIX]{x1D716}\leqslant 2$. At
$\unicode[STIX]{x1D6FE}\geqslant 50^{\circ }$, the branches of
$TR$ and
$LR$ bifurcate again closer to each other. The branches however differ significantly in amplitude and functional form. The
$TR$ branches show non-monotonic behaviour in
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$, e.g. a local maximum at
$\unicode[STIX]{x1D716}=1.8$ and
$\unicode[STIX]{x1D6FE}=90^{\circ }$. Non-monotonic branches of
$TR$ were also computed in vertical convection at
$Pr=0.71$ (Mizushima & Tanaka Reference Mizushima and Tanaka2002a,Reference Mizushima and Tanakab).
The
$LR$ branches monotonically increase in
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ with
$\unicode[STIX]{x1D716}$. For further increasing
$\unicode[STIX]{x1D716}$,
$LR$ appears to eventually approach the same
$\unicode[STIX]{x1D716}$-scaling law as reported for invariant states underlying straight convection rolls in horizontal convection at
$\unicode[STIX]{x1D6FE}=0^{\circ }$ (Waleffe et al. Reference Waleffe, Boonkasame and Smith2015). The large-
$\unicode[STIX]{x1D716}$ behaviour is observed for all
$\unicode[STIX]{x1D6FE}$. The observation that a scaling law of straight convection rolls at
$\unicode[STIX]{x1D6FE}=0^{\circ }$ also describes
$LR$ branches at
$\unicode[STIX]{x1D6FE}\neq 0^{\circ }$ suggests a particular scaling invariance of the nonlinear Oberbeck–Boussinesq equations under changes of inclination angle
$\unicode[STIX]{x1D6FE}$. This scaling invariance is discussed in the following paragraph.
At fixed
$\unicode[STIX]{x1D716}$, all
$\unicode[STIX]{x1D6FE}$ continuations of
$LR$ result in horizontal lines in
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ for
$0^{\circ }\leqslant \unicode[STIX]{x1D6FE}\leqslant \unicode[STIX]{x1D6FE}_{c2}$, e.g. the branch with
$\Vert \unicode[STIX]{x1D703}\Vert _{2}=0.08$ at
$\unicode[STIX]{x1D716}=0.1$ in figure 4. These horizontal lines are a remarkable feature of the bifurcation diagrams and can be explained as a consequence of a scaling invariance of the nonlinear Oberbeck–Boussinesq equations that holds for patterns or states that are steady in time and uniform in
$x$, like
$LR$.
For
$\unicode[STIX]{x1D6FE}<\unicode[STIX]{x1D6FE}_{c2}$, keeping
$\unicode[STIX]{x1D716}=\text{const.}$ implies that
$Ra\sim 1/\cos (\unicode[STIX]{x1D6FE})$ due to (2.12). The laminar solution (2.4)–(2.6) thus scales with
$\unicode[STIX]{x1D6FE}$ as
$U_{0}(z)\sim \sin (\unicode[STIX]{x1D6FE})/\sqrt{\cos (\unicode[STIX]{x1D6FE})}$,
${\mathcal{T}}_{0}(z)\sim 1$ and
$p_{0}(z)\sim -\cos (\unicode[STIX]{x1D6FE})$. Here,
$Pr$ is constant. Inserting the base-fluctuation decomposition
$\boldsymbol{U}=\boldsymbol{U}_{0}+\boldsymbol{u}$ and
${\mathcal{T}}={\mathcal{T}}_{0}+\unicode[STIX]{x1D703}$ into (2.1)–(2.3) and assuming that
$\unicode[STIX]{x2202}_{t}[\boldsymbol{u},\unicode[STIX]{x1D703}]=0$ and
$\unicode[STIX]{x2202}_{x}[\boldsymbol{u},\unicode[STIX]{x1D703}]=0$ for steady stripe/roll states, we observe the following. If temperature and velocity fluctuations are scaled as
$u(y,z)\sim \sin (\unicode[STIX]{x1D6FE})/\sqrt{\cos (\unicode[STIX]{x1D6FE})}$,
$v(y,z)\sim w(y,z)\sim \sqrt{\cos (\unicode[STIX]{x1D6FE})}$ and
$\unicode[STIX]{x1D703}(y,z)\sim 1$ for all components
$[u,v,w,\unicode[STIX]{x1D703},c]$, each term in the governing equations for the respective component is proportional to a common
$\unicode[STIX]{x1D6FE}$-dependent factor. This factor differs between components of the governing equations, but is common for all terms within a single equation:
The common scaling factors
$\sin (\unicode[STIX]{x1D6FE})$,
$\cos (\unicode[STIX]{x1D6FE})$ and
$\sqrt{\cos (\unicode[STIX]{x1D6FE})}$, as indicated above on the right-hand side, can be absorbed. The resulting scaling implies that any equilibrium at one value of
$\unicode[STIX]{x1D6FE}$ corresponds to a whole family of equilibria for
$0^{\circ }\leqslant \unicode[STIX]{x1D6FE}\leqslant 90^{\circ }$. The temperature scaling
$\unicode[STIX]{x1D703}(y,z)\sim 1$ directly implies that
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ of
$LR$ remains invariant under changes in
$\unicode[STIX]{x1D6FE}$ with
$\unicode[STIX]{x1D716}=\text{const.}$ This leads to self-similar
$LR$ branches under
$\unicode[STIX]{x1D716}$ continuation at fixed
$\unicode[STIX]{x1D6FE}$ (figure 3) and horizontal
$LR$ branches under
$\unicode[STIX]{x1D6FE}$ continuation at fixed
$\unicode[STIX]{x1D716}$ (figure 4). Moreover, any
$x$-uniform and steady invariant state for
$0<\unicode[STIX]{x1D6FE}\leqslant 90^{\circ }$ corresponds to a specific invariant state in the horizontal Rayleigh–Bénard case at
$\unicode[STIX]{x1D6FE}=0^{\circ }$. A similar relation has previously been reported for the infinite
$Pr$ limit only (Clever Reference Clever1973). The scaling relation provided here is valid for all
$Pr$ and a property of the full nonlinear Oberbeck–Boussinesq equations.
In the limit of a vertical gap (
$\unicode[STIX]{x1D6FE}\rightarrow 90^{\circ }$), the
$\cos ^{-1}(\unicode[STIX]{x1D6FE})$-scaling implies diverging
$Ra$. In this limit, the amplitude of the fixed
$u(y,z)$-profile diverges and the cross-flow components vanish,
$v,w\rightarrow 0$. The temperature field
$\unicode[STIX]{x1D703}(y,z)$ remains fixed. Consequently in a vertical gap, hot and cold streamwise jets without cross-flow and diverging streamwise velocity amplitude are invariant states in the
$Ra\rightarrow \infty$ limit. Any temperature field of
$LR$ found at
$\unicode[STIX]{x1D6FE}<90^{\circ }$ is a valid temperature field for these jets at infinite
$Ra$.
The subsequent sections discuss selected bifurcation diagrams covering the parameters where temporal dynamics between invariant states has been studied (Reference Reetz and SchneiderRS20). We do not systematically explain the bifurcations at all covered values of the control parameters but rather highlight important features of the bifurcations at selected control parameter values. In each section we summarize key features of the bifurcation structure and relate those to observed spatio-temporally complex dynamics of the flow. The sections are ordered by increasing values of the angle of inclination.
3.1 Skewed varicose state – subcritical connector of bistable rolls
The skewed varicose instability of Rayleigh–Bénard convection, first found as a spatially periodic instability at
$Pr=7$ (Busse & Clever Reference Busse and Clever1979), is experimentally observed to trigger a spatially localized transient pattern at
$Pr=1.07$ with very subtle varicose features (Bodenschatz, Pesch & Ahlers Reference Bodenschatz, Pesch and Ahlers2000, figure 7). In this section we report on a bifurcation from straight convection rolls to an equilibrium state capturing the observed skewed varicose pattern in a periodic domain. The bifurcating branch is subcritical, exists only below
$\unicode[STIX]{x1D716}_{c}$ of the skewed varicose instability, and connects two bistable straight convection rolls at different wavelengths and orientations. The subcritical coexistence of the skewed varicose equilibrium with bistable straight convection rolls may explain the spatial localization of the transiently observed pattern.
3.1.1 Bifurcation branch of skewed varicose states
When convection patterns in experiments or numerical simulations exhibit complex dynamics, we expect the existence of invariant states underlying the pattern dynamics. For the pattern dynamics emerging from the skewed varicose instability of straight convection rolls
$R_{\unicode[STIX]{x1D706}}$ at
$\unicode[STIX]{x1D6FE}=0^{\circ }$, however, we do not find invariant states at the control parameter values where the dynamics is observed. Direct numerical simulations in a minimal periodic domain can reproduce the transient dynamics of the skewed varicose pattern, but previous analysis of the temporal dynamics did not yield an underlying invariant state (Reference Reetz and SchneiderRS20).
An equilibrium state
$SV$ resembling the observed skewed varicose pattern is identified below
$\unicode[STIX]{x1D716}_{c}$ of the skewed varicose instability, as described in § 2.2. Numerical continuation of
$SV$ reveals a subcritical
$\unicode[STIX]{x1D716}$-backward pitchfork bifurcation from
$R_{\unicode[STIX]{x1D706}2}$ at
$\unicode[STIX]{x1D716}_{c}=1.019$. The bifurcation breaks the continuous translation symmetry
$\unicode[STIX]{x1D70F}(a_{x},0)$ of straight convection rolls
$R_{\unicode[STIX]{x1D706}2}$. Here we consider the rolls to be
$x$-aligned and periodic with wavelength
$\unicode[STIX]{x1D706}_{y}$. The bifurcating eigenmode shows a skewed three-dimensional flow structure. The bifurcating equilibrium
$SV$ is
$[4\unicode[STIX]{x1D706}_{x},4\unicode[STIX]{x1D706}_{y}]$-periodic and invariant under transformations of the symmetry group
$S_{SV}=\langle \unicode[STIX]{x03C0}_{xyz},\unicode[STIX]{x1D70F}(0.25,0.25)\rangle$. From the bifurcation point, the
$SV$ branch continues down in
$\unicode[STIX]{x1D716}$, undergoes a sequence of folds, and terminates at
$\unicode[STIX]{x1D716}=0.206$ in a bifurcation from straight convection rolls
$R_{\unicode[STIX]{x1D706}3}$ with wavelength
$\unicode[STIX]{x1D706}=2.8$ (panel
$\unicode[STIX]{x1D6FE}=0^{\circ }$ in figure 3). Thus, the equilibrium state
$SV$ connects two equilibrium states representing straight convection rolls at different wavelengths. The
$SV$ state exists only below the critical threshold parameter
$\unicode[STIX]{x1D716}_{c}$. The pure subcritical existence of
$SV$ explains why no temporal transition to an underlying invariant state at
$\unicode[STIX]{x1D716}>\unicode[STIX]{x1D716}_{c}$ has been found in Reference Reetz and SchneiderRS20.
Since the bifurcation branches are very cluttered at
$\unicode[STIX]{x1D6FE}=0^{\circ }$ in figure 3, we reproduce the bifurcation diagram schematically. In figure 5 the bifurcation branches are plotted in terms of their approximate dominating pattern wavelength
$\unicode[STIX]{x1D706}$ as a function of
$\unicode[STIX]{x1D716}$. Along the
$SV$ branch, convection rolls develop skewed relative orientations (figure 5b) until the rolls pinch-off and reconnect at an oblique orientation (figure 5c). At the bifurcation point,
$R_{\unicode[STIX]{x1D706}3}$ is rotated by
$74.6^{\circ }$ against the orientation of
$R_{\unicode[STIX]{x1D706}2}$ (figure 5a,c). To link these two different roll orientations, the continuous deformations in the skewed varicose pattern skip two instances for potential reconnection to straight rolls with orientations in between. Each of the potential reconnection points corresponds to a pair of folds along the
$SV$ branch. Here, three pairs of folds are observed but this number is specific to the chosen domain size. In between the first two folds at
$0.94<\unicode[STIX]{x1D716}<0.95$, the
$SV$ branch is bistable with
$R_{\unicode[STIX]{x1D706}2}$ and
$R_{\unicode[STIX]{x1D706}3}$ in a symmetry subspace of
$S_{SV}$. The stability of all branches is indicated by the line style. Overall, the bifurcation diagram indicates coexistence of stable (or weakly unstable)
$SV$ with stable
$R_{\unicode[STIX]{x1D706}2}$ and
$R_{\unicode[STIX]{x1D706}3}$ over a range of
$\unicode[STIX]{x1D716}$. The coexistence of invariant states suggests a spatial coexistence of straight convection rolls and the skewed varicose patterns.
The convection pattern along the
$SV$ branch at
$\unicode[STIX]{x1D716}<\unicode[STIX]{x1D716}_{c}$ may be compared to the convection pattern observed transiently in time along a simulated transition at
$\unicode[STIX]{x1D716}=1.05>\unicode[STIX]{x1D716}_{c}$ (dashed line in figure 5d and § 4.3 in Reference Reetz and SchneiderRS20). The midplane temperature contours of
$SV$ along its subcritical bifurcation branch partly match the transient patterns along the supercritical temporal transition. We find matching patterns at initial instances in time when straight convection rolls are observed (figure 5c,g), and at an intermediate time when the transient pattern of skewed varicose pattern emerges (figure 5b,f). Thus,
$SV$ indeed captures the observed transient pattern triggered by the skewed varicose instability, but the comparison is for different
$\unicode[STIX]{x1D716}$. This observation raises the question of how the transient temporal dynamics observed above critical thresholds can be related to an equilibrium state existing only below critical thresholds.
For
$\unicode[STIX]{x1D6FE}=0^{\circ }$ (Rayleigh–Bénard), the subcritical bifurcation of the skewed varicose state connects two equilibrium states of straight convection rolls at different wavelength and orientation. Midplane temperature fields of equilibrium states (a–c) are chosen along the
$SV$ branch, plotted here in a schematic bifurcation diagram in terms of an approximate dominating pattern wavelength
$\unicode[STIX]{x1D706}$ over
$\unicode[STIX]{x1D716}$ in (d). A simulated temporal transition from unstable
$R_{\unicode[STIX]{x1D706}2}$ to stable
$R_{\unicode[STIX]{x1D706}3}$ at supercritical
$\unicode[STIX]{x1D716}>\unicode[STIX]{x1D716}_{c}$ (Reference Reetz and SchneiderRS20) is indicated by the dashed line. Snapshots from the temporal transition (g,f,e) show matching skewed varicose patterns between the equilibrium state (b) for
$\unicode[STIX]{x1D716}<\unicode[STIX]{x1D716}_{c}$ and the transient state (f) at
$\unicode[STIX]{x1D716}>\unicode[STIX]{x1D716}_{c}$. The orientation of
$R_{\unicode[STIX]{x1D706}3}$ differs between the terminating bifurcation branch at
$\unicode[STIX]{x1D716}=0.2$ (a) and the attracted temporal dynamics at
$\unicode[STIX]{x1D716}=1.05$ (e).

3.2 Subharmonic oscillations – standing and travelling waves
Subharmonic oscillations are observed as standing wave patterns emerging in spatially localized patches that may travel across extended domains (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000; Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). Here, the periodic orbit
$SSW$, underlying the standing wave, is found to coexist with a travelling wave state. Standing and travelling wave states always bifurcate together in equivariant Hopf bifurcations. Both standing and travelling waves capture observed patterns of spatially subharmonic oscillations. The existence of a travelling wave state explains the observed travelling dynamics of the pattern.
Symmetry-related pairs of standing wave states
$SSW$ (a) and travelling wave states
$STW$ (b) bifurcate together in equivariant Hopf-bifurcations from
$LR$. The pattern of
$SSW$ and
$STW$ is visualized by temperature contours in three and two dimensions (midplane). The bifurcation diagram in (c) shows the branches of
$SSW$ (black) and
$STW$ (grey) at
$\unicode[STIX]{x1D716}=1.5$. Other branches shown in figure 4 for
$\unicode[STIX]{x1D716}=1.5$ are suppressed for clarity. The inset indicates stable (solid) and unstable (dotted) branches close to the bifurcation at
$\unicode[STIX]{x1D6FE}=8.5^{\circ }$ giving rise to the states in (a) and (b) at
$\unicode[STIX]{x1D6FE}=15^{\circ }$. Here ‘RS20’ labels the control parameter value where the temporal evolution has been studied in Reference Reetz and SchneiderRS20. Panel (d) illustrates how the orbit period
$T_{\mathit{SSW}}$ of
$SSW$ and the travel time
$T_{STW}$ of
$STW$ across
$L_{x}$ approximately follow the mean advection time
$L_{x}/\bar{U}_{0}$ of the laminar base flow
$B$ (dashed–dotted line). Note the ‘hydrolic jumps’ where phase and advection velocity match at
$\unicode[STIX]{x1D6FE}=65^{\circ }$ and at
$\unicode[STIX]{x1D6FE}=110^{\circ }$ of the
$STW$ branch.

3.2.1 Equivariant Hopf bifurcation from longitudinal rolls
Starting from the periodic orbit
$SSW$, found at
$[\unicode[STIX]{x1D6FE},\unicode[STIX]{x1D716}]=[15^{\circ },1.5]$ in Reference Reetz and SchneiderRS20, we continue the branch down in
$\unicode[STIX]{x1D6FE}$. The
$SSW$ branch bifurcates from
$LR$ at
$[\unicode[STIX]{x1D6FE}_{c},\unicode[STIX]{x1D716}_{c}]=[8.5^{\circ },1.5]$. This bifurcation is a
$\unicode[STIX]{x1D6FE}$-forward, supercritical Hopf bifurcation with a critical orbit period of
$T_{\mathit{SSW}}=80.2$. The bifurcation corresponds to four complex eigenvalues crossing the imaginary axis at
$\unicode[STIX]{x1D714}_{i}=\pm 0.078$ controlling the period
$T_{\mathit{SSW}}=2\unicode[STIX]{x03C0}/|\unicode[STIX]{x1D714}_{i}|$. The Hopf bifurcation to
$SSW$ accounts only for one pair of complex neutral eigenmodes. The other pair gives rise to a travelling wave state that we term subharmonic travelling wave (
$STW$) with a critical phase speed of
$c=L_{x}/T_{STW}=0.055$, where
$T_{STW}=T_{\mathit{SSW}}$. Both invariant states, the
$\unicode[STIX]{x03C0}_{y}$-symmetric
$STW$ and the
$\unicode[STIX]{x03C0}_{xyz}$-symmetric
$SSW$ (shown in the left panels of figure 6a,b), each have a counter-propagating sibling state obtained via
$\unicode[STIX]{x03C0}_{xz}$-transformation. Both invariant states capture a subharmonic oscillatory convection pattern invariant under
$\unicode[STIX]{x1D70F}(0.5,0.5)$. An equilibrium pattern very close to
$STW$ can be observed in spatially forced horizontal convection (Weiss, Seiden & Bodenschatz Reference Weiss, Seiden and Bodenschatz2012). The
$STW$ state also resembles the subharmonic ‘sinucose’ state arising from an instability of longitudinal streaks in a pure shear flow (Waleffe Reference Waleffe1997).
At parameters where
$SSW$ and
$STW$ bifurcate locally from
$LR$, they share the same bifurcation point (see panels for
$10^{\circ }\leqslant \unicode[STIX]{x1D6FE}\leqslant 60^{\circ }$ in figure 3). This robust feature in the bifurcation diagram is a consequence of equivariant Hopf bifurcations. It is known that Hopf bifurcations that break the
$O(2)$-symmetry of a flow must result in two branches originating from the bifurcation: a standing wave branch and a travelling wave branch (Knobloch Reference Knobloch1986). At most one of the two branches is stable. In the present case, the Hopf bifurcation breaks the
$O(2)$-symmetry of the
$x$-uniform state
$LR$. The bifurcation at
$[\unicode[STIX]{x1D6FE}_{c},\unicode[STIX]{x1D716}_{c}]=[8.5^{\circ },1.5]$ (inset panel in figure 6c) has both the branches bifurcating supercritically. The
$SSW$ state is initially stable and
$STW$ is unstable. This corresponds to one specific case of the six discussed in Knobloch (Reference Knobloch1986). However, here the bifurcation is secondary. While in Knobloch (Reference Knobloch1986) bifurcations from a non-patterned two-dimensional primary state with a spatial
$O(2)$-symmetry are discussed, here, the bifurcating secondary equilibrium state
$LR$ is a three-dimensional state that is symmetric under transformations of
$O(2)\times Z_{2}$. The additional third dimension and the additional reflection symmetry do not affect the conditions necessary for equivariant Hopf bifurcations (Knobloch Reference Knobloch1986).
Continuations of
$SSW$ and
$STW$ in
$\unicode[STIX]{x1D6FE}$ reveal their existence over a large range of inclination angles
$\unicode[STIX]{x1D6FE}$, including
$\unicode[STIX]{x1D6FE}>90^{\circ }$ where their parent state
$LR$ does not exist anymore (figure 6c). Over this range in
$\unicode[STIX]{x1D6FE}$, the orbit period of
$SSW$ and the propagation time
$T_{STW}=L_{x}/c$ of
$STW$ with phase velocity
$c$ and
$L_{x}=2\unicode[STIX]{x1D706}_{x}$ follow approximately the mean laminar advection time
$L_{x}/\bar{U}_{0}$ (figure 6d). While the state branch of
$STW$ at
$\unicode[STIX]{x1D716}=1.5$, shown as light grey line in figure 6(c), connects two Hopf bifurcations, one at small and one at large
$\unicode[STIX]{x1D6FE}$, the
$SSW$ branches at
$\unicode[STIX]{x1D716}=1.5$, shown as black lines in figure 6(c), originating from these two bifurcations remain disconnected. The branch bifurcating forward at
$\unicode[STIX]{x1D6FE}_{c}=8.5^{\circ }$ undergoes a fold at
$\unicode[STIX]{x1D6FE}=15.2^{\circ }$ (figure 6c), explaining why no temporal state transition from
$LR$ to
$SSW$ was found at
$\unicode[STIX]{x1D6FE}=17^{\circ }$ in Reference Reetz and SchneiderRS20. The fold destabilizes
$SSW$ and connects to a state branch reaching to subcritical parameters where
$LR$ is linearly stable. Similar folds occur under
$\unicode[STIX]{x1D716}$ continuations at
$\unicode[STIX]{x1D6FE}=10^{\circ }$ and
$\unicode[STIX]{x1D6FE}=20^{\circ }$. Beyond these folds,
$SSW$ state branches terminate and show ‘loose ends’ in the bifurcation diagrams. These terminations correspond to global bifurcations which we explain in the next subsection.
3.2.2 Global bifurcation to subharmonic standing waves
The global bifurcation at
$\unicode[STIX]{x1D6FE}=10^{\circ }$ occurs at
$Ra_{c}=2230.25$ (
$\unicode[STIX]{x1D716}_{c}=0.286$), where the pre-periodic orbit
$SSW$, satisfying (2.7) with
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}=\unicode[STIX]{x03C0}_{y}\unicode[STIX]{x1D70F}(0.25,0.25)$ and a pre-period of
$T^{\prime }=T/4$, collides with a heteroclinic cycle between two symmetry related saddle states
$TR$ and
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ (figure 7). At the global bifurcation point, the spectrum of eigenvalues of
$TR$ is computed in a symmetry subspace
$\unicode[STIX]{x1D6F4}_{0}$ given by the
$[2\unicode[STIX]{x1D706}_{x},2\unicode[STIX]{x1D706}_{y}]$-periodic domain with imposed symmetries of
$SSW$, namely
$\unicode[STIX]{x1D70F}(0.5,0.5)$ and
$\unicode[STIX]{x03C0}_{xyz}$. The five leading eigenvalues are real and have the specific numerical values
$[\unicode[STIX]{x1D714}_{1},\unicode[STIX]{x1D714}_{2},\unicode[STIX]{x1D714}_{3},\unicode[STIX]{x1D714}_{4},\unicode[STIX]{x1D714}_{5}]$
$=$
$[0.048,0.045,-0.090,-0.120,-0.138]$. The midplane temperature contours of the associated eigenmodes
$[\boldsymbol{e}_{1}^{u},\boldsymbol{e}_{2}^{u},\boldsymbol{e}_{3}^{s}\boldsymbol{e}_{4}^{s},\boldsymbol{e}_{5}^{s}]$ are given in figure 7(h). Eigenvalues and eigenmodes of
$TR$ do not change significantly when
$Ra$ crosses
$Ra_{c}$. In contrast to the heteroclinic cycle discussed in Reference Reetz and SchneiderRS20, where each of the two symmetry related instances of equilibrium state
$OWR$ (further discussed in § 3.3.2) has a single unstable eigenmode, the present cycle connects symmetry related instances of
$TR$ with two unstable eigenmodes each. Perturbations of
$TR$ with the eigenmode
$\boldsymbol{e}_{1}^{u}$ trigger a state transition
$TR\rightarrow \unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$, while perturbations with
$\boldsymbol{e}_{2}^{u}$ lead to
$LR$ which is dynamically stable in
$\unicode[STIX]{x1D6F4}_{0}$ at these control parameter values. The symmetry relation between
$TR$ and
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ guarantees that
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ has the same eigenvalues as
$TR$ and symmetry related eigenmodes
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}[\boldsymbol{e}_{1}^{u},\boldsymbol{e}_{2}^{u},\boldsymbol{e}_{3}^{s},\boldsymbol{e}_{4}^{s},\boldsymbol{e}_{5}^{s}]$, allowing for the returning transition
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR\rightarrow TR$ to close the heteroclinic cycle.
The pre-periodic orbit
$SSW$ approaches a global bifurcation at
$\unicode[STIX]{x1D6FE}=10^{\circ }$ and
$Ra_{c}=2230.25$ (
$\unicode[STIX]{x1D716}_{c}=0.286$), where it collides with a robust heteroclinic cycle between
$TR$ and the symmetry related equilibrium
$\unicode[STIX]{x1D70E}_{SSW}TR$ with
$\unicode[STIX]{x1D70E}_{SSW}=\unicode[STIX]{x03C0}_{y}\unicode[STIX]{x1D70F}(0.25,0.25)$. (a–e) Sequence of midplane temperature contours along the dynamical connection
$TR\rightarrow \unicode[STIX]{x1D70E}_{SSW}TR$ at
$Ra=2230$. (f) State space projection of the lower (LB) and the upper branch (UB) of
$SSW$ (cf. figure 8a). Inset enlarges the orbit trajectories (black dots) near
$TR$ and the heteroclinic cycle (grey lines). (g)
$L_{2}$-distance of the
$SSW_{UB}$ orbit trajectory from
$TR$ at
$Ra=2230.24$ very close to
$Ra_{c}$. The dynamics of
$SSW_{UB}$ is exponential for most of the pre-period and governed by two eigenvalues of
$TR$. (h) Midplane temperature contours of two unstable and three stable eigenmodes
$\boldsymbol{e}_{i}^{u/s}$ of
$TR$ associated to the five leading eigenvalues
$[\unicode[STIX]{x1D714}_{1},\unicode[STIX]{x1D714}_{2},\unicode[STIX]{x1D714}_{3},\unicode[STIX]{x1D714}_{4},\unicode[STIX]{x1D714}_{5}]=[0.048,0.045,-0.090,-0.120,-0.138]$ of
$TR$.

Direct numerical simulations indicate that states close to the heteroclinic cycle are eventually attracted to
$LR$. To show that this heteroclinic cycle is dynamically unstable but structurally stable, we identify two symmetry subspaces of
$\unicode[STIX]{x1D6F4}_{0}$ in which either
$TR\rightarrow \unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ or
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR\rightarrow TR$ exists as a heteroclinic connection between an equilibrium with a single unstable eigenmode and a dynamically stable equilibrium. By doing so, the heteroclinic cycle is shown to satisfy all conditions of a structurally stable, or robust, heteroclinic cycle between two symmetry related equilibrium states (Krupa Reference Krupa1997), also discussed in Reference Reetz and SchneiderRS20. Subspace
$\unicode[STIX]{x1D6F4}_{1}$ is given by imposing the symmetries in the group
$\langle \unicode[STIX]{x03C0}_{y},\unicode[STIX]{x03C0}_{xz},\unicode[STIX]{x1D70F}(0.5,0.5)\rangle$ and contains the connection
$TR\rightarrow \unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$. Of the five initially considered eigenmodes in
$\unicode[STIX]{x1D6F4}_{0}$,
$TR$ in
$\unicode[STIX]{x1D6F4}_{1}$ has still
$[\boldsymbol{e}_{1}^{u},\boldsymbol{e}_{3}^{s},\boldsymbol{e}_{5}^{s}]$ and
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ in
$\unicode[STIX]{x1D6F4}_{1}$ has still
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}[\boldsymbol{e}_{4}^{s},\boldsymbol{e}_{5}^{s}]$. Subspace
$\unicode[STIX]{x1D6F4}_{\unicode[STIX]{x1D70F}}$ is given by imposing the symmetries in the group
$\langle \unicode[STIX]{x03C0}_{y}\unicode[STIX]{x1D70F}(0.5,0),\unicode[STIX]{x03C0}_{xz}\unicode[STIX]{x1D70F}(0.5,0),\unicode[STIX]{x1D70F}(0.5,0.5)\rangle$ and contains the connection
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR\rightarrow TR$. Of the five initially considered eigenmodes in
$\unicode[STIX]{x1D6F4}_{0}$,
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ in
$\unicode[STIX]{x1D6F4}_{\unicode[STIX]{x1D70F}}$ has still
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}[\boldsymbol{e}_{1}^{u},\boldsymbol{e}_{3}^{s},\boldsymbol{e}_{5}^{s}]$ and
$TR$ in
$\unicode[STIX]{x1D6F4}_{\unicode[STIX]{x1D70F}}$ has still
$[\boldsymbol{e}_{4}^{s},\boldsymbol{e}_{5}^{s}]$. Using the classification of eigenvalues and associated stability theorem in Krupa & Melbourne (Reference Krupa and Melbourne1995), we identify
$\unicode[STIX]{x1D714}_{1}$ as expanding,
$\unicode[STIX]{x1D714}_{2}$ as transverse,
$\unicode[STIX]{x1D714}_{4}$ as contracting and
$\unicode[STIX]{x1D714}_{5}$ as a radial eigenvalue. Eigenvalue
$\unicode[STIX]{x1D714}_{3}$ exists in the same subspace as the expanding eigenvalue
$\unicode[STIX]{x1D714}_{1}$ and, therefore, does not affect the dynamical stability. Since
$\min (-\unicode[STIX]{x1D714}_{4},\unicode[STIX]{x1D714}_{1}-\unicode[STIX]{x1D714}_{2})\ngtr \unicode[STIX]{x1D714}_{1}$, the heteroclinic cycle is not asymptotically stable (Krupa & Melbourne Reference Krupa and Melbourne1995, Theorem 2.7).
Before
$SSW$ disappears in the global bifurcation at
$Ra_{c}$, the solution branch undergoes a fold at
$Ra<Ra_{c}$ (figure 8a). The existence of such a fold near a global bifurcation follows from the dynamical stability of the bifurcating periodic orbit relative to the dynamical stability of the heteroclinic cycle. To analyse the stability, we consider the linearized dynamics around the heteroclinic cycle
$TR\rightarrow \unicode[STIX]{x1D70E}_{\mathit{SSW}}TR\rightarrow TR$ at
$Ra_{c}$ and obtain the Poincaré map (see Bergeon & Knobloch Reference Bergeon and Knobloch2002, for a derivation)
with constant
$c>0$ and control parameter
$\unicode[STIX]{x1D707}\propto Ra_{c}-Ra$. Variable
$\unicode[STIX]{x1D701}_{i}\ll 1$ describes a local coordinate in a Poincaré section located at a distance
$\unicode[STIX]{x1D700}\ll 1$ from
$TR$ and defining a small perturbation around the state vector of
$TR$ as
$\boldsymbol{x}^{\prime }=\boldsymbol{x}_{TR}+\unicode[STIX]{x1D701}_{i}\boldsymbol{e}_{1}^{u}+\unicode[STIX]{x1D700}\boldsymbol{e}_{4}^{s}$. The heteroclinic cycle corresponds to
$\unicode[STIX]{x1D701}_{i}=0$ and is reached at
$\unicode[STIX]{x1D707}=0$. The bifurcating periodic orbit is a fixed point
$\bar{\unicode[STIX]{x1D701}}$ of the map such that
$\bar{\unicode[STIX]{x1D701}}=c\bar{\unicode[STIX]{x1D701}}^{\unicode[STIX]{x1D70C}}+\unicode[STIX]{x1D707}$. Since
$\unicode[STIX]{x1D70C}>1$, a nearby fixed point
$\bar{\unicode[STIX]{x1D701}}$ exists only for
$\unicode[STIX]{x1D707}>0$ and
$Ra<Ra_{c}$, respectively. The graph in figure 8(b) illustrates the map (3.7) and shows that the fixed point
$\bar{\unicode[STIX]{x1D701}}$ is dynamically stable. The stability of
$\bar{\unicode[STIX]{x1D701}}$ assumes that no additional transverse eigendirections are unstable. However, the symmetry subspace that contains
$SSW$ also contains
$\boldsymbol{e}_{2}^{u}$, which must be taken into account. Thus, the above analysis predicts that
$SSW$ bifurcates with a single unstable eigendirection from the heteroclinic cyle, namely
$\boldsymbol{e}_{2}^{u}$. Since
$SSW$ for
$Ra>Ra_{c}$ has two unstable eigenmodes, the periodic orbit must undergo a fold prior to the global bifurcation to stabilize the extra unstable eigendirection. Such a fold also exists at
$\unicode[STIX]{x1D6FE}=20^{\circ }$ (figure 3), but further away from the global bifurcation. Note that unlike the global bifurcation discussed in Bergeon & Knobloch (Reference Bergeon and Knobloch2002), the present bifurcation involves first, a heteroclinic cycle between two symmetry related equilibrium states and second, a dynamically unstable periodic orbit such that the fold may have a stabilising effect.
Before
$SSW$ disappears in a global heteroclinic bifurcation at
$\unicode[STIX]{x1D6FE}=10^{\circ }$ and
$Ra_{c}=2230.25$ (
$\unicode[STIX]{x1D716}_{c}=0.286$), the branch undergoes a saddle node bifurcation. (a) Enlarged bifurcation diagram showing the branch of
$SSW$ indicating the maximum
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ over one cycle (cf. panel
$\unicode[STIX]{x1D6FE}=10^{\circ }$ in figure 3). (b) Graphical representation of the linearized Poincaré map (3.7) for
$\unicode[STIX]{x1D707}=0.1$ around the heteroclinic cycle between symmetry related instances of
$TR$. The fixed point
$\bar{\unicode[STIX]{x1D701}}$ is stable as the slope of the map at
$\bar{\unicode[STIX]{x1D701}}$ is less than one (dashed line). (c,d) Towards the global bifurcation, the orbit period of
$SSW$ approaches an infinite period according to a logarithmic scaling law shown as a function of
$Ra$ in (c) and for rescaled
$Ra$ in (d). This law follows from the linearized map (3.7), as described in the text. The constant
$c_{T}=560$ is determined by fitting the data close to
$Ra_{c}$.

The period of
$SSW$ must increase towards an infinite time period as the periodic orbit approaches the heteroclinic cycle. The map (3.7) suggests an asymptotic scaling law for
$T$ as a function of
$Ra$ close to the global bifurcation. Since
$\unicode[STIX]{x1D701}\ll 1$ and
$\unicode[STIX]{x1D70C}>1$, the periodic orbit is given by the approximation
$\bar{\unicode[STIX]{x1D701}}\approx \unicode[STIX]{x1D707}$. Over a full period of
$SSW$, the orbit trajectory visits both
$TR$ and
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ twice. The time the orbit trajectory spends in the
$\unicode[STIX]{x1D700}$-neighbourhood of
$TR$ or
$\unicode[STIX]{x1D70E}_{\mathit{SSW}}TR$ dominates the entire orbit period
$T$ (figure 7g) such that
$T$ satisfies the approximation
$\bar{\unicode[STIX]{x1D701}}\approx \unicode[STIX]{x1D700}\exp (-\unicode[STIX]{x1D714}_{1}T/4)$. Hence, the period of
$SSW$ is expected to increase as
$T\approx -4/\unicode[STIX]{x1D714}_{1}\ln (Ra_{c}-Ra)+c_{T}$ with constant
$c_{T}=4\ln (\unicode[STIX]{x1D700})/\unicode[STIX]{x1D714}_{1}$ so that asymptotically the period scales as
$T\sim -4/\unicode[STIX]{x1D714}_{1}\ln (Ra_{c}-Ra)$. Using a multi-shooting method,
$SSW$ is continued close to the global bifurcation. The increasing orbit period confirms the predicted scaling law (figure 8c,d).
3.3 Wavy rolls with defects – connecting coexisting state branches
Convection patterns of wavy rolls are observed to quickly incorporate defects in large experimental domains (Daniels & Bodenschatz Reference Daniels and Bodenschatz2002; Daniels et al. Reference Daniels, Brausch, Pesch and Bodenschatz2008). These defects may form interfaces between spatially coexisting wavy rolls at different orientations against the base flow. Here, a bifurcation and stability analysis of
$WR$ reveals four new equilibrium states, including obliquely oriented states and rolls with defects, that all coexist with
$WR$ for the same control parameter values. The multiple states can give rise to the observed spatial coexistence.
3.3.1 Pitchfork bifurcations from longitudinal rolls
Equilibrium states
$WR$ emerge either in pitchfork bifurcations from
$LR$ at inclinations
$20^{\circ }\leqslant \unicode[STIX]{x1D6FE}\leqslant 80^{\circ }$ or in saddle-node bifurcations in the absence of
$LR$ at
$90^{\circ }\leqslant \unicode[STIX]{x1D6FE}\leqslant 100^{\circ }$. The fact that
$WR$ can exist without bifurcating from
$LR$ is known from thermal Couette flow (Clever & Busse Reference Clever and Busse1992). Almost all computed pitchfork bifurcations from
$LR$ to
$WR$ are either
$\unicode[STIX]{x1D716}$-forward or
$\unicode[STIX]{x1D6FE}$-backward. This observation holds even when the
$WR$ branches develop additional folds, like in the bifurcation diagrams at
$\unicode[STIX]{x1D6FE}=80^{\circ }$ or at
$\unicode[STIX]{x1D716}=1.5$. The only
$\unicode[STIX]{x1D6FE}$-forward bifurcation of
$WR$ from
$LR$ is observed at
$\unicode[STIX]{x1D716}=0.1$ (figure 4,
$\unicode[STIX]{x1D716}=0.1$).
3.3.2 Additional bifurcations from wavy rolls
In most cases,
$WR$ branches continue to large
$\unicode[STIX]{x1D716}$. This does not imply the absence of connections to other invariant states. When
$WR$ lose stability, additional bifurcations occur. We demonstrate the increasing number of invariant states and their patterns by following the higher-order instabilities of
$WR$ at
$\unicode[STIX]{x1D6FE}=40^{\circ }$. Arnoldi iteration for
$WR$ indicates the stability of the
$WR$ branch up to
$\unicode[STIX]{x1D716}=0.256$ in a
$[2\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$-periodic domain (figure 9g). At this point, a subcritical pitchfork bifurcation breaks the
$\unicode[STIX]{x03C0}_{xz}$- and
$\unicode[STIX]{x03C0}_{y}$-symmetry and gives rise to an equilibrium state showing disconnected wavy rolls and named
$DWR$. The
$DWR$ state is invariant under
$\unicode[STIX]{x03C0}_{xyz}$. Following the
$DWR$ branch from the pitchfork bifurcation, it undergoes a saddle-node bifurcation at
$\unicode[STIX]{x1D716}=0.254$, becomes bistable with
$WR$ and connects to an equilibrium state of oblique rolls that continues as a stable branch to larger values of
$\unicode[STIX]{x1D716}$. The
$DWR$ state represents stationary roll defects that along its branch breaks the topology of rolls in a double periodic domain (figure 9b), and connects convection rolls at different orientations.
Bifurcation sequence from wavy rolls to oblique wavy rolls (a–d). Bifurcation diagram (e) shows how seven equilibrium states are connected over
$\unicode[STIX]{x1D716}$ and indicates stable (solid) and unstable (dotted) parts of the branches. Dynamical stability is computed by Arnoldi iteration in a
$(2\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y})$-periodic domain. Panel (g) enlarges the parameter region where
$DWR$ connects
$WR$ with
$OR$. Panel (f) enlarges the parameter region where
$OWR_{1}$ with pattern wavenumber
$m=1$ connects
$OWR$ with pattern wavenumber
$m=2$ with
$OR$. Here ‘RS20’ labels the control parameter value at which the temporal evolution has been studied in Reference Reetz and SchneiderRS20.

The
$OR$ state is a secondary state bifurcating from the laminar flow
$B$. The oblique orientation at angle
$\tan ^{-1}(\unicode[STIX]{x1D706}_{y}/2\unicode[STIX]{x1D706}_{x})$ against the laminar flow direction coincides with the diagonal of the
$[2\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$-periodic domain (figure 9c). The
$OR$ state is invariant under transformations of
$S_{OR}=\langle \unicode[STIX]{x03C0}_{xyz},\unicode[STIX]{x1D70F}(a_{x},\pm a_{x})\rangle$, with
$a_{x}\in [0,1)$ corresponding to an
$O(2)$-symmetric state. The sign in front of the continuous shift factor
$a_{x}$ differs between left
$(+)$ and right
$(-)$ oblique rolls, with
$OR^{l}=\unicode[STIX]{x03C0}_{y}OR^{r}$. Here, the superscripts ‘
$l$’ and ‘
$r$’ denote left and right, respectively. When
$OR$ becomes unstable, an
$\unicode[STIX]{x1D716}$-forward pitchfork bifurcation at
$\unicode[STIX]{x1D716}=0.456$ gives rise to stable oblique wavy rolls, see figure 9(d). As
$OR$,
$OWR$ can have a left or right orientation. The symmetry group of
$OWR$ is
$S_{OWR}=\langle \unicode[STIX]{x03C0}_{xyz},\unicode[STIX]{x1D70F}(0.5,\pm 0.5)\rangle$. Equilibrium
$OWR$ with a wavy pattern of wavenumber
$m=2$ along the domain diagonal loses stability at
$\unicode[STIX]{x1D716}=0.463$ to an equilibrium
$OWR_{1}$ with a pattern wavenumber of
$m=1$ and broken
$\unicode[STIX]{x1D70F}(0.5,\pm 0.5)$-symmetry. The branch of
$OWR_{1}$ undergoes a saddle-node bifurcation at
$\unicode[STIX]{x1D716}=0.477$ and terminates on
$OR$ at
$\unicode[STIX]{x1D716}=0.476$ (figure 9f). The small
$\unicode[STIX]{x1D716}$-range with
$\unicode[STIX]{x0394}\unicode[STIX]{x1D716}=0.476-0.456=0.02$ between the two symmetry-breaking bifurcations of
$OWR_{1}$ with
$m=1$ and
$OWR$ with
$m=2$ from
$OR$ suggests a nearby codimension-2 point with spatial 1 : 2 resonance. When stable
$OWR_{1}$ disappear in the saddle-node bifurcation, the temporal dynamics becomes attracted to a robust heteroclinic cycle between unstable instances of
$OWR$ that is discussed in Reference Reetz and SchneiderRS20. The branch of
$OWR$ continues as an unstable branch until it terminates at
$\unicode[STIX]{x1D716}=0.8$ on ribbons (
$RB$), an unstable equilibrium state bifurcating together with
$OR$ from
$B$ in an equivariant pitchfork bifurcation (figure 9e). The detailed properties of
$RB$ are discussed in the following section, but we already note that
$RB$ shows disconnected rolls or plumes, similar to
$DWR$. Thus, we find two instances of bifurcation sequences that may be described as ‘straight rolls bifurcate to wavy rolls, wavy rolls bifurcate to disconnected rolls’. In one instance the sequence happens for longitudinal orientation and in the other instance for oblique orientation. The fact that all of the above states coexist with the WR branch at equal control parameters explains that all patterns represented by the invariant states can spatially coexist with wavy rolls in large domains.
3.4 Knots and ribbons – two different types of bimodal states
Observations of knot patterns exist in horizontal convection (Busse & Whitehead Reference Busse and Whitehead1974; Busse & Clever Reference Busse and Clever1979) and inclined layer convection (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000). They have been described as ‘bimodal convection’ in both cases. Here, the properties of equilibrium states for knots are compared to ribbons, a bimodal state identified in the previous section. A decomposition along their bifurcation branches implies that knots and ribbons are bimodal states that fundamentally differ in their bifurcation structure.
3.4.1 Bifurcations to states for knots and ribbons
The
$KN$ state at
$\unicode[STIX]{x1D6FE}=80^{\circ }$ bifurcates
$\unicode[STIX]{x1D716}$-forward from
$TR$. At smaller
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ than
$TR$,
$KN$ continues without folds and terminates in
$\unicode[STIX]{x1D716}$-backward bifurcations from
$LR$ (figure 3,
$\unicode[STIX]{x1D6FE}=80^{\circ }$). This bifurcation sequence requires
$\unicode[STIX]{x1D6FE}>\unicode[STIX]{x1D6FE}_{c2}$ and was previously analysed using two-mode interactions (Fujimura & Kelly Reference Fujimura and Kelly1993). At
$\unicode[STIX]{x1D6FE}=90^{\circ }$,
$LR$ does not exist at finite
$\unicode[STIX]{x1D716}$ and the
$KN$ branch terminates in a bifurcation from an equilibrium state we term subharmonic lambda plumes and briefly discuss in appendix A.
The
$RB$ state is an equilibrium state found via continuing the state branches of
$OWR$ and
$WR$ that terminate in
$\unicode[STIX]{x1D716}$-bifurcations from
$RB$ at
$\unicode[STIX]{x1D6FE}=40^{\circ }$ and
$\unicode[STIX]{x1D6FE}=50^{\circ }$, respectively (figure 3). The
$RB$ state is invariant under transformations of
$S_{RB}=\langle \unicode[STIX]{x03C0}_{y},\unicode[STIX]{x03C0}_{xz},\unicode[STIX]{x1D70F}(0.5,0.5)\rangle$. Neither experiments nor simulations of ILC observe the pattern of
$RB$ as a dynamically stable pattern at the considered parameters. However, we refer to experimental observations of ‘ribbons’ in Taylor–Couette flow (Tagg et al. Reference Tagg, Edwards, Swinney and Marcus1989). Ribbons in Taylor–Couette flow are analogous to ribbons in ILC. In Taylor–Couette flow, they bifurcate together with oblique spirals (Chossat & Iooss Reference Chossat and Iooss1994) and are connected via oblique wavy cross-spirals (Pinter, Lücke & Hoffmann Reference Pinter, Lücke and Hoffmann2006), two states that are comparable to
$OR$ and
$OWR$ in ILC. As in the Taylor–Couette flow,
$OR$ and
$RB$ in ILC bifurcate robustly together in equivariant bifurcations (Knobloch Reference Knobloch1986) and are connected via
$OWR$. However,
$OR$ and
$RB$ are stationary states and their bifurcation is an equivariant pitchfork bifurcation, unlike equivariant Hopf bifurcations such as those found in Taylor–Couette flow, and those discussed in § 3.2.
3.4.2 Decomposition in terms of straight convection rolls
As a consequence of the stationary equivariant bifurcation, the linear relation
$\boldsymbol{x}_{RB}=\unicode[STIX]{x1D6FC}\boldsymbol{x}_{OR}^{l}+\unicode[STIX]{x1D6FD}\boldsymbol{x}_{OR}^{r}$ with
$\unicode[STIX]{x1D6FC}=\unicode[STIX]{x1D6FD}$ holds at the bifurcation points, where
$\boldsymbol{x}_{RB}$ indicates the state space vector of
$RB$, and
$\boldsymbol{x}_{OR}^{l}$ and
$\boldsymbol{x}_{OR}^{r}$ are the state space vectors of
$OR^{l}$ and
$OR^{r}$, respectively. This linear decomposition is valid for all parameters, where
$RB$ and
$OR$ bifurcate from laminar flow
$B$. Since
$RB$ emerges as a linear superposition of two differently oriented straight convection rolls, we call
$RB$ a ‘bimodal state’. The term ‘bimodal’ has been used previously to describe knot patterns of straight convection rolls at orthogonal orientations in experiments of Rayleigh–Bénard convection (Busse & Whitehead Reference Busse and Whitehead1974) and in experiments of ILC for
$\unicode[STIX]{x1D6FE}>\unicode[STIX]{x1D6FE}_{c2}$ (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000). In line with previously used terminology, we describe
$KN$ and
$RB$ both as bimodal states. However, there are fundamental differences between
$KN$ and
$RB$ bimodal states that are illustrated by the subsequently discussed decomposition analysis.
We consider a bimodal equilibrium state vector
$\boldsymbol{b}(\unicode[STIX]{x1D716})$ depending on continuation parameter
$\unicode[STIX]{x1D716}$ as decomposition
where
$\unicode[STIX]{x1D6FC},\unicode[STIX]{x1D6FD}\in \mathbb{R}$ and
$\boldsymbol{m}_{1},\boldsymbol{m}_{2}$ are state vectors of two differently oriented straight convection rolls, and
$\boldsymbol{d}$ is the difference vector that is necessary to create the composite state
$\boldsymbol{b}$. We simplify the notation by suppressing the dependence of the decomposition on
$\unicode[STIX]{x1D716}$. We seek the optimal coefficients
$\unicode[STIX]{x1D6FC}$ and
$\unicode[STIX]{x1D6FD}$ such that
$\Vert \boldsymbol{d}\Vert _{2}$ is minimal. The optimality condition
$0=\unicode[STIX]{x2202}_{\unicode[STIX]{x1D6FC}}\Vert \boldsymbol{d}\Vert _{2}=\unicode[STIX]{x2202}_{\unicode[STIX]{x1D6FD}}\Vert \boldsymbol{d}\Vert _{2}$ implies that
$\langle \boldsymbol{m}_{1},\boldsymbol{d}\rangle =0$ and
$\langle \boldsymbol{m}_{2},\boldsymbol{d}\rangle =0$, where the inner product
$\langle ,\rangle$ is induced by the full norm (2.11). As a consequence,
$\unicode[STIX]{x1D6FC}$ and
$\unicode[STIX]{x1D6FD}$ may be found via inner products of (3.8) with
$\boldsymbol{m}_{\mathbf{1}}$ and
$\boldsymbol{m}_{\mathbf{2}}$, respectively,
These coupled equations determine the optimal coefficients
$\unicode[STIX]{x1D6FC}$ and
$\unicode[STIX]{x1D6FD}$. The corresponding minimal
$\boldsymbol{d}$ measures nonlinear and non-bimodal effects.
Decomposition of bimodal ribbons into left and right oriented oblique rolls (
$OR^{l,r}$). The bifurcation diagram with details of the decomposition (left) is shown together with visualizations of the temperature and velocity fields on the right (red labels in the diagram indicate panels on the right). The
$RB$ state bifurcates together with
$OR^{l,r}$ in equivariant pitchfork bifurcations from the laminar base flow. At
$\unicode[STIX]{x1D6FE}=40^{\circ }$,
$RB$ emerges with one additional unstable eigendirection, and
$OR^{l,r}$ become dynamically stable in a
$[2\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$-periodic domain at
$\unicode[STIX]{x1D716}=0.282$ (solid line in top left panel). The optimal decomposition
$\boldsymbol{x}_{RB}(\unicode[STIX]{x1D716})=\unicode[STIX]{x1D6FC}(\unicode[STIX]{x1D716})\boldsymbol{x}_{OR}^{l}(\unicode[STIX]{x1D716})+\unicode[STIX]{x1D6FD}(\unicode[STIX]{x1D716})\boldsymbol{x}_{OR}^{r}(\unicode[STIX]{x1D716})+\boldsymbol{d}(\unicode[STIX]{x1D716})$ implies linearly decreasing equal coefficients
$\unicode[STIX]{x1D6FC}=\unicode[STIX]{x1D6FD}$ along the branches (middle panel). The difference vector
$\boldsymbol{d}$ grows linearly in
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ from zero at the bifurcation point (bottom panel). Thus,
$RB$ can be viewed as a bimodal state combining two equally contributing oblique rolls.

The optimal bimodal decomposition (3.8) with (3.9) is calculated for
$\boldsymbol{b}=\boldsymbol{x}_{RB}$ and
$\boldsymbol{m}_{1,2}=\boldsymbol{x}_{OR}^{l,r}$ along the
$\unicode[STIX]{x1D716}$-bifurcation branches at
$\unicode[STIX]{x1D6FE}=40^{\circ }$. The coefficients are found to be equal at all
$\unicode[STIX]{x1D716}$, and to decrease linearly from
$\unicode[STIX]{x1D6FC}=\unicode[STIX]{x1D6FD}=0.3291$ at the bifurcation point (figure 10). The difference vector
$\boldsymbol{d}$ increases linearly in
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ and mostly accounts for corrections to the flow at the streamwise interfaces between hot and cold plumes (figure 10d).
Decomposition of knots into longitudinal and transverse oriented rolls,
$LR$ and
$TR$, respectively. The bifurcation diagram with details of the decomposition (left) is shown together with visualizations of the temperature and velocity fields on the right (red labels in the diagram indicate panels on the right). The
$KN$ state connects
$TR$ and
$LR$ at
$\unicode[STIX]{x1D6FE}=80^{\circ }$ (top). Dynamical stability in a
$[\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$-periodic domain is indicated by the solid lines. Here ‘RS20’ labels the control parameter value where the temporal evolution has been studied in Reference Reetz and SchneiderRS20. Between the bifurcation points, the optimal decomposition
$KN(\unicode[STIX]{x1D716})=\unicode[STIX]{x1D6FC}(\unicode[STIX]{x1D716})TR(\unicode[STIX]{x1D716}=0.024)+\unicode[STIX]{x1D6FD}(\unicode[STIX]{x1D716})LR(\unicode[STIX]{x1D716}=0.22)+\boldsymbol{d}(\unicode[STIX]{x1D716})$ results in decreasing
$\unicode[STIX]{x1D6FC}$, increasing
$\unicode[STIX]{x1D6FD}$ (middle) and parabolically varying
$\boldsymbol{d}$ (bottom). Thus,
$KN$ can be viewed as a bimodal state combining transverse and longitudinal rolls with monotonically changing relative contributions.

The optimal bimodal decomposition (3.8) with (3.9) is calculated again for
$\boldsymbol{b}=\boldsymbol{x}_{KN}$ along the
$KN$ branch at
$\unicode[STIX]{x1D6FE}=80^{\circ }$. Since
$LR$ does not coexist with most of the
$KN$ branch (figure 11), the state vectors
$\boldsymbol{m}_{1}=\boldsymbol{x}_{TR}$ and
$\boldsymbol{m}_{2}=\boldsymbol{x}_{LR}$ in the decomposition are not considered as
$\unicode[STIX]{x1D716}$-dependent. We choose the decomposition
$\boldsymbol{x}_{KN}(\unicode[STIX]{x1D716})=\unicode[STIX]{x1D6FC}(\unicode[STIX]{x1D716})\boldsymbol{x}_{TR}(\unicode[STIX]{x1D716}=0.024)+\unicode[STIX]{x1D6FD}(\unicode[STIX]{x1D716})\boldsymbol{x}_{LR}(\unicode[STIX]{x1D716}=0.22)+\boldsymbol{d}(\unicode[STIX]{x1D716})$. Here,
$\unicode[STIX]{x1D716}$ parametrizes linear interpolation between two bifurcation points. The resulting optimal coefficients
$\unicode[STIX]{x1D6FC}$ and
$\unicode[STIX]{x1D6FD}$ in general differ. While the contribution of the longitudinal rolls monotonically increases, the contribution of the transverse rolls decreases (figure 11). A decomposition with
$\unicode[STIX]{x1D6FC}=\unicode[STIX]{x1D6FD}$ is found at
$\unicode[STIX]{x1D716}=0.095$, approximately half-way between the bifurcation points and close to the maximum of
$\Vert \boldsymbol{d}\Vert _{2}$. Note that
$\boldsymbol{d}$ at the maximum amplitude resembles a ribbon pattern (figure 11d). Towards the bifurcation points,
$\boldsymbol{d}$ decreases parabolically to zero. Since
$\boldsymbol{d}$ combines nonlinear and non-bimodal effects, as well as effects due to interpolation between
$TR$ and
$LR$ at fixed values of
$\unicode[STIX]{x1D716}$, the dominant source for the large values of
$\unicode[STIX]{x2202}\Vert \boldsymbol{d}\Vert _{2}/\unicode[STIX]{x2202}\unicode[STIX]{x1D716}$ at the bifurcations is unclear. Analysis in the weakly nonlinear regime near the knot instability indicates a three-mode interaction (Fujimura & Kelly Reference Fujimura and Kelly1993; Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016), suggesting that beside longitudinal and transverse states a third resonant oblique state is involved. Here, the importance of an oblique mode is evidenced by the significant contribution of
$\boldsymbol{d}$ (figure 11).
The states
$KN$ and
$RB$ differ in their bifurcation structure. The
$KN$ state is a connecting state between
$TR$ and
$LR$, while
$RB$ bifurcates together with
$OR^{l,r}$ in an equivariant pitchfork bifurcation. As a consequence, their bimodal decomposition into two composing straight convection rolls differs significantly. The
$RB$ state is composed of an equal weight superposition of two symmetry related oblique rolls
$OR^{l}=\unicode[STIX]{x03C0}_{y}OR^{r}$. The
$KN$ state in ILC at
$\unicode[STIX]{x1D6FE}\neq 0^{\circ }$ is a mixed mode state, composed of transverse and longitudinal rolls that are not symmetry related and whose weight continuously changes along the branch. At
$\unicode[STIX]{x1D6FE}=0^{\circ }$ however,
$TR$ and
$LR$ become symmetry related via rotation. The knot patterns observed in Rayleigh–Bénard convection (Busse & Whitehead Reference Busse and Whitehead1974) are thus expected to bifurcate in equivariant pitchfork bifurcations, like
$RB$ in ILC.
3.5 Transverse oscillations – continuation towards a chaotic state space
The pattern of obliquely modulated transverse rolls, called ‘switching diamond panes’, shows complex dynamics with chaotically switching pattern orientations (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000). A periodic orbit
$TO$ underlying transverse oscillations has been identified in Reference Reetz and SchneiderRS20 at moderate
$\unicode[STIX]{x1D716}$. The pattern of transverse oscillations seems to capture some aspects of the observed complex dynamics. All
$\unicode[STIX]{x1D716}$ continuations of
$TO$ show that the orbit period of
$TO$ is subject to large and non-monotonic changes, and the number of unstable eigendirections of
$TO$ increases quickly with
$\unicode[STIX]{x1D716}$. This suggests the existence of complex state space structures that support the chaotic dynamics of switching diamond panes.
3.5.1 Bifurcations to transverse oscillations
In all but one analysed parameter continuations, the pre-periodic orbit
$TO$ bifurcates from
$TR$ in a supercritical Hopf bifurcation. The bifurcations are either
$\unicode[STIX]{x1D716}$-forward at
$\unicode[STIX]{x1D716}\approx 0.07$, found at inclination angles
$\unicode[STIX]{x1D6FE}=[80^{\circ },90^{\circ },100^{\circ },110^{\circ }]$, or
$\unicode[STIX]{x1D6FE}$-backward, found at
$\unicode[STIX]{x1D716}=0.1$ and
$\unicode[STIX]{x1D6FE}=132.2^{\circ }$. The latter case represents the upper inclination limit of existence of
$TO$ at
$\unicode[STIX]{x1D716}=0.1$. At the lower limit, just below
$\unicode[STIX]{x1D6FE}_{c2}$,
$TO$ bifurcates as a quaternary state from the transverse subharmonic varicose state
$TSV$ (see inset panel in figure 4). The
$TSV$ state is an equilibrium state discussed briefly in appendix A. A common feature of all
$TO$ branches is that the
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$-maximum over the orbit period remains close to the
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$-value of
$TR$. This agrees with the observation that
$TO$ modulations are sinusoidal oscillations around strictly transverse rolls with the maximum deflection associated to the minimum in
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$ over the orbit period (Reference Reetz and SchneiderRS20).
Continuation of
$TO$ at
$\unicode[STIX]{x1D6FE}=100^{\circ }$ towards a chaotic state space. (a) Midplane temperature of
$TO$ at
$Ra=11\,000$ (
$\unicode[STIX]{x1D716}=0.203$) before instabilities create a turbulent flow (b). The corresponding simulated time series of
$TO$ (d) shows the transition from a periodic to a turbulent signal. The transition is a consequence of the many linear instabilities of
$TO$ (f) that have emerged along the bifurcation branch (c). Solid/dotted bifurcation branches indicate stable and unstable states in the symmetry subspace of
$S=\langle \unicode[STIX]{x03C0}_{y},\unicode[STIX]{x03C0}_{xz},\unicode[STIX]{x1D70F}(0.5,0.5)\rangle$. Here ‘RS20’ labels the control parameter value at which the temporal evolution has been studied in Reference Reetz and SchneiderRS20. The changing relative period
$T^{\prime }$ of
$TO$ along the bifurcation branch is shown in (e) and compared to the time scale of thermal diffusion
$H^{2}/\unicode[STIX]{x1D705}$ and laminar mean advection
$L_{x}/\bar{U}_{0}$ (dashed and dashed–dotted lines).

3.5.2 Numerical continuation of transverse oscillations
The
$\unicode[STIX]{x1D6FE}$ continuation of
$TO$ at
$\unicode[STIX]{x1D716}=0.1$ is numerically straight forward and yields periodic orbits showing weak bending modulations around a purely transverse orientation (Reference Reetz and SchneiderRS20, § 4.2.2). All
$\unicode[STIX]{x1D716}$ continuations are found to be numerically challenging for increasing
$\unicode[STIX]{x1D716}$. We could not continue
$TO$ much beyond
$\unicode[STIX]{x1D716}=0.2$ (figure 12c). The reason for the computational difficulty is two-fold. First, the time period of the orbit drastically changes with
$\unicode[STIX]{x1D716}$, which causes challenges for our shooting method. The pre-periodic orbit
$TO$ satisfying (2.7) with
$\unicode[STIX]{x1D70E}=\unicode[STIX]{x1D70F}(0.5,0)$ oscillates slowly with a relative period
$T^{\prime }\sim O(10^{2})$ close to the heat diffusion time
$H^{2}/\unicode[STIX]{x1D705}$ and the laminar mean advection time
$L_{x}/\bar{U}_{0}$. Along the continuation, the large orbit period is subject to significant and non-monotonic changes over small
$\unicode[STIX]{x1D716}$-intervals (figure 12e). These changes in the orbit period are numerically difficult to trace. Secondly, the iterative solver of the Newton algorithm converges better if the target state is dynamically stable or weakly unstable (Sanchez et al. Reference Sanchez, Net, Garcıia-Archilla and Simo2004). Computing the spectrum of eigenvalues of
$TO$ in the symmetry subspace of
$[12\unicode[STIX]{x1D706}_{x},6\unicode[STIX]{x1D706}_{y}]$-periodicity indicates that the state branch at
$\unicode[STIX]{x1D6FE}=100^{\circ }$ and
$\unicode[STIX]{x1D716}=0.203$ has collected 63 unstable eigenvalues with a broad range of frequencies
$\unicode[STIX]{x1D714}_{i}$ (figure 12f). At these parameters, the single-shooting Newton algorithm converged
$TO$ to a residual of
$\Vert {\mathcal{G}}(\boldsymbol{x})\Vert _{2}<3\times 10^{-11}$ (see (2.11)). When integrating the converged orbit forward in time, unstable directions trigger a transition to a turbulent state after
$t=500$ (figure 12d). This turbulent state has been described as longitudinal bursts within switching diamond panes (Daniels et al. Reference Daniels, Plapp and Bodenschatz2000). We conclude that continuation of
$TO$ for
$\unicode[STIX]{x1D716}>0.2$ is challenging due to the numerical condition of a temporally slow, spatially large and very unstable periodic orbit that competes with many fast and small-scale modes in a chaotic turbulent state space.
4 Discussion
Towards understanding how temporal and spatio-temporally complex dynamics arises in ILC, we have computed three-dimensional invariant states underlying several observed spatially periodic convection patterns in ILC at
$Pr=1.07$. Numerical continuation of these invariant states in two control parameters, the normalized Rayleigh number
$\unicode[STIX]{x1D716}$ and the inclination angle
$\unicode[STIX]{x1D6FE}$, yields 15 bifurcation diagrams covering systematically selected parameter sections in the intervals
$\unicode[STIX]{x1D716}\in [0,2]$ and
$\unicode[STIX]{x1D6FE}\in [0^{\circ },120^{\circ })$. For some selected bifurcating state branches, we have characterized their stability properties and pattern features along the branches. These state branches were selected for a more detailed discussion in the present article for two reasons. First, each selected branch bifurcates at a different secondary instability. Second, they cover the control parameter values at which the temporal dynamics along dynamical connections between stable and unstable invariant states have previously been described (Reference Reetz and SchneiderRS20).
The relevance of the computed invariant states for observed spatio-temporally complex dynamics in ILC depends in general on the type of bifurcation creating the states, the range in control parameters over which state branches exist, and the stability properties of the invariant states along their branches. The dynamical relevance of invariant states in the context of the entire bifurcation structure is discussed below by answering the three specific questions posed in the introduction (Q1–Q3). To describe the role of individual invariant states for temporal pattern dynamics, we can distinguish three different cases.
Case 1: a stable invariant state represents a dynamical attractor at specific control parameter values. This case corresponds, for example, to supercritical
$\unicode[STIX]{x1D716}$-forward bifurcations where the stable bifurcating invariant state is an attractor for the dynamics above the critical values of the control parameters for the bifurcation. For invariant states that have been identified because they represent dynamical attractors at specific control parameter values (Reference Reetz and SchneiderRS20), the present bifurcation analysis indeed confirms supercritical
$\unicode[STIX]{x1D716}$-forward bifurcations (§§ 3.2–3.5).
Case 2: invariant states exist at specific control parameter values but are dynamically unstable. State branches are only stable over a finite range in control parameters. This range is limited by instabilities along state branches. Invariant states which the present study indicates as dynamically unstable at specific values of the control parameters, may still be relevant for the observed temporal dynamics at these control parameter values. One reason is that the range of stability along state branches depends on the considered pattern wavelength. Thus, invariant states might be dynamically stable at other pattern wavelengths not considered here. Another reason is that weakly unstable invariant states may be building blocks for the dynamics supported by a more complex state space attractor. Here, the evolving state vector may transiently visit weakly unstable invariant states by approaching and escaping along their stable and unstable manifolds, respectively (e.g. Suri et al. Reference Suri, Tithof, Grigoriev and Schatz2017). The simplest example for such complex state space attractors is the robust heteroclinic cycle between two weakly unstable instances of symmetry related
$OWR$ described in Reference Reetz and SchneiderRS20.
Case 3: invariant states do not exist at specific control parameter values but their pattern is reminiscent in some state space regions that may be transiently visited by the dynamics. Folds or symmetry-breaking bifurcations may limit the existence of invariant states in parameter space. However, the pattern of the invariant state may still emerge transiently at control parameter values beyond the existence limits. We have observed this case for the transient skewed varicose pattern along a dynamical connection from unstable to stable straight convection rolls in Rayleigh–Bénard convection (§ 3.1), as well as for transient subharmonic oscillations at
$[\unicode[STIX]{x1D716},\unicode[STIX]{x1D6FE}]=[1.5,17^{\circ }]$ (see § 4.2.1 in Reference Reetz and SchneiderRS20) where the
$SSW$ branch does not exist anymore due to a fold (figure 6c). The state space structure supporting such transient dynamics seems related to a state space structure supporting intermittency (Pomeau & Manneville Reference Pomeau and Manneville1980).
Consequently, the patterns of invariant states are often observed because invariant states are stable and attracting, but neither stability nor existence of invariant states is required for observing their pattern.
4.1 Bifurcation types (Q1)
Bifurcations create or destroy invariant states and change the stability along state branches. Thus, bifurcation structures describe how state space structures change across control parameters. In response to question Q1, stated in the introduction, we list all the different types of bifurcations found in the present study and refer to particular examples. Identified bifurcation types are included below.
(i) Pitchfork bifurcation, e.g. from
$TR$ or
$LR$ to
$KN$ along
$\unicode[STIX]{x1D716}$ at
$\unicode[STIX]{x1D6FE}=80^{\circ }$ (figure 11).(ii) Equivariant pitchfork bifurcation, e.g. from
$B$ to
$RB$ and
$OR$ along
$\unicode[STIX]{x1D716}$ at
$\unicode[STIX]{x1D6FE}=40^{\circ }$ (figure 10).(iii) Hopf bifurcation, e.g. from
$TR$ to
$TO$ along
$\unicode[STIX]{x1D716}$ at
$\unicode[STIX]{x1D6FE}=100^{\circ }$ (figure 12).(iv) Equivariant Hopf bifurcation, e.g. from
$LR$ to
$SSW$ and
$STW$ along
$\unicode[STIX]{x1D6FE}$ at
$\unicode[STIX]{x1D716}=1.5$ (figure 6).(v) Saddle-node bifurcation, e.g.
$WR$ along
$\unicode[STIX]{x1D716}$ at
$\unicode[STIX]{x1D6FE}=90^{\circ }$ (figure 3, panel
$\unicode[STIX]{x1D6FE}=90^{\circ }$).(vi) Mutual annihilation of two periodic orbits, e.g. the two folds bounding the
$SSW$ isola along
$\unicode[STIX]{x1D6FE}$ at
$\unicode[STIX]{x1D716}=0.5$ (figure 4).(vii) The global bifurcation of a periodic orbit colliding with a structurally robust heteroclinic cycle, e.g. the
$SSW$ collision with
$TR\rightarrow \unicode[STIX]{x1D70F}_{x}TR\rightarrow TR$ along
$Ra$ at
$\unicode[STIX]{x1D6FE}=10^{\circ }$ (figure 8).
The symmetry-breaking pitchfork and Hopf bifurcations are found as
$\unicode[STIX]{x1D716}$- or
$\unicode[STIX]{x1D6FE}$-forward or -backward bifurcations. The orientation of bifurcations can change when control parameter values are changed, e.g.
$WR$ bifurcates
$\unicode[STIX]{x1D6FE}$-forward from
$LR$ at
$\unicode[STIX]{x1D716}=0.1$, but
$\unicode[STIX]{x1D6FE}$-backward at
$\unicode[STIX]{x1D716}=0.5$ (figure 4). Moreover, pitchfork and Hopf bifurcations can be supercritical or subcritical independent of their orientation. The
$\unicode[STIX]{x1D716}$-backward pitchfork bifurcation from
$R_{\unicode[STIX]{x1D706}2}$ to
$SV$ at
$\unicode[STIX]{x1D6FE}=0^{\circ }$ is subcritical (figure 5d), but the
$\unicode[STIX]{x1D716}$-backward pitchfork bifurcation from
$OR$ to
$DWR$ at
$\unicode[STIX]{x1D6FE}=40^{\circ }$ is supercritical (figure 9f).
The sequential order in which bifurcations occur may depend on the considered path through parameter space. The
$RB$ state at
$\unicode[STIX]{x1D716}=0.5$, for example, can bifurcate in primary or secondary bifurcations along
$\unicode[STIX]{x1D6FE}$. When decreasing
$\unicode[STIX]{x1D6FE}$ towards
$\unicode[STIX]{x1D6FE}=46^{\circ }$,
$RB$ bifurcates from
$B$ in a primary bifurcation. When increasing
$\unicode[STIX]{x1D6FE}$ towards
$\unicode[STIX]{x1D6FE}=24^{\circ }$,
$RB$ bifurcates from
$TR$ in a secondary bifurcation (figure 4, panel
$\unicode[STIX]{x1D716}=0.5$). Thus, describing, for example,
$WR$ as a tertiary state implies a particular parameter path. Since
$WR$ can bifurcate from
$RB$ that may be described as a tertiary state (figure 3, panel
$\unicode[STIX]{x1D6FE}=50^{\circ }$),
$WR$ may also be described as a quaternary state.
The relation between bifurcation structures and spatio-temporally complex dynamics is in general complicated. The various local and global bifurcations can modify the coexisting invariant states and their dynamical connections in various ways. Coexistence of invariant states may result from supercritical or subcritical bifurcations as well as from folds. These bifurcation types exist in ILC at all angles of inclinations. For example, the subcritical coexistence of stable straight convection rolls with unstable
$SV$ (figure 5) or with unstable
$SSW$ (figure 3, panels
$\unicode[STIX]{x1D6FE}=10^{\circ },20^{\circ }$) supports the experimental observation of spatially localized variants of these spatially periodic states (Bodenschatz et al. Reference Bodenschatz, Pesch and Ahlers2000; Daniels et al. Reference Daniels, Plapp and Bodenschatz2000). The supercritical coexistence of
$WR$ with
$DWR$,
$OR$ or
$OWR$ (figure 9) supports the observed pattern defects within the spatially coexisting wavy rolls of different orientations (Daniels & Bodenschatz Reference Daniels and Bodenschatz2002; Daniels et al. Reference Daniels, Brausch, Pesch and Bodenschatz2008). The details of these relations are non-trivial as they require the consideration of spatial dynamics (e.g. Knobloch Reference Knobloch2015).
For a specific bifurcation structure, we see a generic relation to complex temporal dynamics. All computed sequences of primary and secondary supercritical
$\unicode[STIX]{x1D716}$-forward pitchfork or Hopf bifurcations give rise to one of the four sequences of dynamical connections. These are
$B\rightarrow LR\rightarrow SSW,WR$ and
$B\rightarrow TR\rightarrow KN,TO$ as observed in Reference Reetz and SchneiderRS20 and illustrated in figure 2. Consequently, a ‘sequence of bifurcations’ (Busse & Clever Reference Busse, Clever and Riahi1996) that consists of supercritical
$\unicode[STIX]{x1D716}$-forward bifurcations, gives rise to a corresponding ‘sequence of dynamical connections’.
4.2 Connection to instabilities (Q2)
The patterns of the tertiary invariant states
$SV$,
$SSW/STW$,
$WR$,
$KN$ and
$TO$ are similar to the pattern motifs associated to the five secondary instabilities in ILC at
$Pr=1.07$ (Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016). The similarity suggests that the invariant states bifurcate at corresponding secondary instabilities. To confirm this, we compare the bifurcation points of the nonlinear state branches with the critical threshold parameters of the secondary instabilities determined by Subramanian et al. (Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016) using Floquet analysis (compare with question Q2 stated in the introduction). Floquet analysis solves for the pattern wavelengths that first become unstable at the critical threshold
$\unicode[STIX]{x1D716}_{c}$ when
$\unicode[STIX]{x1D716}$ is increased towards
$\unicode[STIX]{x1D716}_{c}$ for fixed
$\unicode[STIX]{x1D6FE}$. For the numerical continuation of invariant states, the pattern wavelength is prescribed. The critical threshold
$\unicode[STIX]{x1D716}_{c}$ is determined by continuing the state branch down in
$\unicode[STIX]{x1D716}$ towards the bifurcation at
$\unicode[STIX]{x1D716}_{c}$ for fixed
$\unicode[STIX]{x1D6FE}$. Consequently, Floquet analysis yields the minimal
$\unicode[STIX]{x1D716}_{c}$ of the instability, while branches of invariant states at prescribed wavelengths bifurcate at higher
$\unicode[STIX]{x1D716}_{c}$. We expect comparable critical thresholds between the two methods if the associated pattern wavelengths are comparable.
Critical thresholds and wavelengths
$[l_{x},l_{y}]$ of secondary instabilities determined by Floquet analysis (Subramanian et al. Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016), and the critical bifurcation points determined by the present bifurcation analysis for prescribed box sizes
$[L_{x},L_{y}]$. The comparison requires to state all critical frequencies
$\unicode[STIX]{x1D714}_{c}^{\dagger }$ in diffusion time scales. Frequencies in free fall time units, used throughout this paper series, are obtained via
$\unicode[STIX]{x1D714}=\unicode[STIX]{x1D714}^{\dagger }/\sqrt{PrRa}$.

In table 2 we compare the results of Floquet analysis and bifurcation analysis in terms of pattern wavelengths
$L_{x}$ and
$L_{z}$, critical thresholds
$\unicode[STIX]{x1D716}_{c}$ and critical frequency
$\unicode[STIX]{x1D714}_{c}$ for Hopf bifurcations. We find clear agreement between the results for skewed varicose, longitudinal subharmonic oscillatory and knot instabilities. Note that Floquet analysis finds the skewed varicose instability for
$\unicode[STIX]{x1D6FE}=0^{\circ }$ at a slightly higher
$\unicode[STIX]{x1D716}_{c}$ than the bifurcation analysis. This suggests that the Floquet analysis did not capture the most unstable wavelengths of the skewed varicose instability.
For the wavy instabilities, the
$\unicode[STIX]{x1D716}_{c}$ obtained from the bifurcation analysis is significantly larger. This discrepancy results from the difference in the streamwise wavelength. Floquet analysis indicates
$l_{x}$ one order of magnitude larger than the
$L_{x}$ prescribed in the bifurcation analysis. We confirmed that
$WR$ bifurcates at identical
$\unicode[STIX]{x1D716}_{c}$ when identical pattern wavelengths are prescribed. Thus, the equilibrium state
$WR$ bifurcates at the previously characterized wavy instability.
For the transverse oscillatory instability, the two methods agree in
$\unicode[STIX]{x1D716}_{c}$ but differ in the critical frequency
$\unicode[STIX]{x1D714}_{c}$. The reasons for this discrepancy are not clear. Continuing the periodic orbit
$TO$ to identical pattern wavelengths does not change the critical frequency much. Thus, we hypothesize that the instability characterized by Floquet analysis corresponds to a different bifurcating periodic orbit as
$TO$. This hypothesis is supported by two observations. First, a Galerkin projection near the transverse oscillatory instability suggests a subcritical
$\unicode[STIX]{x1D716}$-backward bifurcation (Subramanian, Pesch & Schneider Reference Subramanian, Pesch and Schneider2015). The
$TO$ state, however, is always found to bifurcate supercritically and
$\unicode[STIX]{x1D716}$-forward. Second, the pattern of
$TO$ can be described as spatially subharmonic standing wave oscillations. Like the subharmonic standing wave state
$SSW$, also
$TO$ oscillates on the time scale of the laminar mean advection across the pattern
$L_{x}/\bar{U}_{0}$ (§ 2.1). Thus, these subharmonic standing waves satisfy the approximate resonance condition
with
$(m,n)\in \mathbb{N}$. For bifurcations to
$SSW$ at small
$\unicode[STIX]{x1D6FE}$, this approximation holds for
$(m,n)=(2,1)$ with relative errors of about
$\pm 15\,\%$. The nonlinear time scales along the
$SSW$ branch are shown in figure 6(d). For bifurcations to
$TO$, this approximation holds for
$(m,n)=(2,1)$ with relative errors of less than
$\pm 10\,\%$. The nonlinear time scales along the
$TO$ branch are shown in figure 12(e). The transverse oscillatory instability from Floquet analysis, however, satisfies (4.1) best for
$(m,n)=(4,1)$. Due to the different resonance numbers, we suspect other physics than those of subharmonic standing waves to govern the instability described by Floquet analysis. Future research should investigate the possibility for other periodic orbits than
$TO$ to bifurcate, possibly subcritically, at or near the transverse oscillatory instability. Except in the case of
$TO$, the bifurcating invariant states match the characteristics of the secondary instabilities described previously in Subramanian et al. (Reference Subramanian, Brausch, Daniels, Bodenschatz, Schneider and Pesch2016).
Existence regions of invariant states in the considered parameter space at
$Pr=1.07$. Local symmetry-breaking bifurcations are marked and connected to guide the eye:
$SV$ (
),
$WR$ (○) and
$KN$ (♢) bifurcate in pitchfork bifurcations. The
$SSW/STW$ (▫) and
$TO$ (
$\star$) states bifurcate in Hopf bifurcations. Where existence is not limited by symmetry-breaking bifurcations, the limits are due to folds, global bifurcations or, in the case of
$TO$, due to numerical challenges. Hatched regions of existence allow us to better distinguish
$SSW$ (
$\times$) and
$STW$ (
$+$).

4.3 Range of existence (Q3)
The third specific question is about the limits of existence of invariant solutions as control parameters are varied (Q3 stated in the introduction). This problem has been approached by continuing invariant states as far as possible along a priori defined sections across the
$[\unicode[STIX]{x1D6FE},\unicode[STIX]{x1D716}]$-parameter space at
$Pr=1.07$. Since the continuation methods allow tracing invariant states beyond critical threshold parameters of additional instabilities, it was possible to follow bifurcation branches over large intervals of control parameters. The travelling wave
$STW$, for example, is found to exist over a large range of inclinations
$10^{\circ }\leqslant \unicode[STIX]{x1D6FE}\leqslant 110^{\circ }$, covering different flow regimes with small and large laminar shear forces. We identify three invariant states, namely
$SSW$,
$STW$ and
$WR$, whose solution branches persist across the angle of the codimension-2 point
$\unicode[STIX]{x1D6FE}_{c2}$, and for
$\unicode[STIX]{x1D6FE}>90^{\circ }$ where their parent state
$LR$ has disappeared. With the exception of
$SV$, all tertiary invariant states are found to exist for the case of vertical convection with
$\unicode[STIX]{x1D6FE}=90^{\circ }$. All invariant states existing at
$[\unicode[STIX]{x1D6FE},\unicode[STIX]{x1D716}]=[90^{\circ },1.5]$ are briefly discussed and compared with turbulent vertical convection in appendix B. We visually summarize the regions of existence and coexistence of the computed invariant states in figure 13.
Continuation and stability analysis along state branches revealed bifurcations to or from other invariant states. These states are neither clearly observed in experiments or simulations, nor do they correspond to instabilities found by Floquet analysis. The stability analysis along the branch of
$WR$ (§ 3.3) introduced four additional equilibrium states, namely
$DWR$,
$OR$,
$OWR$ and
$RB$. Other invariant states were obtained because continuations terminated at bifurcations from these states. Specifically,
$TO$ may bifurcate from the
$TSV$ equilibrium (figure 4,
$\unicode[STIX]{x1D716}=0.1$),
$KN$ may bifurcate from
$SL$ (figure 3,
$\unicode[STIX]{x1D6FE}=90^{\circ }$), and a global bifurcation of
$SSW$ may involve LSV as the parent state (figure 6c). These three additional states are described in appendix A. We do not distinguish invariant states connected via folded bifurcation branches as upper and lower branch states. Folds exist at all angles of inclinations. See, e.g. the bifurcation branches of
$SV$ at
$\unicode[STIX]{x1D6FE}=0^{\circ }$, and of
$STW$ at
$\unicode[STIX]{x1D6FE}=110^{\circ }$. However, we observe that state branches tend to become more folded when the inclination angle approaches the vertical orientation (compare panels in figure 3).
5 Conclusions
The present bifurcation analysis has identified an extensive network of parametrically connected invariant state branches in inclined layer convection. Overall, 16 different nonlinear three-dimensional invariant states have been discussed in the present paper. Many of them are related to spatio-temporally complex dynamics observed in experiments and simulations. Seven different types of bifurcations were found, including common types like Hopf bifurcations or saddle-node bifurcations, and including less common types like equivariant bifurcations or global collisions between periodic orbits and robust heteroclinic cycles. Computing this many different invariant states and branches just for this work and Reference Reetz and SchneiderRS20 has been straightforward relative to recent integrated efforts of the research community to compute similar numbers of invariant states and branches in other canonical shear flows like plane Couette or pipe flow. Inclined layer convection covers flows from horizontal Rayleigh–Bénard convection to vertical layer convection, which are relevant for engineering applications and which have been widely studied using experiments and simulations. This paper demonstrates that these flows are numerically accessible to nonlinear dynamical systems concepts.
Acknowledgements
The authors are indebted to L. Tuckerman and E. Knobloch for in-depth discussions on the formalization of symmetries, equivariant bifurcations and the structural stability of heteroclinic cycles, as well as for valuable comments on the manuscript. P.S. acknowledges discussions with A. Rucklidge. This work was supported by the Swiss National Science Foundation (SNF) under grant no. 200021-160088, and in parts by a L’Oréal UK and Ireland Fellowship for Women in Science (PS).
Declaration of interests
The authors report no conflict of interest.
Three-dimensional flow structure and midplane temperature contours of three steady equilibrium states participating in bifurcations. (a) Longitudinal subharmonic varicose
$LSV$, (b) transverse subharmonic varicose
$TSV$, (c) subharmonic lambda plumes
$SL$.

Continuation yields eight invariant states in vertical layer convection (
$\unicode[STIX]{x1D6FE}=90^{\circ }$) at
$\unicode[STIX]{x1D716}=1.5$ (
$Ra=21\,266$). (a)
$TR$, (b/f) upper/lower branch of
$WR$, (c/g) upper/lower branch of
$SSW$, (d/h) upper/lower branch of
$STW$, (e)
$KN$. The distinction between upper/lower branches refers to
$\Vert \unicode[STIX]{x1D703}\Vert _{2}$. The upper branches of
$WR$ and
$SSW$ are closer to bifurcations from
$LR$ than the lower branches, for
$STW$vice versa. (i–l) Snapshots from a direct numerical simulation at
$\unicode[STIX]{x1D6FE}=90^{\circ }$ and
$\unicode[STIX]{x1D716}=1.5$ show transiently emerging structures resembling the invariant states above.

Appendix A. Additional invariant states participating in bifurcations
This section briefly discusses three additional invariant states that do not represent previously observed convection patterns at
$Pr=1.07$, but that participate in the bifurcation network by bifurcating to the above discussed invariant states.
The
$\unicode[STIX]{x1D6FE}$ continuation of
$SSW$ at
$\unicode[STIX]{x1D716}=1.5$ approaches an equilibrium state at
$\unicode[STIX]{x1D6FE}=1.9^{\circ }$ that we name LSV, short for longitudinal subharmonic varicose state. The pattern of the equilibrium resembles an instance in time along the orbit
$SSW$ at
$\unicode[STIX]{x1D6FE}=15^{\circ }$ (figure 14a). The LSV state is invariant under transformations of
$S_{\mathit{SSW}}=\langle \unicode[STIX]{x03C0}_{xyz},\unicode[STIX]{x1D70F}(0.5,0.5)\rangle$ and can be continued from
$\unicode[STIX]{x1D6FE}=0^{\circ }$ to
$\unicode[STIX]{x1D6FE}=30^{\circ }$ along which the relative position of the hot and cold plumes changes continuously.
Just below
$\unicode[STIX]{x1D6FE}_{c2}$, a
$\unicode[STIX]{x1D6FE}$-forward Hopf bifurcation generates
$TO$ from an equilibrium state, that we name transverse subharmonic varicose, at
$\unicode[STIX]{x1D716}=0.1$ (inset panel in figure 4). This invariant state represents stationary varicose modulations of the
$TR$ pattern with
$[3\unicode[STIX]{x1D706}_{x},3\unicode[STIX]{x1D706}_{y}]$-periodicity and invariance under transformations of
$S_{\mathit{TV}}=\langle \unicode[STIX]{x03C0}_{xz},\unicode[STIX]{x03C0}_{y}\rangle$ (figure 14b). The
$TSV$ state is similar to the spatially subharmonic state discussed in Clever & Busse (Reference Clever and Busse1995) but is not strictly subharmonic here since
$TSV$ is not invariant under
$\unicode[STIX]{x1D70F}(0.5,0.5)$.
Due to the absence of
$LR$ at finite
$Ra$ at
$\unicode[STIX]{x1D6FE}=90^{\circ }$,
$KN$ do not terminate in a bifurcation from
$LR$ as described in § 3.4 but from
$SL$, an equilibrium that we name subharmonic lambda plumes. The
$[\unicode[STIX]{x1D706}_{x},\unicode[STIX]{x1D706}_{y}]$-periodic
$SL$ emerges in a saddle-node bifurcation at
$\unicode[STIX]{x1D716}=1.670$. From the upper branch of
$SL$,
$KN$ bifurcate
$\unicode[STIX]{x1D716}$-backward at
$\unicode[STIX]{x1D716}=1.672$ (figure 3). The subharmonic flow structure of
$SL$ is invariant under transformations of
$S_{SL}=\langle \unicode[STIX]{x03C0}_{xz}\unicode[STIX]{x1D70F}(0,0.5),\unicode[STIX]{x03C0}_{y}\unicode[STIX]{x1D70F}(0.5,0),\unicode[STIX]{x1D70F}(0.5,0.5)\rangle$ and resembles lambda-shaped plumes at scales of half the gap height (figure 14c).
Appendix B. Invariant states in vertical layer convection
Inclined layer convection at
$\unicode[STIX]{x1D6FE}=90^{\circ }$ is a particular case because vertical layer convection has the largest laminar shear forces of all inclinations. Here, buoyancy provides a body force along the channel domain, acting like a pressure gradient in a pressure-driven channel flow. Despite the absence of a wall-normal buoyancy force, all main invariant states considered in the present study were numerically continued to
$\unicode[STIX]{x1D6FE}=90^{\circ }$, with the exception of
$LR$ existing only at
$Ra=\infty$ due to (3.2)–(3.6) and
$SV$. The flow structures of the invariant states at
$\unicode[STIX]{x1D6FE}=90^{\circ }$ and
$\unicode[STIX]{x1D716}=1.5$ show sharper interfaces and more pointed convective plumes than at the parameters where these states were initially found (figure 15).




































































































































































































