MSC codes. Primary: 86A10, 76U60; Secondary: 37N10, 53Z05, 65N99.
Keywords: Potential vorticity inversion, clouds, wave-vortical decomposition, inertia-gravity waves, water vapor, moisture, geophysical fluid dynamics, phase interface.
Note to the reader: in § 1 below we provide a “shortest path” to the result, providing just enough background to state and motivate the result. The main results and conclusions are summarized in § 7 and § 8 . For a more detailed discussion of the motivation and context we refer the reader to § 2 .
Introduction
Decomposing a vector field, or a state vector, into several constitutive components is ubiquitous in fluid mechanics and atmospheric science. For example, the Helmholtz decomposition will split a vector field into its potential and divergence-free components, and it is used throughout fluid mechanics [16 , 94 , 73 ] . Another example is the vortical–wave decomposition of a state vector which is used in atmospheric and oceanic sciences to understand the fast rotation and strong stratification regime of the Boussinesq equations, e.g., [11 , 4 , 72 ] . Crucially, as discussed in § 2 , both of these decompositions can be viewed from multiple perspectives and share salient features such as reconstruction of the original field via elliptic PDE, a rich geometric interpretation of the decomposition (involving orthogonality and projections), and a connection to slow and fast components of appropriate dynamical systems.
In this paper we consider a moist Boussinesq system from atmospheric dynamics, which brings additional realism and complexity due to clouds. We ask the following question: does this moist, cloudy Boussinesq system also admit such a decomposition?
While these decompositions have applications in slow–fast dynamical systems, they can be investigated without reference to the dynamical evolution, by considering the associated operator. To see this from a general perspective, consider a dynamical system written abstractly as
∂ t U + 1 ϵ D 0 U + D 1 U = 0 , \partial_t U + \frac{1}{\epsilon}\mathcal{D}_0 U + \mathcal{D}_1 U = 0, ∂ t U + ϵ 1 D 0 U + D 1 U = 0 ,
where ϵ \epsilon ϵ is a small parameter, D 0 \mathcal{D}_0 D 0 is the leading-order operator, and D 1 \mathcal{D}_1 D 1 is the next-order contribution to the operator for the dynamical system. To leading order with respect to ϵ \epsilon ϵ , it is the operator D 0 \mathcal{D}_0 D 0 that contains the essential information. Understanding the dynamics can be tantamount to understanding the operator D 0 \mathcal{D}_0 D 0 . Hence, in the remainder of the paper, we will seldom refer to the actual dynamical evolution, and instead the main object of interest is the operator D 0 \mathcal{D}_0 D 0 .
In past cases, such as the Helmholtz decomposition or Boussinesq equations, the leading-order operator is linear (D 0 = L \mathcal{D}_0 = \mathcal{L} D 0 = L ), and the decomposition involves a linear operator and linear eigenmodes. On the other hand, for an atmosphere with clouds, additional nonlinearity arises from phase changes of water, and D 0 = N \mathcal{D}_0 = \mathcal{N} D 0 = N is a nonlinear operator. Hence a substantial challenge in the present paper is from nonlinearity, from the distinction
D 0 = L versus D 0 = N . (1) \mathcal{D}_0 = \mathcal{L} \quad\text{versus} \quad\mathcal{D}_0 = \mathcal{N}. \tag*{(1)} D 0 = L versus D 0 = N . (1)
For a nonlinear operator, it is not clear if eigenmodes can be identified. Consequently, a decomposition
U = U 1 + U 2 + ⋯ + U N e i g (2) U = U_1 + U_2 + \cdots+ U_{N_{\mathrm{eig}}} \tag*{(2)} U = U 1 + U 2 + ⋯ + U N eig (2)
into N e i g N_{\mathrm{eig}} N eig eigenmodes is likely to be impossible.
Despite the challenges of nonlinearity in (1) –(2) , here we forge ahead with a decomposition in the spirit of (2) as the main aim and theme of the paper.
In § 2 we describe the background and main questions in more detail. § 3 , § 4 , § 5 , and § 6 describe main results, and § 7 and § 8 present a summary of main results and conclusions, including links to the main theorems for easy reference. We point the reader who may skip the background in § 2 to note § 2.5 where the simplified notation used in the main body of the paper is recorded.
Background and motivation
In this section we first expound on known decompositions (namely the Helmholtz decomposition and the vortical–wave decomposition of the dry Boussinesq system) in order to set the stage for the decomposition of moist Boussinesq states studied in this paper. We also motivate in more detail the context in which the dry and moist Boussinesq equations arise.
We start with the Helmholtz decomposition for two reasons. First, we present it here for folks familiar with fluid dynamics but not necessarily familiar with geophysical and atmospheric fluid dynamics. Second, we present it here because, as we will detail later, the Helmholtz decomposition plays a central role in the other decompositions, in both the dry and the moist cases.
Helmholtz decomposition
The Helmholtz decomposition of a vector field u u u splits it into two components:
u = ∇ p + σ u = \nabla p + \sigma u = ∇ p + σ
where σ \sigma σ is divergence-free. One of the main reasons for the ubiquity of this decomposition, especially in the realm of fluid mechanics, is that this single decomposition can be viewed through a variety of lenses. Here we recall how it can be viewed geometrically, viewed in relation to elliptic PDE, and viewed in relation to slow-fast dynamical systems. We emphasize these three perspectives because these are precisely the perspectives that coincide with each other in the dry case but diverge in the moist case.
We can view the Helmholtz decomposition geometrically: it decomposes the space of three-dimensional vector fields into two orthogonal pieces, namely potential vector fields and divergence-free vector fields, and thus each component of the decomposition can be understood as a projection onto the corresponding subspace. The Helmholtz decomposition is also related to elliptic PDE since the scalar potential p p p solves
Δ p = ∇ ⋅ u \Delta p = \nabla\cdot u Δ p = ∇ ⋅ u
while, if we write the solenoidal component as σ = ∇ × ψ + c \sigma= \nabla\times\psi+ c σ = ∇ × ψ + c , the vector potential ψ \psi ψ solves
∇ × ( ∇ × ψ ) = ∇ × u \nabla\times(\nabla\times\psi) = \nabla\times u ∇ × ( ∇ × ψ ) = ∇ × u
with c c c given by c = \fint u . Here we use the notation \fint f to denote the average over the spatial domain of any scalar or vector function f f f .
Finally we can view the Helmholtz decomposition as a slow-fast decomposition for an appropriate dynamical system. The relevant system is the low Mach number limit of the compressible Euler equations (for an ideal isentropic gas) [56 , 57 , 71 ] . In this setting the slow dynamics correspond to incompressible flow whereas the fast dynamics correspond to the acoustic wave equation. Crucially: the solenoidal component of the Helmholtz decomposition is precisely the slow piece and the potential component is precisely the fast piece. We postpone a detailed discussion of this slow-fast dynamics to Section A.1: this low Mach number limit is well-known, so we only provide details later once it helps us compare and contrast these classical results (pertaining to Helmholtz and dry Boussinesq) with the results we have obtained in this paper.
Dry Boussinesq and its decomposition
Now we turn our attention to the dry Boussinesq system and its decomposition. As we describe this decomposition we will highlight the properties that it shares with the Helmholtz decomposition discussed above.
In oceanic and atmospheric application problems, one may need a decomposition of observational or model data. The latter is often motivated by, or closely related to, normal-mode analysis of a PDE model such as the dry Boussinesq equations, e.g. [67 , 6 , 64 , 106 ] . Furthermore, a normal-mode decomposition is often the starting point for numerical computations [66 , 4 , 93 HMRP16, 33 , 99 ] . For the Boussinesq equations, the normal-mode analysis is also a wave-vortical decomposition, as further explained below. Although many numerical computations focused on constant rotation and buoyancy frequencies in periodic domains [4 , 93 HMRP16] , other analyses also address more realistic PDE systems, such as variable-stratification Boussinesq dynamics with rigid surface and bottom boundaries [33 ] and the shallow water equations on a sphere [54 , 99 ] .
Let’s stop and discuss this terminology for a second: why is the decomposition of dry Boussinesq states called a vortical-wave decomposition? It is called this because the decomposition has two components. The first component is characterised by the potential vorticity. This is why it is referred to as the “vortical mode”. (In the dry case this corresponds to an honest-to-goodness eigenmode. This is no longer true in the nonlinear moist case, which is why we use the term “component” instead of “mode” in that case.) The second component corresponds to the fast piece of the decomposition. It can then be shown that the dynamics obeyed by this fast piece are that of linear oscillations. This is why this component is known as the “wave mode”. Together the vortical mode and the wave mode thus constitute the vortical-wave decomposition.
The dry Boussinesq system and its fast rotation and strong stratification limit is one of the success stories of atmospheric science. In the associated parameter regime the dry Boussinesq system may be written in terms of conservation laws:
∂ t u + u ⋅ ∇ u + 1 ε ( e 3 × u + ∇ p − θ e 3 ) = 0 and ∂ t θ + u ⋅ ∇ θ + 1 ε u 3 = 0 (3) \begin{aligned}
\partial_t u + u \cdot\nabla u + \frac{1}{\varepsilon}(e_3 \times u + \nabla p - \theta e_3) &= 0 \quad\text{and} \\
\partial_t \theta+ u \cdot\nabla\theta+ \frac{1}{\varepsilon}u_3 &= 0 \tag*{(3)}
\end{aligned} ∂ t u + u ⋅ ∇ u + ε 1 ( e 3 × u + ∇ p − θ e 3 ) ∂ t θ + u ⋅ ∇ θ + ε 1 u 3 = 0 and = 0 (3)
∂ t θ + u ⋅ ∇ θ + 1 ε u 3 = 0 (4) \partial_t \theta+ u \cdot\nabla\theta+ \frac{1}{\varepsilon}u_3 = 0 \tag*{(4)} ∂ t θ + u ⋅ ∇ θ + ε 1 u 3 = 0 (4)
subject the incompressibility assumption ∇ ⋅ u = 0 \nabla\cdot u = 0 ∇ ⋅ u = 0 . Note that for simplicity physical parameters of order one have be set to unity. Here the unknowns are the velocity vector field u u u and the potential temperature θ \theta θ (a scalar field). The first equation corresponds to conservation of linear momentum while the second equation corresponds to the conservation of potential temperature. Of course, as written above (4) makes it look like θ \theta θ is not conserved. This is because there is a large background potential temperature gradient (this is a manifestation of the strong stratification assumption) and so the quantity θ \theta θ used above is actually the anomaly of the potential temperature about that background gradient.
Here the small parameter indicates that we are particularly interested in the behaviour of this system as the stratification is strong and the rotation is fast. In the limit we formally obtain from (4) that the vertical component of the velocity vanishes and we obtain from the (3) the geostrophic balance u h = ∇ h ⊥ p u_h = \nabla_h^\perp p u h = ∇ h ⊥ p and the hydrostatic balance θ = ∂ 3 p \theta= \partial_3p θ = ∂ 3 p . Here v h = ( v 1 , v 2 , 0 ) v_h = (v_1, v_2, 0) v h = ( v 1 , v 2 , 0 ) denotes the horizontal part of a vector v v v and v h ⊥ v_h^\perp v h ⊥ denotes its π 2 \frac{\pi}{2} 2 π –rotation v h ⊥ = ( − v 2 , v 1 , 0 ) v_h^\perp= (-v_2, v_1, 0) v h ⊥ = ( − v 2 , v 1 , 0 ) . Note that the velocity is horizontal since its vertical component vanishes, yet it still depends on all three spatial coordinates (in this sense the flow is not two-dimensional). The discussion is formal here, but rigorous theorems are available (see [EM96] and [76 ] for balanced and general initial data, respectively).
In particular, the slow dynamics are characterised by the potential vorticity P V = ∇ h ⊥ ⋅ u h + ∂ 3 θ PV = \nabla_h^\perp\cdot u_h + \partial_3\theta P V = ∇ h ⊥ ⋅ u h + ∂ 3 θ . This follows from two observations which hold in the limit: (1) we may compute from (3) –(4) that P V PV P V is purely advected by the flow and (2) the geostrophic and hydrostatic balances tell us that p p p solves Δ p = P V \Delta p = PV Δ p = P V . This is the crux of why the limiting Boussinesq dynamics on the slow timescale, known as the (dry) quasi-geostrophic equations, are a success story: everything comes down to a single conserved quantity, the potential vorticity, from which all other information about the system may be recovered (by way of inverting a Laplacian). The potential vorticity not only helps us understand the limit, it also allows us to decompose states of the original dry Boussinesq system into slow and fast components.
We now discuss this decomposition, also known as the vortical-wave decomposition. This decomposition splits the state-vector ( u , θ ) (u,\theta) ( u , θ ) into two pieces:
( u θ ) = ( ∇ h ⊥ p ∂ 3 p ) + ( σ h ⊥ + w e 3 σ 3 ) \begin{pmatrix}
u \\
\theta
\end{pmatrix}
=
\begin{pmatrix}
\nabla_h^\perp p \\
\partial_3p
\end{pmatrix}
+
\begin{pmatrix}
\sigma_h^\perp+ w e_3 \\
\sigma_3
\end{pmatrix} ( u θ ) = ( ∇ h ⊥ p ∂ 3 p ) + ( σ h ⊥ + w e 3 σ 3 )
where σ \sigma σ is divergence-free. The first component is known as the vortical mode and the second component is known as the wave mode . As in the case of the Helmholtz decomposition above, this decomposition has several equivalent interpretations. We can view this decomposition geometrically as projecting into two L 2 L^2 L 2 -orthogonal subspaces: the first piece is the set of balanced states satisfying both geostrophic and hydrostatic balance and the second piece is the set of states with vanishing potential vorticity. This is similar to the Helmholtz decomposition where the first piece, the potential component, satisfies the “balance” u = ∇ p u = \nabla p u = ∇ p while the second piece belongs to the set of states with vanishing divergence. Note that L 2 L^2 L 2 is used here, not by default, but because it is precisely the energy conserved by the dry Boussinesq system (3) –(4) . We can view this decomposition in relation to elliptic PDE since p p p solves
Δ p = P V . \Delta p = PV. Δ p = P V .
Meanwhile, since σ \sigma σ is divergence-free, we can write it in terms of a vector potential as σ = ∇ × ψ + c \sigma= \nabla\times\psi+ c σ = ∇ × ψ + c for some constant c c c where ψ \psi ψ solves
∇ × ( ∇ × ψ ) = ( j ∂ 3 w ) \nabla\times(\nabla\times\psi) =
\begin{pmatrix}
j \\
\partial_3 w
\end{pmatrix} ∇ × ( ∇ × ψ ) = ( j ∂ 3 w )
for j = ∂ 3 u h − ∇ h ⊥ θ j = \partial_3u_h - \nabla_h^\perp\theta j = ∂ 3 u h − ∇ h ⊥ θ denoting the thermal wind imbalance1 , c = − f u h + f θ e 3 c = -f u_h + f \theta e_3 c = − f u h + f θ e 3 , and finally w = u 3 w = u_3 w = u 3 . Finally we can view this decomposition as a slow-fast decomposition for (3) –(4) . Here we describe only cursorily how this may be done, postponing a more detailed discussion to § 4 . Indeed, the vortical mode corresponds precisely to solutions of the limiting quasi-geostrophic system while the wave mode corresponds to the component of the solution with time derivative proportional to 1 / ε 1/\varepsilon 1/ ε (and hence fast). We discuss in more detail why this decomposition works as a slow-fast decomposition in § 4 , where we recall how the fast-wave averaging framework allows us to make this idea precise in order to then use this framework to discuss the slow-fast decomposition in the moist case.
Moist Boussinesq
We have seen in § 2.1 and § 2.2 above that both the Helmholtz decomposition and the vortical-wave decomposition possess several interpretations, be they geometric, related to elliptic PDE, or about the slow-fast nature of appropriate dynamical systems. In this section we introduce the moist Boussinesq system, keeping in mind that our goal is to find a decomposition for that system which possesses properties akin to those discussed above.
First, we motivate the moist Boussinesq system itself. Its dry counterpart is potent for analytical, computational, and predictive purposes because it focuses on two physical features which dominate mid-latitude atmospheric dynamics: stratification and rotation (i.e. the Coriolis effect). The moist Boussinesq system introduces another important physical feature: moisture. Even without incorporating all of the complexities of the water cycle, moisture plays an important role in atmospheric dynamics because of the role it plays in the energy budget: as moist air cools, condensing water releases latent heat energy (and vice-versa as air warms and liquid water evaporates). When incorporating moisture into the dry system from {(2.1a)–(2.1b)}, we obtain
∂ t u + u ⋅ ∇ u + 1 ε ( e 3 × u + ∇ p − θ e 3 ) = 0 , ∂ t θ + u ⋅ ∇ θ + Γ θ ε u 3 = C , ∂ t q v + u ⋅ ∇ q v − Γ q ε u 3 = − C and ∂ t q l + u ⋅ ∇ q l = C . (5) \begin{aligned}
\partial_t u + u \cdot\nabla u + \frac{1}{\varepsilon}(e_3 \times u + \nabla p - \theta e_3) &= 0, \\
\partial_t \theta+ u \cdot\nabla\theta+ \frac{\Gamma_\theta}{\varepsilon}u_3 &= C, \\
\partial_t q_v + u \cdot\nabla q_v - \frac{\Gamma_q}{\varepsilon}u_3 &= -C \text{ and} \\
\partial_t q_l + u \cdot\nabla q_l &= C. \tag*{(5)}
\end{aligned} ∂ t u + u ⋅ ∇ u + ε 1 ( e 3 × u + ∇ p − θ e 3 ) ∂ t θ + u ⋅ ∇ θ + ε Γ θ u 3 ∂ t q v + u ⋅ ∇ q v − ε Γ q u 3 ∂ t q l + u ⋅ ∇ q l = 0 , = C , = − C and = C . (5)
In comparison to the dry Boussinesq system, the moist Boussinesq system includes additional evolution equations for the water vapor mixing ratio, q v q_v q v , in {(2.2c)} and for the liquid water mixing ratio, q l q_l q l , in {(2.2d)}. It also includes a source/sink term C C C for condensation and evaporation, and associated heating and cooling. Condensation occurs for C > 0 C > 0 C > 0 and represents a phase transition from vapor to liquid, and evaporation occurs for C < 0 C < 0 C < 0 and represents a phase transition from liquid to vapor. The constants Γ θ \Gamma_\theta Γ θ and Γ q \Gamma_q Γ q correspond to parameters encoding the strength of the background vertical gradients in potential temperature and water vapor, respectively. We also note that the precise definition of the water content q v q_v q v or q l q_l q l as a mixing ratio is measured as kilograms of total water (vapor plus liquid) per kilogram of dry air [58 , 47 ] . While additional cloud microphysics processes will not be considered here, the moist Boussinesq system above is valuable for more complex scenarios as well, since it provides the starting point for extensions that include rainfall, ice, other precipitation, and other complexities [46 , 58 , 92 , 101 , 100 ] .
Second, we rewrite the moist Boussinesq system in a way that facilitates the main goal here of a slow-fast state decomposition. In particular, note that the source terms C C C in {(2.2b)–(2.2d)} appear to potentially cause complications. However, the system can be rewritten so that the source terms do not explicitly appear, through a convenient change of variables. This reformulation has been used in several past studies [14 , 86 , 47 , 92 , 77 , 63 ] and has facilitated theoretical advances such as an energy decomposition [92 , 77 ] and conservation of potential vorticity [63 ] , even in the presence of clouds and phase changes.
For the convenient way of rewriting {(2.2a)–(2.2d)}, we transform to a different set of thermodynamic variables that is conserved. In particular, define the equivalent potential temperature, θ e \theta_e θ e , and total water mixing ratio, q t q_t q t , as
θ e = θ + q v , q t = q v + q l . (6) \begin{aligned}
\theta_e &= \theta+ q_v, \\
q_t &= q_v + q_l. \tag*{(6)}
\end{aligned} θ e q t = θ + q v , = q v + q l . (6)
Note that the equivalent potential temperature θ e \theta_e θ e may be understood physically as the potential temperature that a parcel of air would have if it were heated by converting all of its water vapour into liquid water. The evolution equations of θ e \theta_e θ e and q t q_t q t can be found by taking appropriate linear combinations of (2.2b)–(2.2d), and they take the form
{ ∂ t θ e + u ⋅ ∇ θ e + 1 ε u 3 = 0 , ∂ t q t + u ⋅ ∇ q t − 1 ε u 3 = 0 , (7) \left\{
\begin{aligned}
\partial_t\theta_e + u \cdot\nabla\theta_e + \frac{1}{\varepsilon}u_3 &= 0,\\
\partial_tq_t + u \cdot\nabla q_t - \frac{1}{\varepsilon}u_3 &= 0,
\end{aligned}
\right. \tag*{(7)} ⎩ ⎨ ⎧ ∂ t θ e + u ⋅ ∇ θ e + ε 1 u 3 ∂ t q t + u ⋅ ∇ q t − ε 1 u 3 = 0 , = 0 , (7)
where we have now set the background gradient parameters from (2.2b) and (2.2c) to be Γ θ = 2 \Gamma_\theta=2 Γ θ = 2 and Γ q = 1 \Gamma_q=1 Γ q = 1 without loss of generality within the stably stratified regime, and for simplicity. The important feature of (2.4a)–(2.4b) is that the source term of condensation and evaporation, C C C , has been eliminated. As a result, (2.4a)–(2.4b) show that θ e + ( z / ε ) \theta_e+(z/\varepsilon) θ e + ( z / ε ) and q t − ( z / ε ) q_t-(z/\varepsilon) q t − ( z / ε ) (which combine the anomalies θ e \theta_e θ e and q t q_t q t with their background profiles z / ε z/\varepsilon z / ε and − z / ε -z/\varepsilon − z / ε , respectively) are conserved along fluid parcel trajectories. Physically, the equivalent potential temperature is conserved because losses of water vapor q v q_v q v are compensated by gains in heat θ \theta θ , as indicated by the definition in (2.3a); and the total water mixing ratio, q t = q v + q l q_t=q_v+q_l q t = q v + q l , is conserved because losses of water vapor q v q_v q v are compensated by gains in liquid water q l q_l q l .
To complete the rewriting, the buoyancy term θ / ε \theta/\varepsilon θ / ε from (2.2a) must also be rewritten in terms of θ e \theta_e θ e and q t q_t q t . To do so, the following transformation is used:
{ θ = θ e − min ( q t , q v s ) , q v = min ( q t , q v s ) , q l = max ( 0 , q t − q v s ) , (8) \left\{
\begin{aligned}
\theta&= \theta_e-\min(q_t,q_{vs}),\\
q_v &= \min(q_t,q_{vs}),\\
q_l &= \max(0,q_t-q_{vs}),
\end{aligned}
\right. \tag*{(8)} ⎩ ⎨ ⎧ θ q v q l = θ e − min ( q t , q v s ) , = min ( q t , q v s ) , = max ( 0 , q t − q v s ) , (8)
which is the reverse of the transformation in (2.3a)–(2.3b). The saturation water vapor value q v s q_{vs} q v s appears here for distinguishing between the unsaturated phase (q v < q v s q_v<q_{vs} q v < q v s ), when no cloud is present, and the saturated phase (q v = q v s q_v=q_{vs} q v = q v s ), when a cloud is present. Note that q v s q_{vs} q v s is not appearing here for the first time. It is already implicit in the condensation/evaporation source term C C C in (2.2b)–(2.2d). The value of q v s q_{vs} q v s will be constant here for simplicity. For explicit expressions for θ \theta θ in each phase, one can rewrite (2.5a) as
θ = { θ e − q t if q t < q v s , θ e − q v s if q t ≥ q v s , (9) \theta=
\begin{cases}
\theta_e-q_t & \text{if } q_t<q_{vs},\\
\theta_e-q_{vs} & \text{if } q_t\ge q_{vs},
\end{cases} \tag*{(9)} θ = { θ e − q t θ e − q v s if q t < q v s , if q t ≥ q v s , (9)
which can be used for another expression of the buoyancy in terms of the variables θ e \theta_e θ e and q t q_t q t .
The end result of the rewriting is the moist Boussinesq system in the following form:
{ ∂ t u + u ⋅ ∇ u + 1 ε ( e 3 × u + ∇ p − ( θ e − min 0 q t ) e 3 ) = 0 , ∂ t θ e + u ⋅ ∇ θ e = − 1 ε u 3 , ∂ t q t + u ⋅ ∇ q t = 1 ε u 3 , (10) \left\{
\begin{aligned}
\partial_tu+u\cdot\nabla u+\frac{1}{\varepsilon}(e_3\times u+\nabla p-(\theta_e-\min_0 q_t)e_3)&=0,\\
\partial_t\theta_e+u\cdot\nabla\theta_e&=-\frac{1}{\varepsilon}u_3,\\
\partial_tq_t+u\cdot\nabla q_t&=\frac{1}{\varepsilon}u_3,
\end{aligned}
\right. \tag*{(10)} ⎩ ⎨ ⎧ ∂ t u + u ⋅ ∇ u + ε 1 ( e 3 × u + ∇ p − ( θ e − 0 min q t ) e 3 ) ∂ t θ e + u ⋅ ∇ θ e ∂ t q t + u ⋅ ∇ q t = 0 , = − ε 1 u 3 , = ε 1 u 3 , (10)
subject to ∇ ⋅ u = 0 \nabla\cdot u=0 ∇ ⋅ u = 0 , and following from (2.2a), (2.4a), (2.4b), and (2.5a). Some additional specifications are as follows. The notation min 0 q t = min ( 0 , q t ) \min_0q_t=\min(0,q_t) min 0 q t = min ( 0 , q t ) has been introduced in (2.7a), and the saturation water vapor parameter q v s q_{vs} q v s has been set to zero without loss of generality, and for simplicity. Throughout this paper we work on the three-dimensional torus (i.e. with triply-periodic boundary conditions). Also recall that we set the background gradient parameters from (2.2b) and (2.2c) to be Γ θ = 2 \Gamma_\theta=2 Γ θ = 2 and Γ q = 1 \Gamma_q=1 Γ q = 1 without loss of generality within the stably stratified regime, and for simplicity. One can now see the value in rewriting the moist Boussinesq system: while the ( θ , q v , q l ) (\theta,q_v,q_l) ( θ , q v , q l ) formulation in (2.2a)–(2.2d) is helpful for seeing the additional cloud physics that is added to the dry system, the ( θ e , q t ) (\theta_e,q_t) ( θ e , q t ) formulation in (2.7a)–(2.7c) is helpful for writing the moist system in a way that is mathematically similar to the dry system. The main differences between the moist and dry systems are that the moist system has two thermodynamic variables (θ e \theta_e θ e and q t q_t q t ) instead of one (θ \theta θ ), and the buoyancy term in the moist case involves the nonlinear expression θ e − min 0 q t \theta_e-\min_0q_t θ e − min 0 q t .
What makes the moist Boussinesq system particularly interesting is that moisture introduces phase boundaries . In concrete terms: here we have assumed without loss of generality that water saturation occurs at q t = 0 q_t=0 q t = 0 and so different physics are in play depending on the sign of q t q_t q t , i.e. depending on whether water is above or below saturation (this is discussed in more detail in § 2 ). We therefore refer to the set { q t = 0 } \{q_t=0\} { q t = 0 } as the phase boundary since it delineates the separation between water in vapour form (outside of a cloud) and water in vapour-plus-liquid form (inside of a cloud).
Finally, to end this subsection, we review some properties of the moist Boussinesq system that are potentially relevant to the aim here—i.e., to the search for a state decomposition.
As ε → 0 \varepsilon\to0 ε → 0 , solutions of the moist Boussinesq system (2.7a)–(2.7c) are forced to satisfy two balances, namely geostrophic balance u = u h = ∇ h ⊥ p u = u_h = \nabla_h^\perp p u = u h = ∇ h ⊥ p and hydrostatic balance θ e − min 0 q t = ∂ 3 p \theta_e - \min_0 q_t = \partial_3 p θ e − min 0 q t = ∂ 3 p , which come from the fast rotation and strong stratification, respectively. This warrants the following definition.
Definition 2.1 (State space and balanced set). We denote by L σ 2 : = ( L σ 2 ) 3 × L 2 × L 2 \mathbb{L}^2_\sigma:= (L^2_\sigma)^3 \times L^2 \times L^2 L σ 2 := ( L σ 2 ) 3 × L 2 × L 2 the state space of the moist Boussinesq system (2.7a)–(2.7c). Here u ∈ ( L σ 2 ) 3 u \in(L^2_\sigma)^3 u ∈ ( L σ 2 ) 3 means that u ∈ L 2 ( T 3 ; R 3 ) u \in L^2(\mathbb{T}^3;\mathbb{R}^3) u ∈ L 2 ( T 3 ; R 3 ) is divergence-free. We then define the balanced set
B : = { ( u , θ e , q t ) ∈ L σ 2 : u h = ∇ h ⊥ p , u 3 = 0 , and θ e − min 0 q t = ∂ 3 p for some p ∈ H ˚ 1 } (11) \mathcal{B} := \left\{(u,\theta_e,q_t) \in\mathbb{L}^2_\sigma: u_h = \nabla_h^\perp p,\ u_3 = 0,\ \text{and }\theta_e - \min_0 q_t = \partial_3 p\text{ for some }p \in\mathring{H}^1\right\} \tag*{(11)} B := { ( u , θ e , q t ) ∈ L σ 2 : u h = ∇ h ⊥ p , u 3 = 0 , and θ e − 0 min q t = ∂ 3 p for some p ∈ H ˚ 1 } (11) where H ˚ 1 \mathring{H}^1 H ˚ 1 is the set of scalar fields in H 1 H^1 H 1 with vanishing averages.
We note that a state belongs to the balanced set precisely when it satisfies both geostrophic and hydrostatic balance.
Proceeding once again by analogy with the dry case, one may now identify two slow quantities : the equivalent potential vorticity P V e = ∇ h ⊥ ⋅ u h + ∂ 3 θ e PV_e = \nabla_h^\perp\cdot u_h + \partial_3\theta_e P V e = ∇ h ⊥ ⋅ u h + ∂ 3 θ e , defined now in terms of the equivalent potential temperature , and the moist variable M = θ e + q t M = \theta_e + q_t M = θ e + q t [92 ] . Similar to the dry case, now p p p and M M M fully determine a balanced state by way of the P V e PV_e P V e -and-M M M inversion
Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) = P V e , (12) \Delta p + \frac{1}{2}\partial_3\min_0\left(M-\partial_3p\right)=PV_e, \tag*{(12)} Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) = P V e , (12)
which is obtained as before by inserting the balances into the definition of P V e PV_e P V e . This is a nonlinear elliptic PDE whose well-posedness was established recently [89 ] . Using this PDE for P V e PV_e P V e -and-M M M inversion, one can then phrase the limiting moist (and potentially precipitating ) quasi-geostrophic system as a nonlinear transport equation for P V e PV_e P V e and M M M [92 ] .
It is important to note that the nonlinear elliptic PDE in (12) only has meaning for the balanced states from (11) , at this point in the story. This is in contrast to the cases of the Helmholtz decomposition and dry Boussinesq system from § 2.1 and § 2.2 , respectively, for which an elliptic PDE is part of a state decomposition for general states . One aim of the present paper is to investigate whether the nonlinear P V e PV_e P V e -and-M M M inversion in (12) may have additional meaning as part of a state decomposition for general states .
The question is now this: can we find a decomposition for the moist Boussinesq system which has the same favourable properties as the Helmholtz and vortical–wave decompositions?
Literature on atmospheric dynamics with moisture and clouds, and mathematical theory
In this section we review some literature on topics related to the main focus of the paper, i.e., state decomposition for a moist atmosphere with clouds. We mention both atmospheric sciences literature and mathematics literature.
Decompositions have been useful for numerous applications in atmospheric and science, and they are commonly based on decomposing the dry portions ( u , θ ) (u,\theta) ( u , θ ) of the atmospheric state. Applications include analysis of observational or model data [51 ZKT+15] , data assimilation [69 , 29 , 31 , 104 , 2 ] , and nonlinear wave interactions [41 , 74 , 12 , 87 , 39 ] . Other references are also described at the beginning of § 2.2 . One aim here is to extend the ideas of dry decompositions to the realm of a moist atmosphere with clouds.
For a moist atmosphere with clouds, decompositions can potentially provide valuable insight into the behavior of rainfall and precipitation [60 ] . Past work has commonly used linearisation, and either dry or moist eigenmodes. By adapting or modifying the dry linear eigenmodes, some studies have analyzed observational and model data for understanding moisture-coupled phenomena such as convectively coupled equatorial waves (CCEW) [102 , 43 , 82 , 75 KGH+22] , tropical intraseasonal oscillations such as the Madden–Julian Oscillation (MJO) [103 ] , and the Walker circulation [91 ] . Moist eigenmodes, obtained by linearisation of moist dynamics, have also been used for analyzing observational and model data for the MJO [81 , 90 ] , and for asymptotic analysis of weakly nonlinear wave interactions [20 , 21 ] . Balanced and unbalanced moisture decompositions have been investigated with a type of linearisation that applies with clouds present [101 , 100 ] . Applications to data assimilation have also investigated linear eigenmodes that decompose moisture into its balanced and unbalanced components [84 ] .
Here we aim for a decomposition that does not rely on linearisation, so that it may potentially include a more faithful representation of the nonlinearity and the physics associated with clouds and latent heating.
The PDE community has recently provided rigorous treatment of moist atmospheric dynamics with phase changes. A model with water vapor and liquid water was analyzed in [27 , 25 , 7 ] , where the velocity field is prescribed. Then followed results for two-phase flow with dynamical velocity evolving via the primitive equations [26 , 68 , 97 , 98 ] . An additional category of water was included in [17 , 95 , 49 , 50 ] , such that liquid water is divided into cloud water (that does not fall) and precipitating water. The work on warm-rain microphysics flow may be separated by the microphysics model determining how water is partitioned between vapor, cloud water and rain water, as well as by whether the velocity is prescribed or governed by the primitive equations. The case of prescribed velocity is considered in [17 ] for Grabowski microphysics [45 ] , and in [49 ] for Klein-Majda microphysics [58 ] . For primitive-equation velocity, theorems for the Grabowski and Klein-Majda models are found, respectively, in [95 ] and [50 ] . Incorporation of ice into water microphysics is discussed in [18 ] .
We note that the aforementioned rigorous studies of atmospheric flows with phase changes treat the velocity as either prescribed, or as governed by the primitive equations. Following [23 ] , the barotropic-baroclinic decomposition of the velocity is the basis for analysis of the primitive equations and its extensions. Recently, the current authors obtained existence, uniqueness and regularity results [89 ] for the nonlinear elliptic PDE (12) underlying Boussinesq velocity dynamics in the limit of asymptotically fast rotation rate and buoyancy frequencies [92 , 109 ] . Here, we show that the same nonlinear elliptic PDE is the foundation for decomposition of the Boussinesq equations more generally, in the regime of finite parameter values.
Pertaining to the dry Boussinesq evolution equations in the limit of asymptotically fast rotation rate and buoyancy frequency, a series of rigorous results established the decoupling between slow, balanced motions and fast, unbalanced waves [EM96, BANZ97, EM98, ME98, Maj03]. For moist atmospheric dynamics with phase changes, first steps to use the fast-wave averaging framework [EM96, EM98, ME98, Maj03] appear in [108 , 107 ] . The latter references present formal asymptotic analysis [108 ] and supporting numerical analyses [107 ] . The inclusion of boundary layers and the diabatic layer have been investigated in formal asymptotic analysis as well [62 BIHK23] . We note that rigorous fast-wave-averaging analysis for multi-phase atmospheric flows remains open. In § 4 , we explain the path forward for rigorous fast-wave-averaging analysis using the decomposition defined in § 3 . To date, the evidence suggests that nonlinear oscillations lead to coupling between fast and slow flow components, possibly even in the limiting dynamics when distinct phases co-exist (see Figure 13 in the appendix and [107 ] ).
Figure 13. An exact solution of q ′ ′ + q + min 0 q = 0 q'' + q + \min_0 q = 0 q ′′ + q + min 0 q = 0 which has nonzero averages (over its period). Note that this is not a sine wave shifted upwards slightly: the oscillation frequencies and amplitudes are different when q < 0 q < 0 q < 0 and when q ≥ 0 q \ge0 q ≥ 0 .
Simplified notation
To simplify the notation used throughout the rest of the paper, we will from now on use θ \theta θ to denote the equivalent potential temperature θ e \theta_e θ e (which is not an issue since the potential temperature, denoted θ \theta θ in the text above, will not make any appearances in the sequel). Similarly we will write P V PV P V for the equivalent potential temperature P V e PV_e P V e and q q q to mean q t q_t q t , the total water content.
Decomposition
As discussed in § 1 and § 2 , we seek a decomposition adapted to the leading-order operator of the moist Boussinesq system. We can see from (2.7a)–(2.7c) that this operator is
N ( u , θ , q ) = ( P L ( u h ⊥ − ( θ − min 0 q ) e 3 ) u 3 − u 3 ) (13) \mathcal{N}(\mathbf{u},\theta,q)=
\begin{pmatrix}
\mathbb{P}_{L}\left(u_h^\perp-(\theta-\min_{0}q)e_3\right)\\
u_3\\
-u_3
\end{pmatrix} \tag*{(13)} N ( u , θ , q ) = P L ( u h ⊥ − ( θ − min 0 q ) e 3 ) u 3 − u 3 (13)
where we note again that θ e \theta_e θ e and q t q_t q t are written as θ \theta θ and q q q for notational simplicity, as specified in § 2.5 . Here P L \mathbb{P}_{L} P L denotes the Leray projector onto the space of divergence-free vector fields. On the three-dimensional torus it takes the simple form P L u = − ∇ × ( Δ − 1 ∇ × u ) \mathbb{P}_{L}u=-\nabla\times(\Delta^{-1}\nabla\times u) P L u = − ∇ × ( Δ − 1 ∇ × u ) .
Given that the operator N \mathcal{N} N is nonlinear, it is likely impossible to define and characterize a full set of the nonlinear eigenfunctions and eigenvalues. Hence it is likely impossible to define a full eigenfunction decomposition such as in (2) . An alternative route to a decomposition must be sought.
Here we propose to use the notions of null set and image as concepts that may allow us to extend from the linear to the nonlinear realm. We propose a decomposition that will split any state in the state space L σ 2 L^2_\sigma L σ 2 into two pieces: one piece residing in the null set of N \mathcal{N} N and another piece residing in the image of N \mathcal{N} N . In other words we will perform the decomposition
L σ 2 = { N = 0 } + im N . (14) \mathbb{L}^2_\sigma= \{\mathcal{N}=0\}+\operatorname{im}\mathcal{N}. \tag*{(14)} L σ 2 = { N = 0 } + im N . (14)
If the operator N \mathcal{N} N were linear , then this decomposition would be a (coarse) eigendecomposition, into the eigenspace of the zero eigenvalue and the eigenspace of all nonzero eigenvalues. Such eigenspace ideas are helpful as motivation for pursuing the null set and image as fundamental concepts for defining a decomposition here. However, the operator N \mathcal{N} N is nonlinear, so the full meaning of (14) is not immediately clear. We aim to demonstrate in this section that (14) is able to characterize the desired decomposition in the nonlinear realm.
Note that in this section we will only present and discuss this decomposition itself. We will not discuss the various interpretations of this decomposition. This will be done in later sections. In § 4 we discuss how this decomposition is related to slow-fast dynamics, in § 5 we discuss how this decomposition relates to the conserved energy and projections, and finally in § 6 we discuss iterative methods used to compute this decomposition in practice and their geometric interpretation.
Since we wish to perform the decomposition foreshadowed in (14) we must begin by understanding both the null set and the image of N \mathcal{N} N as well as possible. We first turn our attention to the null set. As shown in the result below, this null set is precisely the set of states consisting of states in both hydrostatic and geostrophic balance. This is why the null set of N \mathcal{N} N will be referred to, from now on, as the balanced set .
Proposition 3.1 (Alternate characterisations of the balanced set). For any state X ∈ L σ 2 \mathcal{X}\in\mathbb{L}^2_\sigma X ∈ L σ 2 the following are equivalent.
X \mathcal{X} X is balanced, i.e., N ( u , θ , q ) = 0 \mathcal{N}(u,\theta,q)=0 N ( u , θ , q ) = 0 .
X = ( u , θ , q ) \mathcal{X}=(u,\theta,q) X = ( u , θ , q ) satisfies
∂ 3 u h − ∇ h ⊥ ( θ − min 0 q ) = 0 , u 3 = 0 , ∫ u h ⊥ = 0 , and ∫ θ − min 0 q = 0. (15) \partial_3u_h-\nabla_h^\perp(\theta-\min_0q)=0,\quad u_3=0,\quad\int u_h^\perp=0,\quad\text{and}\quad\int\theta-\min_0q=0. \tag*{(15)} ∂ 3 u h − ∇ h ⊥ ( θ − 0 min q ) = 0 , u 3 = 0 , ∫ u h ⊥ = 0 , and ∫ θ − 0 min q = 0. (15) X \mathcal{X} X may be parameterised as X = Φ ( p , M ) \mathcal{X}=\Phi(p,M) X = Φ ( p , M ) for some unique p ∈ H ˚ 1 p\in\mathring{H}^1 p ∈ H ˚ 1 and M ∈ L 2 M\in L^2 M ∈ L 2 , where
Φ ( p , M ) = ( ∇ h ⊥ p ∂ 3 p + 1 2 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 1 2 min 0 ( M − ∂ 3 p ) ) . \Phi(p,M)=\begin{pmatrix}
\nabla_h^\perp p\\
\partial_3p+\frac{1}{2}\min_0(M-\partial_3p)\\
M-\partial_3p-\frac{1}{2}\min_0(M-\partial_3p)
\end{pmatrix}. Φ ( p , M ) = ∇ h ⊥ p ∂ 3 p + 2 1 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 2 1 min 0 ( M − ∂ 3 p ) .
Note that this proposition characterizes the balanced set in three ways:
it is the null set of N \mathcal{N} N ,
it is characterised as the null set of the three measurements j j j , a a a , and w w w given by
j ( u , θ , q ) = ∂ 3 u h ⊥ − ∇ h ⊥ ( θ − min 0 q ) , w ( u , θ , q ) = u 3 j(u,\theta,q)=\partial_3u_h^\perp-\nabla_h^\perp(\theta-\min_0q),\quad w(u,\theta,q)=u_3 j ( u , θ , q ) = ∂ 3 u h ⊥ − ∇ h ⊥ ( θ − 0 min q ) , w ( u , θ , q ) = u 3
and
a ( u , θ , q ) = − ∫ u h ⊥ + ∫ ( θ − min 0 q ) e 3 , a(u,\theta,q)=-\int u_h^\perp+\int(\theta-\min_0q)e_3, a ( u , θ , q ) = − ∫ u h ⊥ + ∫ ( θ − 0 min q ) e 3 ,
and
The measurement j j j is commonly referred to as the thermal wind imbalance since it quantifies how far the state is from obeying both geostrophic and hydrostatic balanced. Further down we will see that the image of N \mathcal{N} N also admits three comparable descriptions. The measurement a a a depends on some spatial averages of the state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) . Why it depends on these particular averages (e.g. it depends on the horizontal velocity average but not the vertical velocity average) and why these averages are combined in this way to constitute a a a is explained in Remark 3.2.
Proof of Proposition 3.1 . It will be particularly convenient to consider the new unknown
v : = − u h ⊥ + ( θ − min 0 q ) e 3 . v:=-u_h^\perp+(\theta-\min_0q)e_3. v := − u h ⊥ + ( θ − 0 min q ) e 3 .
Indeed, we see that using this unknown we may rephrase items 1, 2, and 3 in a much simpler way. The first item may now be written as
v = ∇ p and u 3 = 0 for some p ∈ H ˚ 1 v=\nabla p\quad\text{and}\quad u_3=0\text{ for some }p\in\mathring{H}^1 v = ∇ p and u 3 = 0 for some p ∈ H ˚ 1
while the second item may be written, using Lemma A.5 , as
∇ × v = 0 , ∫ v = 0 , and u 3 = 0. \nabla\times v = 0,\quad\int v = 0,\quad\text{and }u_3 = 0. ∇ × v = 0 , ∫ v = 0 , and u 3 = 0.
It then follows directly from the Helmholtz decomposition (see Corollary A.2) that items 1 and 2 are equivalent. Finally we note that, by virtue of Lemma A.6 , item 3 is equivalent to
{ u = ∇ h ⊥ p , θ + q = M , and θ − min 0 q = ∂ 3 p , \left\{
\begin{aligned}
u &= \nabla_h^\perp p,\\
\theta+ q &= M,\quad\text{and}\\
\theta- \min_0 q &= \partial_3p,
\end{aligned}
\right. ⎩ ⎨ ⎧ u θ + q θ − 0 min q = ∇ h ⊥ p , = M , and = ∂ 3 p ,
which implies that v = ∇ p v = \nabla p v = ∇ p and that u 3 = 0 u_3 = 0 u 3 = 0 , from which item 1 follows.
Now that we understand the null set quite well we turn our attention to the image of N \mathcal{N} N . This set will actually be called the wave set W \mathcal{W} W . This is because the dynamics induced on that set by the moist Boussinesq system are oscillatory (in the dry case, the dynamics induced on the dry analog of the wave set by the dry Boussinesq system are precisely the dynamics of linear oscillations, i.e. waves).
Proposition 3.3 (Characterisation of the image of N \mathcal{N} N ). For any state X ∈ L σ 2 \mathcal{X}\in L^2_\sigma X ∈ L σ 2 the following are equivalent.
X \mathcal{X} X belongs to the image of N \mathcal{N} N .
X = ( u , θ , q ) \mathcal{X}=(u,\theta,q) X = ( u , θ , q ) satisfies
P V = ∇ h ⊥ ⋅ u h + ∂ 3 θ = 0 and M = θ + q = 0. (16) PV=\nabla_h^\perp\cdot u_h+\partial_3\theta=0\quad\text{and}\quad M=\theta+q=0. \tag*{(16)} P V = ∇ h ⊥ ⋅ u h + ∂ 3 θ = 0 and M = θ + q = 0. (16) X \mathcal{X} X may be parameterised as X = Ψ ( σ , w ) \mathcal{X}=\Psi(\sigma,w) X = Ψ ( σ , w ) for some unique σ ∈ H σ 1 \sigma\in H^1_\sigma σ ∈ H σ 1 and w ∈ L 2 w\in L^2 w ∈ L 2 satisfying ∇ h ⊥ ⋅ σ h = ∂ 3 w \nabla_h^\perp\cdot\sigma_h=\partial_3w ∇ h ⊥ ⋅ σ h = ∂ 3 w , where
Ψ ( σ , w ) = ( σ h ⊥ + w e 3 σ 3 − σ 3 ) . \Psi(\sigma,w)=
\begin{pmatrix}
\sigma_h^\perp+w e_3\\
\sigma_3\\
-\sigma_3
\end{pmatrix}. Ψ ( σ , w ) = σ h ⊥ + w e 3 σ 3 − σ 3 . In either case we say that X ∈ W \mathcal{X}\in\mathcal{W} X ∈ W for the wave set W \mathcal{W} W . Why this set warrants the name “wave set” will be discussed in § 4 below when we discuss slow-fast decompositions in more detail.
Note that, as was done for the balanced set above, Proposition 3.3 characterises the wave set in three ways:
it is the image of N \mathcal{N} N ,
it is characterised as the null set of the two measurements P V PV P V and M M M , and
it may be parameterised via the coordinates σ \sigma σ and w w w .
Proof. First we suppose that item 1 holds, i.e. that
{ u = P L ( v h ⊥ + ( ϕ − min 0 r ) e 3 ) , θ = v 3 , and q = − v 3 \left\{
\begin{aligned}
u &= \mathbb{P}_{L}\left(v_h^\perp+(\phi-\min_0 r)e_3\right),\\
\theta&= v_3,\ \text{and}\\
q &= -v_3
\end{aligned}
\right. ⎩ ⎨ ⎧ u θ q = P L ( v h ⊥ + ( ϕ − 0 min r ) e 3 ) , = v 3 , and = − v 3
for some ( v , ϕ , r ) ∈ L σ 2 (v,\phi,r)\in L^2_\sigma ( v , ϕ , r ) ∈ L σ 2 . This means that u = v h ⊥ + ( ϕ − min 0 q ) e 3 + ∇ π u=v_h^\perp+(\phi-\min_0 q)e_3+\nabla\pi u = v h ⊥ + ( ϕ − min 0 q ) e 3 + ∇ π for some π ∈ H ˚ 1 \pi\in\mathring{H}^1 π ∈ H ˚ 1 and so
∇ h ⊥ ⋅ u h + ∂ 3 θ = ∇ h ⊥ ⋅ ( v h ⊥ + ∇ h π ) + ∂ 3 v 3 = ∇ ⋅ v = 0 \nabla_h^\perp\cdot u_h+\partial_3\theta=\nabla_h^\perp\cdot\left(v_h^\perp+\nabla_h\pi\right)+\partial_3v_3=\nabla\cdot v=0 ∇ h ⊥ ⋅ u h + ∂ 3 θ = ∇ h ⊥ ⋅ ( v h ⊥ + ∇ h π ) + ∂ 3 v 3 = ∇ ⋅ v = 0
while clearly θ + q = 0 \theta+q=0 θ + q = 0 . Thus item 2 holds.
Now suppose that item 2 holds. We may introduce the new unknown v = − u h ⊥ + θ e 3 v=-u_h^\perp+\theta e_3 v = − u h ⊥ + θ e 3 such that now item 2 reads
∇ ⋅ v = 0 and θ + q = 0 \nabla\cdot v=0\quad\text{and}\quad\theta+q=0 ∇ ⋅ v = 0 and θ + q = 0
(by virtue of Lemma A.5 ). In other words v = σ v=\sigma v = σ for some divergence-free σ \sigma σ while θ = − q \theta=-q θ = − q , and so
{ u = σ h ⊥ + u 3 e 3 , θ = σ 3 , and q = − σ 3 . \left\{
\begin{aligned}
u &= \sigma_h^\perp+u_3e_3,\\
\theta&= \sigma_3,\ \text{and}\\
q &= -\sigma_3.
\end{aligned}
\right. ⎩ ⎨ ⎧ u θ q = σ h ⊥ + u 3 e 3 , = σ 3 , and = − σ 3 .
Therefore, for w : = u 3 w:=u_3 w := u 3 , we see that item 3 holds as desired since ∇ h ⊥ ⋅ σ h − ∂ 3 w = − ∇ ⋅ u = 0 \nabla_h^\perp\cdot\sigma_h-\partial_3w=-\nabla\cdot u=0 ∇ h ⊥ ⋅ σ h − ∂ 3 w = − ∇ ⋅ u = 0 .
Finally we suppose that item 3 holds. Then item 1 follows immediately from choosing v = σ v=\sigma v = σ , ϕ = w \phi=w ϕ = w , and r = 0 r=0 r = 0 since σ h ⊥ + w e 3 \sigma_h^\perp+we_3 σ h ⊥ + w e 3 is divergence-free and so the Leray projection of this vector field is equal to itself.
At this stage we have a good understanding of both the balanced set (which is the null set of N \mathcal{N} N ) and of the wave set (which is the image of N \mathcal{N} N ). We now seek to leverage this understanding to produce the decomposition formally foreshadowed in (14) .
In order to do this in a manner that will be useful later, especially when it comes to dynamics, we need to discuss dynamics briefly. Recall that, at leading order, the dynamics of interest are those governed by ∂ t + N = 0 \partial_t+\mathcal{N}=0 ∂ t + N = 0 . So how do the null set and image of N \mathcal{N} N relate to this dynamical system?
Well, note that Proposition 3.1 tells us that producing balanced states is easy: it suffices to specify p p p and M M M since then a balanced state may be reconstructed from these two quantities. If the leading operator N \mathcal{N} N were linear, then this would be enough since balanced solutions of ∂ t + N \partial_t+\mathcal{N} ∂ t + N would then be guaranteed to be slow, meaning that they would have vanishing time derivatives. However, due to the presence of moisture and phase boundaries, N \mathcal{N} N is not linear. This means that we must additionally impose that balanced solutions be slow. In more prosaic terms: since any choice of the coordinates p p p and M M M produces a balanced state, we must make sure that these coordinates are carefully chosen to also be slow .
This is where our understanding of the image of N \mathcal{N} N , i.e. of the wave set, comes in handy. Indeed: we have characterised this wave set as the set of states whose P V PV P V and M M M vanish. This is another way of saying that P V PV P V and M M M are precisely a complete set of time-dependent slow measurements! (Any constant measurement would be trivially slow.)
We have found our solution: to guarantee that our balanced component determined by p p p and M M M is also slow, it must be reconstructed solely from the slow measurements, P V PV P V and M M M . If we think about this geometrically, this is really a matter of transversality : we need the wave set W \mathcal{W} W to be transverse to the balanced set B \mathcal{B} B so that balanced states in B \mathcal{B} B may be fully characterised by the measurements annihilating the wave set W \mathcal{W} W . This is taken care of in the result below.
We can think about why the balanced set and the wave set need to be transverse another way. Recall that we are after a decomposition of the form (14) . This decomposition is only unique if its two components are transverse, otherwise the same state could be decomposed in several different ways.
Lemma 3.4 (PV and M characterise balanced states). Given P V ∈ H − 1 PV \in H^{-1} P V ∈ H − 1 and M ∈ L 2 M \in L^2 M ∈ L 2 , where H − 1 H^{-1} H − 1 denotes the dual of H ˚ 1 \mathring{H}^1 H ˚ 1 , there exists a unique X ∈ B \mathcal{X} \in\mathcal{B} X ∈ B for which P V ( X ) = P V PV(\mathcal{X}) = PV P V ( X ) = P V and M ( X ) = M M(\mathcal{X}) = M M ( X ) = M .
Proof. By virtue of Proposition 3.1 it suffices to show that there exist a unique p ∈ H ˚ 1 p \in\mathring{H}^1 p ∈ H ˚ 1 and a unique M ‾ ∈ L 2 \overline{M} \in L^2 M ∈ L 2 for which ( P V ∘ Φ ) ( p , M ‾ ) = P V (PV \circ\Phi)(p,\overline{M}) = PV ( P V ∘ Φ ) ( p , M ) = P V and ( M ∘ Φ ) ( p , M ‾ ) = M (M \circ\Phi)(p,\overline{M}) = M ( M ∘ Φ ) ( p , M ) = M . The latter identity tells us that M ‾ = M \overline{M} = M M = M , and since
( P V ∘ Φ ) ( p , M ) = Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) (PV \circ\Phi)(p,M) = \Delta p + \frac{1}{2}\partial_3\min_0\left(M-\partial_3p\right) ( P V ∘ Φ ) ( p , M ) = Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) the result follows from the uniqueness of a solution to nonlinear P V PV P V -and-M M M inversion as proved in [89 ] .
Now that we have verified that the wave set and the balanced set are transverse, we are ready to obtain the decomposition that this entire section is building towards.
Theorem 3.5 (Pre-decomposition). We have the decomposition L σ 2 = B + W L^2_\sigma= \mathcal{B} + \mathcal{W} L σ 2 = B + W in the sense that, for any X ∈ L σ 2 \mathcal{X} \in L^2_\sigma X ∈ L σ 2 there exist unique X B ∈ B \mathcal{X}_B \in\mathcal{B} X B ∈ B and X W ∈ W \mathcal{X}_W \in\mathcal{W} X W ∈ W , called the balanced component of X \mathcal{X} X and the wave component of X \mathcal{X} X respectively, such that
X = X B + X W \mathcal{X} = \mathcal{X}_B + \mathcal{X}_W X = X B + X W and
P V ( X ) = P V ( X B ) PV(\mathcal{X}) = PV(\mathcal{X}_B) P V ( X ) = P V ( X B ) and M ( X ) = M ( X B ) M(\mathcal{X}) = M(\mathcal{X}_B) M ( X ) = M ( X B ) for M ( u , θ , q ) = θ + q M(u,\theta,q) = \theta+ q M ( u , θ , q ) = θ + q and P V ( u , θ , q ) = ∇ h ⊥ ⋅ u h + ∂ 3 θ PV(u,\theta,q) = \nabla_h^\perp\cdot u_h + \partial_3\theta P V ( u , θ , q ) = ∇ h ⊥ ⋅ u h + ∂ 3 θ .
Moreover X B \mathcal{X}_B X B and X W \mathcal{X}_W X W may be computed explicitly as follows.
X B = Φ ( p , M ) \mathcal{X}_B = \Phi(p,M) X B = Φ ( p , M ) for Φ \Phi Φ as in Proposition 3.1 , where M : = θ + q M := \theta+ q M := θ + q and where p ∈ H ˚ 1 p \in\mathring{H}^1 p ∈ H ˚ 1 is the unique solution of
Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) = P V . (17) \Delta p + \frac{1}{2}\partial_3\min_0\left(M-\partial_3p\right) = PV. \tag*{(17)} Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) = P V . (17) X W = X − X B \mathcal{X}_W = \mathcal{X} - \mathcal{X}_B X W = X − X B , and, in light of Proposition 3.3 , we also have X W = Ψ ( σ , w ) \mathcal{X}_W = \Psi(\sigma,w) X W = Ψ ( σ , w ) where w = u 3 w = u_3 w = u 3 and
σ = − ( u h ⊥ − u B , h ⊥ ) + ( θ − θ B ) e 3 \sigma= -\left(u_h^\perp-u_{B,h}^\perp\right)+(\theta-\theta_B)e_3 σ = − ( u h ⊥ − u B , h ⊥ ) + ( θ − θ B ) e 3 = − ( u h ⊥ − ∇ h p ) + ( θ − ∂ 3 p − 1 2 min 0 ( M − ∂ 3 p ) ) e 3 . = -\left(u_h^\perp-\nabla_hp\right)+\left(\theta-\partial_3p-\frac{1}{2}\min_0\left(M-\partial_3p\right)\right)e_3. = − ( u h ⊥ − ∇ h p ) + ( θ − ∂ 3 p − 2 1 0 min ( M − ∂ 3 p ) ) e 3 . In particular σ \sigma σ satisfies
∇ × ( A B − 1 σ ) = ∂ 3 u h − ∇ h ⊥ ( θ − H B q ) + ( ∂ 3 w ) e 3 , \nabla\times\left(A_B^{-1}\sigma\right)=\partial_3u_h-\nabla_h^\perp\left(\theta-H_Bq\right)+(\partial_3w)e_3, ∇ × ( A B − 1 σ ) = ∂ 3 u h − ∇ h ⊥ ( θ − H B q ) + ( ∂ 3 w ) e 3 , where H B : = 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) H_B := \mathbb{1}(q_B<0)=\mathbb{1}(M<\partial_3p) H B := 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) and A B : = I − 1 2 H B e 3 ⊗ e 3 A_B := I-\frac{1}{2}H_Be_3\otimes e_3 A B := I − 2 1 H B e 3 ⊗ e 3 , such that A B − 1 = I + H B e 3 ⊗ e 3 A_B^{-1}=I+H_Be_3\otimes e_3 A B − 1 = I + H B e 3 ⊗ e 3 , subject to ∇ ⋅ σ = 0 \nabla\cdot\sigma=0 ∇ ⋅ σ = 0 and
∫ A B − 1 σ = − ∫ u h ⊥ + ∫ ( θ − H B q ) e 3 . \int A_B^{-1}\sigma=-\int u_h^\perp+\int\left(\theta-H_Bq\right)e_3. ∫ A B − 1 σ = − ∫ u h ⊥ + ∫ ( θ − H B q ) e 3 .
Proof. The uniqueness of this decomposition follows from Lemma 3.4 and the fact that, as per Proposition 3.3 , elements of the wave set W \mathcal{W} W have vanishing P V PV P V and M M M . The existence of the decomposition follows from the existence of solutions to nonlinear P V PV P V -and-M M M inversion since finding X B \mathcal{X}_B X B is, by virtue of Proposition 3.1 , equivalent to finding p p p and M M M and since p p p may be chosen to be the solution of nonlinear P V PV P V -and-M M M inversion.
The latter characterisation of σ \sigma σ follows from Lemmas A.5 and A.7 :
∇ × ( A B − 1 σ ) = ∇ × [ − u h ⊥ + ( θ − H B q ) e 3 ] + ∇ × ∇ p = ∂ 3 u h − ∇ h ⊥ ( θ − H B q ) + ( ∂ 3 w ) e 3 \nabla\times\left(A_B^{-1}\sigma\right)=\nabla\times\left[-u_h^\perp+\left(\theta-H_Bq\right)e_3\right]+\nabla\times\nabla p
=\partial_3u_h-\nabla_h^\perp\left(\theta-H_Bq\right)+(\partial_3w)e_3 ∇ × ( A B − 1 σ ) = ∇ × [ − u h ⊥ + ( θ − H B q ) e 3 ] + ∇ × ∇ p = ∂ 3 u h − ∇ h ⊥ ( θ − H B q ) + ( ∂ 3 w ) e 3
and
∫ A B − 1 σ = − ∫ u h ⊥ + ∫ ( θ − H B q ) e 3 . \int A_B^{-1}\sigma=-\int u_h^\perp+\int\left(\theta-H_Bq\right)e_3. ∫ A B − 1 σ = − ∫ u h ⊥ + ∫ ( θ − H B q ) e 3 .
It is interesting to see ∇ × ( A B − 1 σ ) \nabla\times\left(A_B^{-1}\sigma\right) ∇ × ( A B − 1 σ ) arise in the pre-decomposition above, where matrix A B − 1 A_B^{-1} A B − 1 depends on the balanced moisture, q B q_B q B . It would be difficult to guess the form of ∇ × ( A B − 1 σ ) \nabla\times\left(A_B^{-1}\sigma\right) ∇ × ( A B − 1 σ ) , with its coefficient matrix that depends on balanced moisture, q B q_B q B . Here, the appearance of A B − 1 A_B^{-1} A B − 1 follows from the decomposition into nullspace and image, and from seeking a curl-like operator for determining the unbalanced coordinate σ \sigma σ . Also note that ∇ × ( A B − 1 σ ) \nabla\times\left(A_B^{-1}\sigma\right) ∇ × ( A B − 1 σ ) involves a linear differential operator, once q B q_B q B is known, but it is a nonlinear transformation from ( u , θ , q ) (u,\theta,q) ( u , θ , q ) to σ \sigma σ .
As an immediate corollary of this decomposition we make the following observation: since both the balanced set and the wave set could be parameterised and since any state can be decomposed into a balanced and a wave piece, this means that we have obtained a global parametrisation of the entire state space.
Corollary 3.6. Any state X ∈ L σ 2 \mathcal{X} \in L^2_\sigma X ∈ L σ 2 can be uniquely parameterised by
{ u = ∇ h ⊥ p + σ h ⊥ + w e 3 , θ = ∂ 3 p + 1 2 min 0 ( M − ∂ 3 p ) + σ 3 , and q = M − ∂ 3 p − 1 2 min 0 ( M − ∂ 3 p ) − σ 3 \left\{
\begin{aligned}
u &= \nabla_h^\perp p + \sigma_h^\perp+ w e_3,\\
\theta&= \partial_3p + \frac{1}{2}\min_0(M-\partial_3p) + \sigma_3,\quad\text{and}\\
q &= M-\partial_3p-\frac{1}{2}\min_0(M-\partial_3p)-\sigma_3
\end{aligned}
\right. ⎩ ⎨ ⎧ u θ q = ∇ h ⊥ p + σ h ⊥ + w e 3 , = ∂ 3 p + 2 1 0 min ( M − ∂ 3 p ) + σ 3 , and = M − ∂ 3 p − 2 1 0 min ( M − ∂ 3 p ) − σ 3 for p ∈ H ˚ 1 p \in\mathring{H}^1 p ∈ H ˚ 1 , M , w ∈ L 2 M,w \in L^2 M , w ∈ L 2 , and σ ∈ ( L 2 ) σ 3 \sigma\in(L^2)^3_\sigma σ ∈ ( L 2 ) σ 3 satisfying ∇ h ⊥ ⋅ σ h = ∂ 3 w \nabla_h^\perp\cdot\sigma_h = \partial_3w ∇ h ⊥ ⋅ σ h = ∂ 3 w .
Proof. Theorem 3.5 tells us that any X ∈ L σ 2 \mathcal{X} \in L^2_\sigma X ∈ L σ 2 may be uniquely decomposed as
X = X B + X W ∈ B + W . \mathcal{X}=\mathcal{X}_B+\mathcal{X}_W\in\mathcal{B}+\mathcal{W}. X = X B + X W ∈ B + W .
Then Propositions 3.1 and 3.3 tell us that X B = Φ ( p , M ) \mathcal{X}_B=\Phi(p,M) X B = Φ ( p , M ) and X W = Ψ ( σ , w ) \mathcal{X}_W=\Psi(\sigma,w) X W = Ψ ( σ , w ) for some uniquely determined p p p , M M M , σ \sigma σ , and w w w . So indeed
X = Φ ( p , M ) + Ψ ( σ , w ) \mathcal{X}=\Phi(p,M)+\Psi(\sigma,w) X = Φ ( p , M ) + Ψ ( σ , w )
as desired.
We do not even have to stop there: both the balanced set and the wave set had two dual descriptions. They were described using a parametrisation, which produced Corollary 3.6 above. But they were also described using the measurements that annihilate them. This suggests that using the full set of measurements, namely both those characterising the balanced set and those characterising the wave set, would yield a complete description of the state space. We verify this in the result below.
Proposition 3.7 (Global nonlinear change of coordinates). Consider the measurement space
M : = H − 1 × L 2 × ( H − 1 ) 2 × L b 2 × R 3 . \mathcal{M}:=H^{-1}\times L^2\times(H^{-1})^2\times L^2_b\times\mathbb{R}^3. M := H − 1 × L 2 × ( H − 1 ) 2 × L b 2 × R 3 . The map M : L σ 2 → M \mathcal{M}:L^2_\sigma\to\mathcal{M} M : L σ 2 → M given by
M ( u , θ , q ) = ( ∇ h ⊥ ⋅ u h + ∂ 3 θ θ + q ∂ 3 u h − ∇ h ⊥ ( θ − min 0 q ) u 3 − ∫ u h ⊥ + ∫ ( θ − min 0 q ) e 3 ) = ( P V M j w a ) \mathcal{M}(u,\theta,q)=
\begin{pmatrix}
\nabla_h^\perp\cdot u_h+\partial_3\theta\\
\theta+q\\
\partial_3u_h-\nabla_h^\perp(\theta-\min_0q)\\
u_3\\
-\int u_h^\perp+\int(\theta-\min_0q)e_3
\end{pmatrix}
=
\begin{pmatrix}
PV\\
M\\
j\\
w\\
a
\end{pmatrix} M ( u , θ , q ) = ∇ h ⊥ ⋅ u h + ∂ 3 θ θ + q ∂ 3 u h − ∇ h ⊥ ( θ − min 0 q ) u 3 − ∫ u h ⊥ + ∫ ( θ − min 0 q ) e 3 = P V M j w a is invertible.
Proof. To verify the invertibility of this change of coordinates we will use the coordinates π \pi π , M M M , ζ \zeta ζ , and w w w defined implicitly by
{ u = ∇ h ⊥ π + ζ h ⊥ + w e 3 , θ = ∂ 3 π + ζ 3 + 1 2 min 0 ( M − ∂ 3 π − ζ 3 ) , and q = M − ∂ 3 π − ζ 3 − 1 2 min 0 ( M − ∂ 3 π − ζ 3 ) (18) \left\{
\begin{aligned}
u &= \nabla_h^\perp\pi+\zeta_h^\perp+w e_3,\\
\theta&= \partial_3\pi+\zeta_3+\frac{1}{2}\min_0(M-\partial_3\pi-\zeta_3),\quad\text{and}\\
q &= M-\partial_3\pi-\zeta_3-\frac{1}{2}\min_0(M-\partial_3\pi-\zeta_3)
\end{aligned}
\right. \tag*{(18)} ⎩ ⎨ ⎧ u θ q = ∇ h ⊥ π + ζ h ⊥ + w e 3 , = ∂ 3 π + ζ 3 + 2 1 0 min ( M − ∂ 3 π − ζ 3 ) , and = M − ∂ 3 π − ζ 3 − 2 1 0 min ( M − ∂ 3 π − ζ 3 ) (18)
where ∇ h ⊥ ⋅ ζ h = ∂ 3 w \nabla_h^\perp\cdot\zeta_h=\partial_3w ∇ h ⊥ ⋅ ζ h = ∂ 3 w and ∇ ⋅ ζ = 0 \nabla\cdot\zeta=0 ∇ ⋅ ζ = 0 . Note that the coordinates ( π , M , ζ , w ) (\pi,M,\zeta,w) ( π , M , ζ , w ) differ from the coordinates ( p , M , σ , w ) (p,M,\sigma,w) ( p , M , σ , w ) used in Corollary 3.6. The key advantage of the coordinates used here is that
sign q = sign ( M − ∂ 3 π − ζ 3 ) , (19) \operatorname{sign}q=\operatorname{sign}(M-\partial_3\pi-\zeta_3), \tag*{(19)} sign q = sign ( M − ∂ 3 π − ζ 3 ) , (19)
which is why they are particularly useful here when it comes to inverting M \mathcal{M} M .
However these coordinates have the less advantageous property that π \pi π depends on P V PV P V , M M M , and ζ \zeta ζ , as we will see below. This is different from p p p (from Corollary 3.6), which only depends on P V PV P V and M M M . That is the reason why these π \pi π and ζ \zeta ζ coordinates are not used for the decomposition (instead of p p p and σ \sigma σ , respectively): the resulting balanced component computed from π \pi π and M M M would not be slow.
We close this parenthesis regarding the choice of coordinates and go back to the matter of invertibility. In light of (3.7), Lemma A.6 and Lemma A.7 tell us that (3.6) is equivalent to
{ − u h ⊥ + ( θ − min 0 q ) e 3 = ∇ π + ζ , u 3 = w , and θ + q = M . (20) \left\{
\begin{aligned}
-u_h^\perp+(\theta-\min_0q)e_3&=\nabla\pi+\zeta,\\
u_3&=w,\quad\text{and}\\
\theta+q&=M.
\end{aligned}
\right. \tag*{(20)} ⎩ ⎨ ⎧ − u h ⊥ + ( θ − 0 min q ) e 3 u 3 θ + q = ∇ π + ζ , = w , and = M . (20)
In particular it follows from the uniqueness of the Helmholtz decomposition that these coordinates are uniquely determined by u u u , θ \theta θ , and q q q .
So now it remains to show that, for any tuple ( P V , M , j , w , a ) (PV, M, j, w, a) ( P V , M , j , w , a ) , corresponding coordinates ( π , M , ζ , w ) (\pi, M, \zeta, w) ( π , M , ζ , w ) exist. Plugging (18) into the definition of the measurement map M \mathcal{M} M we see, using (20) , that
{ P V = ∇ π + 1 2 ∂ 3 min 0 ( M − ∂ 3 π − ζ 3 ) , M = M , j = ∂ 3 ζ h − ∇ h ⊥ ζ 3 , w = w , and a = ∫ ζ . \left\{
\begin{aligned}
PV &= \nabla\pi+\frac{1}{2}\partial_{3}\min_{0}(M-\partial_{3}\pi-\zeta_{3}),\\
M &= M,\\
j &= \partial_{3}\zeta_{h}-\nabla_{h}^{\perp}\zeta_{3},\\
w &= w,\quad\text{and}\\
a &= \int\zeta.
\end{aligned}
\right. ⎩ ⎨ ⎧ P V M j w a = ∇ π + 2 1 ∂ 3 0 min ( M − ∂ 3 π − ζ 3 ) , = M , = ∂ 3 ζ h − ∇ h ⊥ ζ 3 , = w , and = ∫ ζ .
Since ∇ h ⊥ ⋅ ζ h = ∂ 3 w \nabla_{h}^{\perp}\cdot\zeta_{h}=\partial_{3}w ∇ h ⊥ ⋅ ζ h = ∂ 3 w we see using Lemma A.4 that
∇ × ζ = j + ( ∂ 3 w ) e 3 . \nabla\times\zeta=j+(\partial_{3}w)e_{3}. ∇ × ζ = j + ( ∂ 3 w ) e 3 .
We are thus guaranteed a (unique) solution ζ \zeta ζ of
{ ∇ × ζ = j + ( ∂ 3 w ) e 3 , ∇ ⋅ ζ = 0 , and ∫ ζ = a . (21) \left\{
\begin{aligned}
\nabla\times\zeta&= j+(\partial_{3}w)e_{3},\\
\nabla\cdot\zeta&= 0,\quad\text{and}\\
\int\zeta&= a.
\end{aligned}
\right. \tag*{(21)} ⎩ ⎨ ⎧ ∇ × ζ ∇ ⋅ ζ ∫ ζ = j + ( ∂ 3 w ) e 3 , = 0 , and = a . (21)
The solvability of nonlinear P V PV P V -and-M M M inversion [89 ] then guarantees a unique solution π \pi π in H ˙ 1 \dot{H}^{1} H ˙ 1 of
Δ π + 1 2 ∂ 3 min 0 [ ( M − ζ 3 ) − ∂ 3 π ] = P V . (22) \Delta\pi+\frac{1}{2}\partial_{3}\min_{0}\left[(M-\zeta_{3})-\partial_{3}\pi\right]=PV. \tag*{(22)} Δ π + 2 1 ∂ 3 0 min [ ( M − ζ 3 ) − ∂ 3 π ] = P V . (22)
In other words, if we define
E ~ : = { ( π , M , ζ , w ) ∈ H ˙ 1 × L 2 × ( L 2 ) σ 2 × L 2 : ∇ h ⊥ ⋅ ζ h = ∂ 3 w } \widetilde{\mathcal{E}}:=\left\{(\pi,M,\zeta,w)\in\dot{H}^{1}\times L^{2}\times(L^{2})^{2}_{\sigma}\times L^{2}:\nabla_{h}^{\perp}\cdot\zeta_{h}=\partial_{3}w\right\} E := { ( π , M , ζ , w ) ∈ H ˙ 1 × L 2 × ( L 2 ) σ 2 × L 2 : ∇ h ⊥ ⋅ ζ h = ∂ 3 w }
to be the space where the coordinates ( π , M , ζ , w ) (\pi,M,\zeta,w) ( π , M , ζ , w ) reside and denote by ( u , θ , q ) = C ( π , M , ζ , w ) (u,\theta,q)=\mathcal{C}(\pi,M,\zeta,w) ( u , θ , q ) = C ( π , M , ζ , w ) the map defined by (18) , we have shown that for any ( P V , M , j , w , a ) ∈ E (PV,M,j,w,a)\in\mathcal{E} ( P V , M , j , w , a ) ∈ E there exists a unique ( π , M , ζ , w ) ∈ E ~ (\pi,M,\zeta,w)\in\widetilde{\mathcal{E}} ( π , M , ζ , w ) ∈ E such that ( M ∘ C ) ( π , M , ζ , w ) = ( P V , M , j , w , a ) (\mathcal{M}\circ\mathcal{C})(\pi,M,\zeta,w)=(PV,M,j,w,a) ( M ∘ C ) ( π , M , ζ , w ) = ( P V , M , j , w , a ) . This means that we can take ( u , θ , q ) = C ( π , M , ζ , w ) (u,\theta,q)=\mathcal{C}(\pi,M,\zeta,w) ( u , θ , q ) = C ( π , M , ζ , w ) to satisfy M ( u , θ , q ) = ( P V , M , j , w , a ) \mathcal{M}(u,\theta,q)=(PV,M,j,w,a) M ( u , θ , q ) = ( P V , M , j , w , a ) . Moreover, since we know that this state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) is uniquely determined by its coordinates $(\pi,M,\zeta,w) this proves that there is a unique state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) satisfying M ( u , θ , q ) = ( P V , M , j , w , a ) \mathcal{M}(u,\theta,q)=(PV,M,j,w,a) M ( u , θ , q ) = ( P V , M , j , w , a ) . This proves that M \mathcal{M} M is invertible, as desired. Moreover the inverse may be computed by solving, in order, (21) then (22) and then plugging the solutions into (18) .
Remark 3.8 (The global change of coordinates in the dry case). The proof of Proposition 3.7 above is a bit of a mouthful. In order to shed some light on the underlying ideas, without having to worry about the details of the proof itself, we discuss here what an analog global change of coordinates would look like in the dry case.
In that case we could break down the change of coordinates as a composition of invertible pieces, thus illustrating more clearly what the moving pieces are. Indeed, in the dry case we have the following chain of inversions:
( u θ ) ↔ ( u h u 3 θ ) ↔ ( − u h ⊥ + θ e 3 u 3 ) = ( g w ) ↔ ( 1 ) ( ∇ ⋅ q ∇ × g ∫ g w ) ≡ ( 2 ) ( ∇ h ⋅ u h + ∂ 3 θ ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 − ∫ u h ⊥ + ∫ θ e 3 w ) = ( P V j + ( ∂ 3 w ) e 3 a w ) ↔ ( P V j a w ) . \begin{aligned}
\begin{pmatrix}
u\\
\theta
\end{pmatrix}
\leftrightarrow
\begin{pmatrix}
u_{h}\\
u_{3}\\
\theta
\end{pmatrix}
\leftrightarrow
\begin{pmatrix}
-u_{h}^{\perp}+\theta e_{3}\\
u_{3}
\end{pmatrix}
=
\begin{pmatrix}
g\\
w
\end{pmatrix}
\mathrel{\mathop{\leftrightarrow}^{(1)}}
\begin{pmatrix}
\nabla\cdot q\\
\nabla\times g\\
\int g\\
w
\end{pmatrix}
\mathrel{\mathop{\equiv}^{(2)}}
\begin{pmatrix}
\nabla_{h}\cdot u_{h}+\partial_{3}\theta\\
\partial_{3}u_{h}-\nabla_{h}^{\perp}\theta+(\partial_{3}w)e_{3}\\
-\int u_{h}^{\perp}+\int\theta e_{3}\\
w
\end{pmatrix} \\
&=
\begin{pmatrix}
PV\\
j+(\partial_{3}w)e_{3}\\
a\\
w
\end{pmatrix}
\leftrightarrow
\begin{pmatrix}
PV\\
j\\
a\\
w
\end{pmatrix}.
\end{aligned} ( u θ ) ↔ u h u 3 θ ↔ ( − u h ⊥ + θ e 3 u 3 ) = ( g w ) ↔ ( 1 ) ∇ ⋅ q ∇ × g ∫ g w ≡ ( 2 ) ∇ h ⋅ u h + ∂ 3 θ ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 − ∫ u h ⊥ + ∫ θ e 3 w = P V j + ( ∂ 3 w ) e 3 a w ↔ P V j a w .
where (1) follows from the Helmholtz decomposition (recorded in Theorem A.1 ) and (2) follows from Lemma A.5 and the fact that the velocity field u u u is divergence-free.
It is therefore tempting to assume that a similar chain of invertible maps produces the global nonlinear change of coordinates for the moist case recorded in Proposition 3.7. This is not the case. Of course, we would need to perform the following inversions initially in the moist case, by virtue of Lemma A.6 :
( u θ q ) ↔ ( u h u 3 θ + q θ − min 0 q ) = : ( w M b ) . \begin{pmatrix} u \\ \theta\\ q \end{pmatrix}
\leftrightarrow
\begin{pmatrix} u_h \\ u_3 \\ \theta+q \\ \theta-\min_0 q \end{pmatrix}
=:
\begin{pmatrix} w \\ M \\ b \end{pmatrix}. u θ q ↔ u h u 3 θ + q θ − min 0 q =: w M b .
This is not the source of the issue. The source of the issue is that, in the moist case, the potential vorticity P V PV P V and the thermal wind imbalance j j j come from different good unknowns. Indeed:
P V = ∇ ⋅ ( − u h ⊥ + θ e 3 ) while j = [ ∇ × ( − u h ⊥ + ( θ − min 0 q ) e 3 ) ] h . PV=\nabla\cdot\left(-u_h^\perp+\theta e_3\right)
\quad\text{while}\quad
j=\left[\nabla\times\left(-u_h^\perp+(\theta-\min_0 q)e_3\right)\right]_h. P V = ∇ ⋅ ( − u h ⊥ + θ e 3 ) while j = [ ∇ × ( − u h ⊥ + ( θ − 0 min q ) e 3 ) ] h .
The good unknowns
− u h ⊥ + θ e 3 and − u h ⊥ + ( θ − min 0 q ) e 3 -u_h^\perp+\theta e_3
\quad\text{and}\quad
-u_h^\perp+(\theta-\min_0 q)e_3 − u h ⊥ + θ e 3 and − u h ⊥ + ( θ − 0 min q ) e 3
are not the same. Note that this is not a technical issue. This comes from the dynamics in the moist case being fundamentally different. In particular, the operator in the moist case loses the skew-symmetry present in the dry case, and this is precisely the source of the discrepancy between the good unknowns used for P V PV P V and j j j . Indeed, the ( θ , q ) (\theta,q) ( θ , q ) –dependence in the equation governing ∂ t u \partial_tu ∂ t u is of the form
∂ t u = ⋯ + ( θ − min 0 q ) e 3 = ⋯ + ( θ − q H ) e 3 \partial_tu=\cdots+(\theta-\min_0 q)e_3=\cdots+(\theta-qH)e_3 ∂ t u = ⋯ + ( θ − 0 min q ) e 3 = ⋯ + ( θ − q H ) e 3
for H : = 1 ( q < 0 ) H:=\mathbf{1}(q<0) H := 1 ( q < 0 ) while the u u u –dependence in the equation governing ∂ t θ \partial_t\theta ∂ t θ and ∂ t q \partial_tq ∂ t q is of the form
∂ t θ = ⋯ − u 3 and ∂ t q = ⋯ + u 3 . \partial_t\theta=\cdots-u_3
\quad\text{and}\quad
\partial_tq=\cdots+u_3. ∂ t θ = ⋯ − u 3 and ∂ t q = ⋯ + u 3 .
In matrix form this reads
∂ t ( u θ q ) = ( ⋯ e 3 − H e 3 − e 3 T ⋯ ⋯ e 3 T ⋯ ⋯ ) ( u θ q ) . \partial_t
\begin{pmatrix} u \\ \theta\\ q \end{pmatrix}
=
\begin{pmatrix}
\cdots& e_3 & -He_3 \\
-e_3^{T} & \cdots& \cdots\\
e_3^{T} & \cdots& \cdots
\end{pmatrix}
\begin{pmatrix} u \\ \theta\\ q \end{pmatrix}. ∂ t u θ q = ⋯ − e 3 T e 3 T e 3 ⋯ ⋯ − H e 3 ⋯ ⋯ u θ q .
This means that the good unknown for P V PV P V comes from the first column, while the good unknown for j j j comes from the first row. Said another way: P V PV P V comes from the image of the operator, whereas j j j comes from the null set of the operator. Since that matrix is not skew-symmetric, i.e. since the operator is not skew-adjoint, it follows that P V PV P V and j j j come from different good unknowns.
Remark 3.9. To conclude this section, we note that we now have three descriptions of the state space available to us:
the “primitive” description in terms of the unknowns u u u , θ \theta θ , and q q q ,
the description in terms of the coordinates p p p , M M M , σ \sigma σ , and w w w (see Corollary 3.6), and
the description in terms of the measurements P V PV P V , M M M , j j j , a a a , and w w w (see Proposition 3.7).
Crucially: the latter two descriptions are compatible with the decomposition! We can summarize this with the following diagram, where we give a name to the maps converting between each of these descriptions.
( p , M , σ , w ) ∈ C ← C ( P V , M , j , a , w ) ∈ M ↗ S ↘ M ( u , θ , q ) ∈ L σ 2 \begin{array}{ccc}
(p,M,\sigma,w)\in\mathcal{C} & \xleftarrow{\mathcal{C}} & (PV,M,j,a,w)\in\mathcal{M} \\
& \nearrow\mathcal{S} \quad\searrow\mathcal{M} & \\
& (u,\theta,q)\in L^2_\sigma&
\end{array} ( p , M , σ , w ) ∈ C C ↗ S ↘ M ( u , θ , q ) ∈ L σ 2 ( P V , M , j , a , w ) ∈ M
Recall that L σ 2 L^2_\sigma L σ 2 is the state space introduced in Definition 2.1 and that M \mathcal{M} M is the measurement space defined in Proposition 3.7. Here C \mathcal{C} C is the coordinate space defined as follows:
C = { ( p , M , σ , w ) ∈ H 1 × L 2 × ( L σ 2 ) ⊥ × L 2 : ∇ h ⊥ ⋅ σ h = ∂ 3 w } . \mathcal{C}=\left\{(p,M,\sigma,w)\in H^1\times L^2\times(L^2_\sigma)^\perp\times L^2:\nabla_h^\perp\cdot\sigma_h=\partial_3w\right\}. C = { ( p , M , σ , w ) ∈ H 1 × L 2 × ( L σ 2 ) ⊥ × L 2 : ∇ h ⊥ ⋅ σ h = ∂ 3 w } .
The state-to-measurement map M \mathcal{M} M is then as defined in Proposition 3.7, where it is shown to be invertible, the coordinate-to-state map S \mathcal{S} S corresponds to the parametrisation recorded in Corollary 3.6, which is invertible since Theorem 3.5 tells us how to compute its inverse, and the measurement-to-coordinate map C \mathcal{C} C is well-defined by virtue of the invertibility of the other two maps.
Remark 3.10 (Vertically Sheared Horizontal Flows). This remark discusses vertically sheared horizontal flows, which are particular solutions to the Boussinesq system characterized by vanishing thermodynamic variables (namely θ \theta θ and q q q in the moist case) and by a purely z z z -dependent and horizontal velocity u = u h ( z ) u = u_h(z) u = u h ( z ) .
The purpose of this remark is two-fold. First it shows what the global change of coordinates of Corollary 3.6 and Proposition 3.7 looks like in a specific case, namely when it is applied to these vertically sheared horizontal flow. Second it shows one benefit of the decomposition presented here: these flows are accounted for naturally .
First, we recall that combining Corollary 3.6 and Proposition 3.7 gives us three descriptions of any given state.
The description in physical space in terms of ( u , θ , q ) (u,\theta,q) ( u , θ , q ) .
The description in coordinate space in terms of ( p , M , σ , w ) (p,M,\sigma,w) ( p , M , σ , w ) .
The description in measurement space in terms of ( P V , M , j , w , a ) (PV,M,j,w,a) ( P V , M , j , w , a ) .
Let us now see what vertically sheared horizontal flows look like in each of these three descriptions.
In physical space a vertically sheared horizontal flow corresponds to any state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) where θ \theta θ and q q q vanish and where u = u h ( z ) u=u_h(z) u = u h ( z ) , i.e. u 3 = 0 u_3=0 u 3 = 0 and u u u only depends on z z z .
In coordinate space a vertically sheared horizontal flow corresponds to any state described by ( p , M , σ , w ) (p,M,\sigma,w) ( p , M , σ , w ) where p p p , M M M , and w w w vanish and where σ = σ h ( z ) \sigma=\sigma_h(z) σ = σ h ( z ) .
In measurement space a vertically sheared horizontal flow corresponds to any any state described by ( P V , M , j , w , a ) (PV,M,j,w,a) ( P V , M , j , w , a ) where P V PV P V , M M M , and w w w vanish, where j = j ( z ) j=j(z) j = j ( z ) , and where a = ( a 1 , a 2 , 0 ) a=(a_1,a_2,0) a = ( a 1 , a 2 , 0 ) is any purely horizontal vector in R 3 \mathbb{R}^3 R 3 .
In other words, since a vertically sheared horizontal flows is fully characterized by a map γ : T 1 → R 2 \gamma:\mathbb{T}^1\to\mathbb{R}^2 γ : T 1 → R 2 , here are three descriptions of the same such flow.
Table 2.
Second, we note that as described above our decomposition accounts for these vertically sheared horizontal flows very naturally. This comes from the fact that our decomposition leverages a three-dimensional Helmholtz decomposition of the mixed vector fields − u h ⊥ + θ e 3 -u_h^\perp+\theta e_3 − u h ⊥ + θ e 3 (for P V PV P V ) and − u h ⊥ + ( θ − min 0 q ) e 3 -u_h^\perp+(\theta-\min_0q)e_3 − u h ⊥ + ( θ − min 0 q ) e 3 (for j j j ). This is different from what is often done where a two-dimensional Helmholtz decomposition is performed solely on the horizontal velocity u h u_h u h . This two-dimensional Helmholtz decomposition then leaves a two-dimensional average term of the form u h ( z ) u_h(z) u h ( z ) to be determined, which is precisely the term accounting for these vertically sheared horizontal flows. In other words: our decomposition accounts for these intrinsically, instead of having to treat them as a separate add-on after the fact.
Remark 3.11. It is important to note that the decomposition discussed in this section is always valid, even when the state in L σ 2 L^2_\sigma L σ 2 thus decomposed is not a solution of the moist Boussinesq dynamical system. In other words: the decomposition of Theorem 3.5 is valid for any state in the state space L σ 2 L^2_\sigma L σ 2 , similarly to how the Helmholtz decomposition is valid for any three-dimensional vector field.
However, as discussed in the next section, the decomposition we introduced in this context does also have additional meaning in the context of the dynamics of the moist Boussinesq system ((2.7a)–(2.7c)).
Decomposition and slow-fast dynamics
In § 3 above we have introduced a preliminary decomposition of the leading-order dynamics. In particular we have discussed how to then make sure that the balanced component is slow. The key step in doing so was ensuring that the pressure, which acts as a streamfunction for the balanced component of the velocity, be slow. This differs from previous work [108 , 107 , 101 , 100 ] where the streamfunction constructed was approximately slow but not precisely slow.
A question remains: “did we extract all of the slow components of a state in this manner?”. In the dry case, that question may be answered in the affirmative since the fast dynamics are that of linear oscillations which do not feedback into the slow component. In the moist case, however, the fast oscillations are nonlinear and so they may feed back into the slow component. To make this idea precise we appeal to the framework of fast-wave averaging. Fast-wave averaging is a technique used to identify the limiting behaviour of certain dynamical systems which admit two timescales. The prototypical case in which two timescales would be present is the following dynamical system in R n 1 + n 2 \mathbb{R}^{n_1+n_2} R n 1 + n 2 :
∂ t ( x 1 x 2 ) = ( f 1 ( x 1 ) 1 ε f 2 ( x 2 ) ) \partial_t
\begin{pmatrix}
x_1 \\
x_2
\end{pmatrix}
=
\begin{pmatrix}
f_1(x_1) \\
\frac{1}{\varepsilon}f_2(x_2)
\end{pmatrix} ∂ t ( x 1 x 2 ) = ( f 1 ( x 1 ) ε 1 f 2 ( x 2 ) )
for ( x 1 , x 2 ) ∈ R n 1 + n 2 (x_1,x_2)\in\mathbb{R}^{n_1+n_2} ( x 1 , x 2 ) ∈ R n 1 + n 2 , where ε > 0 \varepsilon>0 ε > 0 is a small parameter. In this case we would say that the dynamics of x 1 x_1 x 1 are slow since they are of order one, whereas the dynamics of x 2 x_2 x 2 are fast since they are of order 1 / ε 1/\varepsilon 1/ ε . In general, dynamical systems will not admit an exact two timescale decomposition as in the prototypical example above. In general, that decomposition is only approximately valid for finite ε \varepsilon ε and we are then interested in the limiting behaviour of the slow and fast components as ε → 0 \varepsilon\to0 ε → 0 .
Where does fast-wave averaging fit in all this? Well: possessing two timescales is a feature a dynamical system may or may not manifest, whereas fast-wave averaging is a technique used to identify the limiting behaviour of the slow and fast timescales (for such dynamical systems). Fast-wave averaging happens in four steps.
Assume a two-timescale asymptotic expansion of the unknowns (which requires a sub-linear growth condition).
Plug this expansion into the equations of motion to identify the resulting equations at leading-order 1 / ε 1/\varepsilon 1/ ε and at next-order.
Solve the leading-order equations over the fast timescale in terms of a yet-to-be-determined “background profile” at the slow timescale.
Determine this background profile at the slow timescale by using a solvability condition for the next-order equations.
We will now run through these four steps for the moist Boussinesq system (2.7a)–(2.7c). At each step we will highlight the differences with the dry case. This is because the dry case is well-understood however in the moist case we only provide here a sketch of the fast-wave averaging process, leaving a rigorous treatment for future work.
First we assume a two-timescale expansion of the unknowns. This is done exactly as in the dry case and we assume that, for X = ( u , θ , q ) \mathcal{X}=(u,\theta,q) X = ( u , θ , q ) ,
X = X 0 ( t , τ ) ∣ τ = t / ε + ε X 1 ( t , τ ) ∣ τ = t / ε + O ( ε 2 ) (23) \mathcal{X}=\mathcal{X}^{0}(t,\tau)\big|_{\tau=t/\varepsilon}+\varepsilon\mathcal{X}^{1}(t,\tau)\big|_{\tau=t/\varepsilon}+O(\varepsilon^{2}) \tag*{(23)} X = X 0 ( t , τ ) τ = t / ε + ε X 1 ( t , τ ) τ = t / ε + O ( ε 2 ) (23)
for some X 0 \mathcal{X}^{0} X 0 and X 1 \mathcal{X}^{1} X 1 to be determined where t t t is the slow timescale and τ = t / ε \tau=t/\varepsilon τ = t / ε is the fast timescale, subject to the sub-linear growth condition
X 1 ( t , τ ) = o ( τ ) in L 2 uniformly in t as τ → ∞ , (24) \mathcal{X}^{1}(t,\tau)=o(\tau)\text{ in }L^{2}\text{ uniformly in }t\text{ as }\tau\to\infty, \tag*{(24)} X 1 ( t , τ ) = o ( τ ) in L 2 uniformly in t as τ → ∞ , (24)
or equivalently for every δ > 0 \delta>0 δ > 0 there exists τ ∗ > 0 \tau_{*}>0 τ ∗ > 0 such that if τ > τ ∗ \tau>\tau_{*} τ > τ ∗ then
∥ X 1 ( t , τ ) ∥ L 2 τ < δ for all t . \frac{\lVert\mathcal{X}^{1}(t,\tau)\rVert_{L^{2}}}{\tau}<\delta\text{ for all }t. τ ∥ X 1 ( t , τ ) ∥ L 2 < δ for all t .
Note that this sub-linearity condition is imposed for the following reason. If the next-order term X 1 \mathcal{X}^{1} X 1 is allowed to grow linearly in τ \tau τ then we would have that
ε X 1 ( t , τ ) ∣ τ = t / ε ∼ ε ( X 1 ( t , 0 ) + t ε ) = ε X 1 ( t , 0 ) + t , \varepsilon\mathcal{X}^{1}(t,\tau)\big|_{\tau=t/\varepsilon}\sim\varepsilon\left(\mathcal{X}^{1}(t,0)+\frac{t}{\varepsilon}\right)=\varepsilon\mathcal{X}^{1}(t,0)+t, ε X 1 ( t , τ ) τ = t / ε ∼ ε ( X 1 ( t , 0 ) + ε t ) = ε X 1 ( t , 0 ) + t ,
i.e. the term t t t of order one would appear in an expression that ought to be of order ε \varepsilon ε .
The next step is to insert the two-timescale assumption (23) into the moist Boussinesq system (2.7a)–(2.7c) and read off the leading-order and next-order equations that come out. In order to carry this out we record here two preliminary computations. First we note that
∂ t [ f ( t , τ ) ∣ τ = t / ε ] = [ ( ∂ t f ) ( t , τ ) + 1 ε ( ∂ τ f ) ( t , τ ) ] τ = t / ε . \partial_t\left[f(t,\tau)\big|_{\tau=t/\varepsilon}\right]
=
\left[\left(\partial_t f\right)(t,\tau)+\frac{1}{\varepsilon}\left(\partial_\tau f\right)(t,\tau)\right]_{\tau=t/\varepsilon}. ∂ t [ f ( t , τ ) τ = t / ε ] = [ ( ∂ t f ) ( t , τ ) + ε 1 ( ∂ τ f ) ( t , τ ) ] τ = t / ε .
Second we note that, introducing the notation H : = 1 ( q 0 + ε q 1 < 0 ) H := \mathbb{1}(q^{0}+\varepsilon q^{1}<0) H := 1 ( q 0 + ε q 1 < 0 ) and H 0 : = 1 ( q 0 < 0 ) H^{0}:=\mathbb{1}(q^{0}<0) H 0 := 1 ( q 0 < 0 ) ,
min 0 ( q 0 + ε q 1 ) = ( q 0 + ε q 1 ) H = q 0 H 0 + q 0 ( H − H 0 ) ⏟ min 0 q 0 + ε q 1 H \min_{0}(q^{0}+\varepsilon q^{1})=(q^{0}+\varepsilon q^{1})H=q^{0}H^{0}+\underbrace{q^{0}(H-H^{0})}_{\min_{0}q^{0}}+\varepsilon q^{1}H 0 min ( q 0 + ε q 1 ) = ( q 0 + ε q 1 ) H = q 0 H 0 + m i n 0 q 0 q 0 ( H − H 0 ) + ε q 1 H
where
q 0 ( H − H 0 ) = q 0 [ 1 ( q 0 + ε q 1 < 0 ) − 1 ( q 0 < 0 ) ] = q 0 [ 1 ( q 0 + ε q 1 < 0 , q 0 ≥ 0 ) − 1 ( q 0 < 0 , q 0 + ε q 1 ≥ 0 ) ] = q 0 [ 1 ( 0 ≤ q 0 < − ε q 1 ) − 1 ( − ε q 1 ≤ q 0 < 0 ) ] = ε ⋅ q 0 ε [ 1 ( 0 ≤ q 0 ε < − q 1 ) − 1 ( − q 1 ≤ q 0 ε < 0 ) ] = : ε R ( q 0 ; q 1 ) . \begin{aligned}
q^{0}(H-H^{0})&=q^{0}\left[\mathbb{1}(q^{0}+\varepsilon q^{1}<0)-\mathbb{1}(q^{0}<0)\right]\\
&=q^{0}\left[\mathbb{1}(q^{0}+\varepsilon q^{1}<0,q^{0}\ge0)-\mathbb{1}(q^{0}<0,q^{0}+\varepsilon q^{1}\ge0)\right]\\
&=q^{0}\left[\mathbb{1}(0\le q^{0}<-\varepsilon q^{1})-\mathbb{1}(-\varepsilon q^{1}\le q^{0}<0)\right]\\
&=\varepsilon\cdot\frac{q^{0}}{\varepsilon}\left[\mathbb{1}\left(0\le\frac{q^{0}}{\varepsilon}<-q^{1}\right)-\mathbb{1}\left(-q^{1}\le\frac{q^{0}}{\varepsilon}<0\right)\right]\\
&=:\varepsilon R(q^{0};q^{1}).
\end{aligned} q 0 ( H − H 0 ) = q 0 [ 1 ( q 0 + ε q 1 < 0 ) − 1 ( q 0 < 0 ) ] = q 0 [ 1 ( q 0 + ε q 1 < 0 , q 0 ≥ 0 ) − 1 ( q 0 < 0 , q 0 + ε q 1 ≥ 0 ) ] = q 0 [ 1 ( 0 ≤ q 0 < − ε q 1 ) − 1 ( − ε q 1 ≤ q 0 < 0 ) ] = ε ⋅ ε q 0 [ 1 ( 0 ≤ ε q 0 < − q 1 ) − 1 ( − q 1 ≤ ε q 0 < 0 ) ] =: εR ( q 0 ; q 1 ) .
In particular we note that R ( q 0 ; q 1 ) ∈ O ( 1 ) R(q^{0};q^{1})\in O(1) R ( q 0 ; q 1 ) ∈ O ( 1 ) provided that q 1 q^{1} q 1 is bounded.
Using these two observations we plug (23) into (2.7a)–(2.7c) and read off the leading-order equations, at order 1 / ε 1/\varepsilon 1/ ε , to be
∂ τ u 0 + e 3 × u 0 + ∇ p 0 − θ 0 e 3 + min 0 q 0 e 3 = 0 , ∂ τ θ 0 + u 3 0 = 0 , and ∂ τ q 0 − u 3 0 = 0 (25) \begin{aligned}
\partial_{\tau}u^{0}+e_{3}\times u^{0}+\nabla p^{0}-\theta^{0}e_{3}+\min_{0}q^{0}e_{3}&=0,\\
\partial_{\tau}\theta^{0}+u_{3}^{0}&=0,\quad\text{and}\\
\partial_{\tau}q^{0}-u_{3}^{0}&=0 \tag*{(25)}
\end{aligned} ∂ τ u 0 + e 3 × u 0 + ∇ p 0 − θ 0 e 3 + 0 min q 0 e 3 ∂ τ θ 0 + u 3 0 ∂ τ q 0 − u 3 0 = 0 , = 0 , and = 0 (25)
while the next-order equations, at order one (i.e. order ε 0 \varepsilon^{0} ε 0 ) to be
∂ τ u 1 + e 3 × u 1 + ∇ p 1 − θ 1 e 3 + q 1 H e 3 = − ∂ t u 0 − u 0 ⋅ ∇ u 0 + R ( q 0 ; q 1 ) , ∂ τ θ 1 + u 3 1 = − ∂ t θ 0 − u 0 ⋅ ∇ θ 0 , and ∂ τ q 1 − u 3 1 = − ∂ t q 0 − u 0 ⋅ ∇ q 0 . (26) \begin{aligned}
\partial_{\tau}u^{1}+e_{3}\times u^{1}+\nabla p^{1}-\theta^{1}e_{3}+q^{1}He_{3}&=-\partial_{t}u^{0}-u^{0}\cdot\nabla u^{0}+R(q^{0};q^{1}),\\
\partial_{\tau}\theta^{1}+u_{3}^{1}&=-\partial_{t}\theta^{0}-u^{0}\cdot\nabla\theta^{0},\quad\text{and}\\
\partial_{\tau}q^{1}-u_{3}^{1}&=-\partial_{t}q^{0}-u^{0}\cdot\nabla q^{0}. \tag*{(26)}
\end{aligned} ∂ τ u 1 + e 3 × u 1 + ∇ p 1 − θ 1 e 3 + q 1 H e 3 ∂ τ θ 1 + u 3 1 ∂ τ q 1 − u 3 1 = − ∂ t u 0 − u 0 ⋅ ∇ u 0 + R ( q 0 ; q 1 ) , = − ∂ t θ 0 − u 0 ⋅ ∇ θ 0 , and = − ∂ t q 0 − u 0 ⋅ ∇ q 0 . (26)
We can also write these equations more compactly as
∂ τ X 0 + N ( X 0 ) = 0 , and ∂ τ X 1 + L H ( X 1 ) = − ∂ t X 0 − u 0 ⋅ ∇ X 0 + R ( X 0 ; X 1 ) , (27) \begin{aligned}
\partial_{\tau}\mathcal{X}^{0}+\mathcal{N}(\mathcal{X}^{0})&=0,\quad\text{and}\\
\partial_{\tau}\mathcal{X}^{1}+\mathcal{L}_{H}(\mathcal{X}^{1})&=-\partial_{t}\mathcal{X}^{0}-u^{0}\cdot\nabla\mathcal{X}^{0}+R(\mathcal{X}^{0};\mathcal{X}^{1}), \tag*{(27)}
\end{aligned} ∂ τ X 0 + N ( X 0 ) ∂ τ X 1 + L H ( X 1 ) = 0 , and = − ∂ t X 0 − u 0 ⋅ ∇ X 0 + R ( X 0 ; X 1 ) , (27)
for N \mathcal{N} N as defined in (13) , L H \mathcal{L}_{H} L H its formal linearisation given by
L H ( u θ q ) = ( P L [ e 3 × u − ( θ − q H ) e 3 ] u 3 − u 3 ) , (28) \mathcal{L}_{H}
\begin{pmatrix}
u\\
\theta\\
q
\end{pmatrix}
=
\begin{pmatrix}
\mathbb{P}_{L}\left[e_{3}\times u-(\theta-qH)e_{3}\right]\\
u_{3}\\
-u_{3}
\end{pmatrix}, \tag*{(28)} L H u θ q = P L [ e 3 × u − ( θ − q H ) e 3 ] u 3 − u 3 , (28)
and the remainder term R R R given by R ( u 0 , θ 0 , q 0 ; u 1 , θ 1 , q 1 ) = ( R ( q 0 ; q 1 ) , 0 , 0 ) R(u^{0},\theta^{0},q^{0};u^{1},\theta^{1},q^{1})=(R(q^{0};q^{1}),0,0) R ( u 0 , θ 0 , q 0 ; u 1 , θ 1 , q 1 ) = ( R ( q 0 ; q 1 ) , 0 , 0 ) .
Crucially: we see that the dynamics at leading-order are governed by ∂ τ + N = 0 \partial_{\tau}+\mathcal{N}=0 ∂ τ + N = 0 for N \mathcal{N} N as introduced in (13) . The sets B \mathcal{B} B and W \mathcal{W} W characterised in Propositions 3.1 and 3.3 thus play an essential role since they come from that same operator N \mathcal{N} N . Indeed, as discussed in Proposition A.9, a solution X \mathcal{X} X of ∂ τ X + N ( X ) = 0 \partial_{\tau}\mathcal{X}+\mathcal{N}(\mathcal{X})=0 ∂ τ X + N ( X ) = 0 may be written as
X ( τ ) = X B ( 0 ) + e − τ L W , H B X W ( 0 ) . (29) \mathcal{X}(\tau)=\mathcal{X}_{B}(0)+e^{-\tau\mathcal{L}_{W,H_{B}}}\mathcal{X}_{W}(0). \tag*{(29)} X ( τ ) = X B ( 0 ) + e − τ L W , H B X W ( 0 ) . (29)
Note that here we abused notation a little bit in order to be reminiscent of what happens in the dry case. In the dry case the leading order dynamics are ∂ τ + L \partial_{\tau}+\mathcal{L} ∂ τ + L for L \mathcal{L} L given by
L ( u , θ ) = ( P L ( u h ⊥ − θ e 3 ) u 3 ) . (30) \mathcal{L}(u,\theta)=
\begin{pmatrix}
\mathbb{P}_{L}(u_{h}^{\perp}-\theta e_{3})\\
u_{3}
\end{pmatrix}. \tag*{(30)} L ( u , θ ) = ( P L ( u h ⊥ − θ e 3 ) u 3 ) . (30)
as per (3) –(4) . Since the operator L \mathcal{L} L is linear and constant, it follows that the solution map of ∂ τ + L \partial_{\tau}+\mathcal{L} ∂ τ + L is precisely given by the semigroup e − t L e^{-t\mathcal{L}} e − t L . The operator L W , H B \mathcal{L}_{W,H_{B}} L W , H B in the moist case is linear but with variable coefficients (namely through its dependence on H B H_{B} H B ), yet we use the same notation to represent the solution map, keeping it mind that it is slightly abusive notation since the solution map is not a semigroup.
We are thus ready to carry out the third step of the fast-wave averaging process. We deduce from (29) that the solution of the leading-order equations (27) is given by
X 0 ( t , τ ) = X B 0 ( t ) + e − τ L W , H B X W 0 ( t ) \mathcal{X}^{0}(t,\tau)=\mathcal{X}_{B}^{0}(t)+e^{-\tau\mathcal{L}_{W,H_{B}}}\mathcal{X}_{W}^{0}(t) X 0 ( t , τ ) = X B 0 ( t ) + e − τ L W , H B X W 0 ( t )
where the background slow profile X 0 ( t ) = X B 0 ( t ) + X W 0 ( t ) \mathcal{X}^{0}(t)=\mathcal{X}^{0}_{B}(t)+\mathcal{X}^{0}_{W}(t) X 0 ( t ) = X B 0 ( t ) + X W 0 ( t ) is to be determined (this is done in the next step of the fast-wave averaging process). This is where the first essential difference with the dry case manifests itself. In the dry case, the leading-order solution takes the form
X 0 ( t , τ ) = X B 0 ( t ) + e − τ L W X W 0 ( t ) , (31) \mathcal{X}^{0}(t,\tau)=\mathcal{X}^{0}_{B}(t)+e^{-\tau\mathcal{L}_{W}}\mathcal{X}^{0}_{W}(t), \tag*{(31)} X 0 ( t , τ ) = X B 0 ( t ) + e − τ L W X W 0 ( t ) , (31)
where L W \mathcal{L}_{W} L W denotes the restriction of L \mathcal{L} L from (30) onto its image. Crucially: since the operator L W \mathcal{L}_{W} L W is skew-adjoint in L 2 L^{2} L 2 the oscillations in (31) are linear oscillations with constant frequencies. Here, in the moist case, the frequencies depend on the phase.
To see why this matters, suppose for now that we have completed the fourth and last step of the fast-wave averaging process. That is, suppose we have determined the dynamics of the slow part X 0 ( t ) \mathcal{X}^{0}(t) X 0 ( t ) . Since we are interested in the limiting behaviour of solutions, we then wish to study the limiting behaviour of X 0 ( t , τ ) \mathcal{X}^{0}(t,\tau) X 0 ( t , τ ) as ε → 0 \varepsilon\to0 ε → 0 . In the dry case this means studying the limiting behaviour of
e − τ L W X W 0 ( t ) . e^{-\tau\mathcal{L}_{W}}\mathcal{X}^{0}_{W}(t). e − τ L W X W 0 ( t ) .
Crucially: since the dry oscillations have constant frequencies, these oscillations have vanishing time averages, and so the weak limit of this term is simply zero!
In the moist case we must by contrast study the limiting behaviour of
e − τ L W , H B X W 0 ( t ) . e^{-\tau\mathcal{L}_{W,H_{B}}}\mathcal{X}^{0}_{W}(t). e − τ L W , H B X W 0 ( t ) .
But now, as shown in an example first recorded in [107 ] (see Corollary A.22 , and the simple example it arises from in Proposition A.19 ), phase-dependent oscillations do not have vanishing time averages. This wave term therefore converges weakly to a non-zero time average!
This tells us exactly how to augment our “pre-decomposition” above, in Theorem 3.5 , to now make sure we include all of the slow components: we take the slow balanced piece from the previous section and add to it the time-average of the wave piece.
This produces the following slow-fast decomposition, where for any space S S S , L 1 S L^{1}S L 1 S is used to denote L 1 ( [ 0 , ∞ ) ; S ) L^{1}([0,\infty);S) L 1 ([ 0 , ∞ ) ; S ) .
Theorem 4.1 (Slow-fast decomposition). We have the decomposition
L 1 L σ 2 = L 1 B + W ‾ + W ~ . L^{1}L^{2}_{\sigma}=L^{1}B+\overline{W}+\widetilde{W}. L 1 L σ 2 = L 1 B + W + W . where
W ‾ : = { X ∈ L 1 W : X is independent of time τ } and W ~ : = { X ∈ L 1 W : lim τ → ∞ 1 τ ∫ 0 τ X ( τ ′ ) d τ ′ = 0 } , \overline{W}:=\left\{\mathcal{X}\in L^{1}W:\mathcal{X}\text{ is independent of time }\tau\right\}\text{ and }\widetilde{W}:=\left\{\mathcal{X}\in L^{1}W:\lim_{\tau\to\infty}\frac{1}{\tau}\int_{0}^{\tau}\mathcal{X}(\tau')\,\mathrm{d}\tau'=0\right\}, W := { X ∈ L 1 W : X is independent of time τ } and W := { X ∈ L 1 W : τ → ∞ lim τ 1 ∫ 0 τ X ( τ ′ ) d τ ′ = 0 } , in the sense that, for every X ∈ L 1 L σ 2 \mathcal{X}\in L^{1}L^{2}_{\sigma} X ∈ L 1 L σ 2 there exist unique X B ∈ L 1 B \mathcal{X}_{B}\in L^{1}B X B ∈ L 1 B , X ‾ W ∈ W ‾ \overline{\mathcal{X}}_{W}\in\overline{W} X W ∈ W , and X ~ W ∈ W ~ \widetilde{\mathcal{X}}_{W}\in\widetilde{W} X W ∈ W such that
X = X B + X ‾ W + X ~ W \mathcal{X}=\mathcal{X}_{B}+\overline{\mathcal{X}}_{W}+\widetilde{\mathcal{X}}_{W} X = X B + X W + X W and
P V ( X ) = P V ( X B ) PV(\mathcal{X})=PV(\mathcal{X}_{B}) P V ( X ) = P V ( X B ) and M ( X ) = M ( X B ) \mathcal{M}(\mathcal{X})=\mathcal{M}(\mathcal{X}_{B}) M ( X ) = M ( X B ) for P V PV P V and M \mathcal{M} M as in Theorem 3.5 .
Moreover X B ( τ ) \mathcal{X}_{B}(\tau) X B ( τ ) and X W ( τ ) \mathcal{X}_{W}(\tau) X W ( τ ) may be computed explicitly at every instant τ \tau τ as in Theorem 3.5 such that then X ‾ W : = lim τ → ∞ 1 τ ∫ 0 τ X W ( τ ′ ) d τ ′ \overline{\mathcal{X}}_{W}:=\lim_{\tau\to\infty}\frac{1}{\tau}\int_{0}^{\tau}\mathcal{X}_{W}(\tau')\,\mathrm{d}\tau' X W := lim τ → ∞ τ 1 ∫ 0 τ X W ( τ ′ ) d τ ′ and X ~ : = X − X ‾ W \widetilde{\mathcal{X}}:=\mathcal{X}-\overline{\mathcal{X}}_{W} X := X − X W .
Proof. This follows immediately from applying Theorem 3.5 to X ( τ ) \mathcal{X}(\tau) X ( τ ) at every instant τ ≥ 0 \tau\geq0 τ ≥ 0 and then constructing X ‾ W \overline{\mathcal{X}}_{W} X W and X ~ W \widetilde{\mathcal{X}}_{W} X W as indicated.
Note that the technical aspects of this theorem were already taken care of in the proof of the “pre-decomposition” of Theorem 3.5 . So there is nothing new to prove here. Nonetheless, as explained in this section in the lead-up to Theorem 4.1 , this statement now has an interpretation that Theorem 3.5 did not. Once viewed through the lens of fast-wave averaging, we can indeed interpret Theorem 4.1 as a slow-fast decomposition:
X = X B + X ‾ W ⏟ X s l o w + X ~ W ⏟ X f a s t . \mathcal{X}=\underbrace{\mathcal{X}_{B}+\overline{\mathcal{X}}_{W}}_{\mathcal{X}_{\mathrm{slow}}}+\underbrace{\widetilde{\mathcal{X}}_{W}}_{\mathcal{X}_{\mathrm{fast}}}. X = X slow X B + X W + X fast X W .
In other words: the fast-wave averaging framework now, at least formally, ensures that this decomposition accounts for the entirety of the slow components of the solution!
Finally, for completeness we record the fourth and final step of fast-wave averaging. The idea is that we can solve (4.5b) for X 1 \mathcal{X}^{1} X 1 , but we are not particularly interested in what that solution looks like. What we care about is making sure that this next-order solution satisfies the sub-linear growth condition (4.2.
Ultimately, this comes down to a condition on the right-hand side forcing terms of (4.5b). Since these forcing terms depend on the leading-order solution X 0 \mathcal{X}^{0} X 0 , enforcing the sub-linear growth condition on X 1 \mathcal{X}^{1} X 1 turns into a condition on X 0 \mathcal{X}^{0} X 0 .
In the dry case, this condition takes the form
∫ 0 τ ∂ t X 0 + e τ ′ L ( e − τ ′ L 0 u ‾ 0 ⋅ ∇ e − τ ′ L X 0 ) d τ ′ = 0 \int_{0}^{\tau} \partial_{t}\mathcal{X}^{0}+e^{\tau'\mathcal{L}}\left(e^{-\tau'\mathcal{L}_{0}}\overline{u}_{0}\cdot\nabla e^{-\tau'\mathcal{L}}\mathcal{X}_{0}\right)\,\mathrm{d}\tau'=0 ∫ 0 τ ∂ t X 0 + e τ ′ L ( e − τ ′ L 0 u 0 ⋅ ∇ e − τ ′ L X 0 ) d τ ′ = 0
since we may use the fact that the semigroup e − τ L e^{-\tau\mathcal{L}} e − τ L is norm-preserving. In the moist case however the solution map e − τ L W , H B e^{-\tau\mathcal{L}_{W,H_{B}}} e − τ L W , H B is not norm-preserving (see Proposition A.12 ) and so we cannot perform the same trick. Identifying what the solvability condition looks like in the moist case is therefore left open for future work.
Decomposition and energy
In the dry case, the decomposition of states into their slow and fast components has the additional properties that it is an orthogonal decomposition, where the inner product used is the one induced by the conserved energy.
In the moist case, which is the focus of this paper, we therefore ask the following question: how does our decomposition of Theorem 3.5 relate to the energy conserved by the moist Boussinesq system? The energy in question is given by
E = 1 2 ∫ ∣ u ∣ 2 + θ 2 + min 0 2 q + M 2 , (32) E=\frac{1}{2}\int\lvert u\rvert^{2}+\theta^{2}+\min_{0}^{2}q+M^{2}, \tag*{(32)} E = 2 1 ∫ ∣ u ∣ 2 + θ 2 + 0 min 2 q + M 2 , (32)
where an arbitrary coefficient of the M 2 M^{2} M 2 term has been set to 1 for simplicity. For a demonstration that this energy is conserved, see Proposition A.10 .
In the dry case, since the decomposition comes down to linear projections, and since the energy gives rise to an inner product, the relation between the decomposition and the energy can be phrased as a simple question: are these projections orthogonal with respect to this inner product? In the moist case, both the decomposition and the energy are nonlinear and so it is not clear how one would go about even asking if the two are related. We offer several options in § 5.1 –§ 5.3 , summarizing and contrasting each option, along with their benefits and inconveniences, in § 5.4 .
Note that, in each of the three cases discussed below, the strategy is the same. Namely we seek to identify a way to quantify how far apart two states ( u 1 , θ 1 , q 1 ) (u_{1},\theta_{1},q_{1}) ( u 1 , θ 1 , q 1 ) and ( u 2 , θ 2 , q 2 ) (u_{2},\theta_{2},q_{2}) ( u 2 , θ 2 , q 2 ) are in such a manner that
this agrees with the decomposition of Theorem 3.5 , and gives it further meaning as a projection, in the sense that the closest balanced state to any given state must be its balanced component and
this agrees with the energy (when the two are comparable).
Finally, note that the authors believe that the quantification provided in § 5.3 is really the right one to use for describing the decomposition from a projection perspective. However, it may appear a bit odd at first. To better motivate it we therefore discuss, in § 5.1 and 5.2, alternative approaches which are less conclusive but help us understand why the good approach, discussed in § 5.3 , must take the form it does. Also, the alternative approaches in § 5.1 and 5.2 have value in their own right, for other applications beyond the focus here on a nonlinear eigenspace decomposition.
Before we dive into the details of each of these three approaches, we must elucidate the following question: what does it mean for two states to be comparable using the energy?
The question is: given two states X 1 \mathcal{X}_1 X 1 and X 2 \mathcal{X}_2 X 2 , when can we use the energy to quantify the discrepancy between X 1 \mathcal{X}_1 X 1 and X 2 \mathcal{X}_2 X 2 ?
A naive attempt would be to use the energy of their difference , i.e.
E ( X 1 − X 2 ) = ∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + min 0 2 ( q 1 − q 2 ) + ( ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 . (33) E(\mathcal{X}_1-\mathcal{X}_2)=\int|u_1-u_2|^2+(\theta_1-\theta_2)^2+\min_0^2(q_1-q_2)+((\theta_1+q_1)-(\theta_2+q_2))^2. \tag*{(33)} E ( X 1 − X 2 ) = ∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + 0 min 2 ( q 1 − q 2 ) + (( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 . (33)
This is not physically meaningful. Why? Because of the nonlinearity. When computing E ( X 1 − X 2 ) E(\mathcal{X}_1-\mathcal{X}_2) E ( X 1 − X 2 ) the nonlinear term min 0 2 ( q 1 − q 2 ) \min_0^2(q_1-q_2) min 0 2 ( q 1 − q 2 ) only turns on when q 1 − q 2 > 0 q_1-q_2>0 q 1 − q 2 > 0 . However this nonlinearity is not supposed to care about the sign of q 1 − q 2 q_1-q_2 q 1 − q 2 . Instead the nonlinearity is there to account for the fact that, depending on the signs of q 1 q_1 q 1 and q 2 q_2 q 2 , the energetic weights of q 1 q_1 q 1 and q 2 q_2 q 2 change.
What is the takeaway? The only time we can directly compare two states X 1 \mathcal{X}_1 X 1 and X 2 \mathcal{X}_2 X 2 using the energy is when the signs of q 1 q_1 q 1 and q 2 q_2 q 2 agree everywhere in the spatial domain. The sign of q q q is a mathematical proxy used to tell us which phase the state is in at any point in the spatial domain. We may thus rephrase the observation above as follows.
The only time we can directly compare two states X 1 \mathcal{X}_1 X 1 and X 2 \mathcal{X}_2 X 2 is when their phases coincide everywhere in the spatial domain.
It will therefore be convenient to encode the spatial location of a state’s phases. This is done by using the indicator function
H : = 1 ( q < 0 ) . H:=1(q<0). H := 1 ( q < 0 ) .
Note that since the water content q = q ( t , x ) q=q(t,x) q = q ( t , x ) is a function of t t t and x x x , then the indicator H H H also depends on t t t and x x x . As above, we will often use notation which suppresses the ( t , x ) (t,x) ( t , x ) -dependence in an effort to improve readability.
So let us consider two states X 1 \mathcal{X}_1 X 1 and X 2 \mathcal{X}_2 X 2 whose phases agree, i.e. for which H 1 = H 2 = : H H_1=H_2=:H H 1 = H 2 =: H . On one hand
E ( X 1 ) = ∫ ∣ u 1 ∣ 2 + θ 1 2 + min 0 2 q 1 + ( θ 1 + q 1 ) 2 = ∫ ∣ u 1 ∣ 2 + θ 1 2 + q 1 2 H + ( θ 1 + q 1 ) 2 \begin{aligned}
E(\mathcal{X}_1)=\int|u_1|^2+\theta_1^2+\min_0^2q_1+(\theta_1+q_1)^2 \\
&=\int|u_1|^2+\theta_1^2+q_1^2H+(\theta_1+q_1)^2
\end{aligned} E ( X 1 ) = ∫ ∣ u 1 ∣ 2 + θ 1 2 + 0 min 2 q 1 + ( θ 1 + q 1 ) 2 = ∫ ∣ u 1 ∣ 2 + θ 1 2 + q 1 2 H + ( θ 1 + q 1 ) 2
while on the other hand
E ( X 2 ) = ∫ ∣ u 2 ∣ 2 + θ 2 2 + min 0 2 q 2 + ( θ 2 + q 2 ) 2 = ∫ ∣ u 2 ∣ 2 + θ 2 2 + q 2 2 H + ( θ 2 + q 2 ) 2 . \begin{aligned}
E(\mathcal{X}_2)=\int|u_2|^2+\theta_2^2+\min_0^2q_2+(\theta_2+q_2)^2 \\
&=\int|u_2|^2+\theta_2^2+q_2^2H+(\theta_2+q_2)^2.
\end{aligned} E ( X 2 ) = ∫ ∣ u 2 ∣ 2 + θ 2 2 + 0 min 2 q 2 + ( θ 2 + q 2 ) 2 = ∫ ∣ u 2 ∣ 2 + θ 2 2 + q 2 2 H + ( θ 2 + q 2 ) 2 .
A natural way to compare X 1 \mathcal{X}_1 X 1 and X 2 \mathcal{X}_2 X 2 using the energy is then to treat the indicator function H H H as fixed in the expressions above for both E ( X 1 ) E(\mathcal{X}_1) E ( X 1 ) and E ( X 2 ) E(\mathcal{X}_2) E ( X 2 ) and to consider
∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 . (34) \int|u_1-u_2|^2+(\theta_1-\theta_2)^2+(q_1-q_2)^2H+((\theta_1+q_1)-(\theta_2+q_2))^2. \tag*{(34)} ∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + (( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 . (34)
This is not the same as E ( X 1 − X 2 ) E(\mathcal{X}_1-\mathcal{X}_2) E ( X 1 − X 2 ) , as can be seen by comparing (33) with (34) . Nonetheless the quantity appearing in (34) is, so far, the only way we have of using the energy to compare two states, as long as their phases agree everywhere.
So we give this quantity a name.
Definition 5.2. For any indicator function H H H and any state X = ( u , θ , q ) ∈ L σ 2 \mathcal{X}=(u,\theta,q)\in L^2_\sigma X = ( u , θ , q ) ∈ L σ 2 let
E H ( X ) : = ∫ ∣ u ∣ 2 + θ 2 + q 2 H + ( θ + q ) 2 . E_H(\mathcal{X}):=\int|u|^2+\theta^2+q^2H+(\theta+q)^2. E H ( X ) := ∫ ∣ u ∣ 2 + θ 2 + q 2 H + ( θ + q ) 2 .
Note that this notation may then be used even when (especially when!) the sign of q q q does not agree with H H H . That is precisely what happens in (34) : H H H encodes the signs of q 1 q_1 q 1 and q 2 q_2 q 2 , not the sign of q 1 − q 2 q_1-q_2 q 1 − q 2 (that was precisely the naive nonphysical approach we sought to avoid!)
Here we have motivated this “fixed H H H ” energy by talking about restricting our attention to a subset of states whose phases agree. There is another reason to look at this energy: it is the quadratic approximation of the conserved energy about a state whose phase are encoded in H H H . So this energy also appears in situations where linearisation are used, since linearising the dynamical equations is essentially the same as working with the quadratic approximation of the conserved energy. This interpretation of the “fixed H H H ” energy will be more prevalent in later sections, such as § 6.2 where geometrical notions relying on tangent spaces (a.k.a. linearisations) are discussed. Linearisations and the “fixed H H H ” energy are also relevant for applications such as adjoint models, sensitivity analysis, and variational data assimilation, which rely on tangent linear approximations to the dynamics [65 , 85 , 37 , 36 , 70 , 5 , 1 , 30 ] .
PDE-centric approach
We first discuss a PDE-centric approach to relating the decomposition and the energy. Why start here? There are two reasons to begin with a PDE-centric approach.
The first reason to take such an approach is that elliptic PDEs underpin the decomposition in both the case of the Helmholtz decomposition and in the dry case. Since a nonlinear elliptic PDE is one of the few structures that have already been identified to be present in the moist case, this PDE therefore presents a natural starting point.
The second reason to take a PDE-centric approach is that a similar approach was fruitful in identifying the slow–fast decomposition. Indeed, the slow–fast decomposition was obtained by recognizing the important role played by the slow measurements P V PV P V and M M M , and then inverting the nonlinear P V PV P V -and-M M M inversion PDE (17) in order to obtain p p p . The variables p p p and M M M were then used to characterise the balanced part of a state.
Equipped with Propositions 3.1 and 3.3 we can actually go a step further: the slow measurements P V PV P V and M M M characterise the balanced component of a state while the fast measurements j j j , w w w , and a a a characterise its wave component. Indeed, we have the following.
j ( u , θ , q ) = ∂ 3 u h − ∇ h ⊥ ( θ − min 0 q ) , w ( u , θ , q ) = u 3 , j(u,\theta,q)=\partial_3u_h-\nabla_h^\perp(\theta-\min_0 q),\quad w(u,\theta,q)=u_3, j ( u , θ , q ) = ∂ 3 u h − ∇ h ⊥ ( θ − 0 min q ) , w ( u , θ , q ) = u 3 ,
and
a ( u , θ , q ) = − ∫ u h ⊥ + ∫ ( θ − min 0 θ ) e 3 a(u,\theta,q)=-\int u_h^\perp+\int(\theta-\min_0\theta)e_3 a ( u , θ , q ) = − ∫ u h ⊥ + ∫ ( θ − 0 min θ ) e 3
are transversal to the balanced set B \mathcal{B} B , and so we may expect them to characterize elements of the wave set W \mathcal{W} W .
P V ( u , θ , q ) = ∇ h ⊥ ⋅ u h + ∂ 3 θ and M ( u , θ , q ) = θ + q PV(u,\theta,q)=\nabla_h^\perp\cdot u_h+\partial_3\theta\quad\text{and}\quad M(u,\theta,q)=\theta+q P V ( u , θ , q ) = ∇ h ⊥ ⋅ u h + ∂ 3 θ and M ( u , θ , q ) = θ + q
are transversal to the wave set W \mathcal{W} W , and so we may similarly expect them to characterize elements of the balanced set B \mathcal{B} B .
Both of these assertions turn out to be true, and thus give rise to the global change of coordinates recorded in Proposition 3.7 which allows us to translate between a state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) and the tuple of measurements ( P V , M , j , w , a ) (PV,M,j,w,a) ( P V , M , j , w , a ) .
We will now use this change of variable from Proposition 3.7 to define a distance on the state space L σ 2 \mathbb{L}^2_\sigma L σ 2 . We will do this by using the change of variables to pull-back a metric on the measurement space M \mathcal{M} M and thus produce a metric on the state space L σ 2 \mathbb{L}^2_\sigma L σ 2 (see Proposition A.13 ). The execution of this strategy requires a few steps. In order to have a clearer idea of where we are headed, we record below what would happen if we tried doing the same thing in the dry case. In the dry case we can write the energy in terms of the measurements (namely P V PV P V , j j j , w w w , and a a a ). In the language that will be used in this section, we can therefore view the energy, and the L 2 L^2 L 2 metric it corresponds to, as the pullback of a metric on the (dry) measurement space! Since we will ultimately, in this section, define the distance of interest in terms of the coordinates (think p p p , M M M , σ \sigma σ , and w w w ), we also record in Proposition 5.3 below how the energy may be written in a third way, in terms of the coordinates (which are p p p , σ \sigma σ , and w w w in the dry case).
Proposition 5.3 (Parseval identity in the dry case). For any u ∈ ( L 2 ) σ 3 u\in(L^2)^3_\sigma u ∈ ( L 2 ) σ 3 and θ ∈ L 2 \theta\in L^2 θ ∈ L 2 we have that
∫ ∣ u ∣ 2 + θ 2 ⏟ E = ∫ ∣ ∇ Δ − 1 P V ∣ 2 + ∣ ( ∇ × ) − 1 ( j + ( ∂ 3 w ) e 3 , a ) ∣ 2 = ∫ ∣ ∇ p ∣ 2 + ∣ σ ∣ 2 + w 2 ⏟ ∼ d 2 \underbrace{\int|u|^2+\theta^2}_{E}=\int|\nabla\Delta^{-1}PV|^2+\left|(\nabla\times)^{-1}\left(j+(\partial_3w)e_3,a\right)\right|^2=\underbrace{\int|\nabla p|^2+|\sigma|^2+w^2}_{\sim d^2} E ∫ ∣ u ∣ 2 + θ 2 = ∫ ∣∇ Δ − 1 P V ∣ 2 + ( ∇ × ) − 1 ( j + ( ∂ 3 w ) e 3 , a ) 2 = ∼ d 2 ∫ ∣∇ p ∣ 2 + ∣ σ ∣ 2 + w 2 where p p p solves
Δ p = ∇ h ⊥ ⋅ u h + ∂ 3 θ ⏟ P V \Delta p=\underbrace{\nabla_h^\perp\cdot u_h+\partial_3\theta}_{PV} Δ p = P V ∇ h ⊥ ⋅ u h + ∂ 3 θ
while σ \sigma σ solves
∇ × σ = ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 subject to ∫ σ = ∫ − u h ⊥ + ∫ θ e 3 , \nabla\times\sigma= \partial_3 u_h - \nabla_h^\perp\theta+ (\partial_3 w)e_3 \quad\text{subject to} \quad\int\sigma= \int-u_h^\perp+ \int\theta e_3, ∇ × σ = ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 subject to ∫ σ = ∫ − u h ⊥ + ∫ θ e 3 ,
and where w : = u 3 w := u_3 w := u 3 . Here we write σ = ( ∇ × ) − 1 ( ω , a ) \sigma= (\nabla\times)^{-1}(\omega, a) σ = ( ∇ × ) − 1 ( ω , a ) to mean that σ \sigma σ solves
∇ × σ = ω subject to ∇ ⋅ σ = 0 and ∫ σ = a . \nabla\times\sigma= \omega\quad\text{subject to} \quad\nabla\cdot\sigma= 0 \quad\text{and} \quad\int\sigma= a. ∇ × σ = ω subject to ∇ ⋅ σ = 0 and ∫ σ = a .
Proof. We recall that the decomposition in the dry case is
( u θ ) = ( ∇ h ⊥ p ∂ 3 p ) + ( σ h ⊥ + w e 3 σ 3 ) . \begin{pmatrix} u \\ \theta\end{pmatrix}
=
\begin{pmatrix} \nabla_h^\perp p \\ \partial_3 p \end{pmatrix}
+
\begin{pmatrix} \sigma_h^\perp+ w e_3 \\ \sigma_3 \end{pmatrix}. ( u θ ) = ( ∇ h ⊥ p ∂ 3 p ) + ( σ h ⊥ + w e 3 σ 3 ) .
where σ \sigma σ is divergence-free and satisfies ∇ h ⊥ ⋅ σ h = ∂ 3 w \nabla_h^\perp\cdot\sigma_h = \partial_3 w ∇ h ⊥ ⋅ σ h = ∂ 3 w . We may then compute that
P V = ∇ h ⊥ ⋅ u h + ∂ 3 θ = Δ p PV = \nabla_h^\perp\cdot u_h + \partial_3 \theta= \Delta p P V = ∇ h ⊥ ⋅ u h + ∂ 3 θ = Δ p
while Lemma A.4 tells us that
∇ × σ = ∂ 3 σ h ⊥ − ∇ h ⊥ σ 3 + ( ∇ h ⊥ ⋅ σ h ) e 3 = ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 , \nabla\times\sigma= \partial_3 \sigma_h^\perp- \nabla_h^\perp\sigma_3 + (\nabla_h^\perp\cdot\sigma_h)e_3 = \partial_3 u_h - \nabla_h^\perp\theta+ (\partial_3 w)e_3, ∇ × σ = ∂ 3 σ h ⊥ − ∇ h ⊥ σ 3 + ( ∇ h ⊥ ⋅ σ h ) e 3 = ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 ,
with moreover
∫ σ = ∫ ( − u h ⊥ + θ e 3 ) , \int\sigma= \int\left(-u_h^\perp+ \theta e_3\right), ∫ σ = ∫ ( − u h ⊥ + θ e 3 ) ,
as claimed. Finally we may compute that, since σ \sigma σ is divergence-free and hence ∫ σ ⋅ ∇ p = 0 \int\sigma\cdot\nabla p = 0 ∫ σ ⋅ ∇ p = 0 ,
E = 1 2 ∫ ∣ u ∣ 2 + θ 2 = 1 2 ∫ ∣ u h ∣ 2 + u 3 2 + θ 2 = 1 2 ∫ ∣ ∇ h p + σ h ∣ 2 + w 2 + ( ∂ 3 p + σ 3 ) 2 = 1 2 ∫ ∣ ∇ p ∣ 2 + ∣ σ ∣ 2 + w 2 , \begin{aligned}
E &= \frac{1}{2}\int|u|^2 + \theta^2 = \frac{1}{2}\int|u_h|^2 + u_3^2 + \theta^2 \\
&= \frac{1}{2}\int|\nabla_h p + \sigma_h|^2 + w^2 + (\partial_3 p + \sigma_3)^2 \\
&= \frac{1}{2}\int|\nabla p|^2 + |\sigma|^2 + w^2,
\end{aligned} E = 2 1 ∫ ∣ u ∣ 2 + θ 2 = 2 1 ∫ ∣ u h ∣ 2 + u 3 2 + θ 2 = 2 1 ∫ ∣ ∇ h p + σ h ∣ 2 + w 2 + ( ∂ 3 p + σ 3 ) 2 = 2 1 ∫ ∣∇ p ∣ 2 + ∣ σ ∣ 2 + w 2 ,
as desired.
The benefit of the approach carried out in this section, where we pull back a metric from measurement space onto state space, is that it will immediately produce a metric which agrees with the decomposition. The drawback of this approach is that it does not, inherently, relate to the conserved energy in any way. To see why a metric produced in this way will automatically agree with the decomposition relies on the following observation: if two states have the same P V PV P V and M M M , then they must have the same balanced component. This is proved in Lemma 5.4 below. Therefore projecting a state onto the balanced set is the same as keeping its P V PV P V and M M M measurements fixed and sending its remaining measurements, namely j j j , w w w , and a a a , to zero. This is why a metric pulled back from the measurement space M \mathcal{M} M will necessarily agree with the decomposition. As mentioned above, this relies on the following result.
Lemma 5.4 (P V PV P V and M M M characterize balanced components). Given any X ∈ L σ 2 \mathcal{X} \in L_\sigma^2 X ∈ L σ 2 the unique balanced state in B \mathcal{B} B whose P V PV P V and M M M agree with the P V PV P V and M M M of X \mathcal{X} X is precisely the balanced component X B \mathcal{X}_{B} X B .
Proof. Let X ∈ L σ 2 \mathcal{X} \in L_\sigma^2 X ∈ L σ 2 and let Y ∈ B \mathcal{Y} \in\mathcal{B} Y ∈ B be a balanced state whose P V PV P V and M M M agrees with those of X \mathcal{X} X , i.e.
P V ( X ) = P V ( Y ) and M ( X ) = M ( Y ) . PV(\mathcal{X}) = PV(\mathcal{Y}) \quad\text{and} \quad M(\mathcal{X}) = M(\mathcal{Y}). P V ( X ) = P V ( Y ) and M ( X ) = M ( Y ) .
We know from Proposition 3.1 that since Y \mathcal{Y} Y is balanced it is uniquely characterised by p ∈ H ˚ 1 p \in\mathring{H}^1 p ∈ H ˚ 1 and M ∈ L 2 M \in L^2 M ∈ L 2 via Y = Φ ( p , M ) \mathcal{Y} = \Phi(p,M) Y = Φ ( p , M ) . Moreover we may compute that
M ( Φ ( p , M ) ) = M and P V ( Φ ( p , M ) ) = Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) . M(\Phi(p,M)) = M \quad\text{and} \quad PV(\Phi(p,M)) = \Delta p + \frac{1}{2}\partial_3 \min_0(M-\partial_3p). M ( Φ ( p , M )) = M and P V ( Φ ( p , M )) = Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) .
By the well-posedness of nonlinear P V PV P V -and-M M M inversion, we know that there is a unique p p p solving
Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) = P V ( Y ) = P V ( X ) , \Delta p + \frac{1}{2}\partial_3 \min_0(M-\partial_3p) = PV(\mathcal{Y}) = PV(\mathcal{X}), Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) = P V ( Y ) = P V ( X ) ,
and Proposition 3.1 tells us that the resulting state is precisely X B \mathcal{X}_{B} X B . This verifies that indeed Y = X B \mathcal{Y} = \mathcal{X}_{B} Y = X B .
Using Lemma 5.4 above we could then immediately deduce that a distance on L σ 2 L^2_\sigma L σ 2 defined via
d ( X 1 , X 2 ) 2 : = ∥ M ( X 1 ) − M ( X 2 ) ∥ m 2 = ∥ P V 1 − P V 2 ∥ H − 1 2 + ∥ M 1 − M 2 ∥ L 2 2 + ∥ j 1 − j 2 ∥ H − 1 2 + ∥ w 1 − w 2 ∥ L 2 2 + ∣ a 1 − a 2 ∣ 2 (35) \begin{aligned}
d(\mathcal{X}_1,\mathcal{X}_2)^2 := \lVert\mathcal{M}(\mathcal{X}_1)-\mathcal{M}(\mathcal{X}_2)\rVert_{\mathfrak{m}}^2 \\
&= \lVert PV_1-PV_2\rVert_{H^{-1}}^2+\lVert M_1-M_2\rVert_{L^2}^2+\lVert j_1-j_2\rVert_{H^{-1}}^2+\lVert w_1-w_2\rVert_{L^2}^2+\lvert a_1-a_2\rvert^2 \tag*{(35)}
\end{aligned} d ( X 1 , X 2 ) 2 := ∥ M ( X 1 ) − M ( X 2 ) ∥ m 2 = ∥ P V 1 − P V 2 ∥ H − 1 2 + ∥ M 1 − M 2 ∥ L 2 2 + ∥ j 1 − j 2 ∥ H − 1 2 + ∥ w 1 − w 2 ∥ L 2 2 + ∣ a 1 − a 2 ∣ 2 (35)
would automatically agree with the decomposition. However this distance has no chance of agreeing with the energy! This is because its dimensions, or units, are all wrong. Indeed: this naive distance above treats P V PV P V and M M M on the same footing even though they do not have the same dimensions. In order to produce a metric on the state space L σ 2 L^2_\sigma L σ 2 pulled back from the measurement space M \mathcal{M} M which has a chance of agreeing with the energy we must therefore dimensionalize the naive distance from (35) .
In order to perform this dimensionalisation we will use the fact that the “fixed H H H ” energy introduced in Definition 5.2 may sometimes be written not in terms of the state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) but in terms of its coordinates ( p , M , σ , w ) (p,M,\sigma,w) ( p , M , σ , w ) . To see this, we first define the following.
Definition 5.5. Fix an indicator function H H H . We define, for any ( p , M , σ , w ) ∈ C (p,M,\sigma,w)\in\mathcal{C} ( p , M , σ , w ) ∈ C ,
E ~ H ( p , M , σ , w ) : = 1 2 ∫ A H ∇ p ⋅ ∇ p + ( 1 + H 2 ) M 2 + A H − 1 σ ⋅ σ + w 2 , \widetilde{E}_H(p,M,\sigma,w):=\frac{1}{2}\int A_H\nabla p\cdot\nabla p+\left(1+\frac{H}{2}\right)M^2+A_H^{-1}\sigma\cdot\sigma+w^2, E H ( p , M , σ , w ) := 2 1 ∫ A H ∇ p ⋅ ∇ p + ( 1 + 2 H ) M 2 + A H − 1 σ ⋅ σ + w 2 , where A H = I − H 2 e 3 ⊗ e 3 A_H=I-\frac{H}{2}e_3\otimes e_3 A H = I − 2 H e 3 ⊗ e 3 and so A H − 1 = I + H e 3 ⊗ e 3 A_H^{-1}=I+He_3\otimes e_3 A H − 1 = I + H e 3 ⊗ e 3 .
In general the form of the fixed Heaviside energy provided in Definition 5.5 does not agree with the form of the fixed Heaviside energy provided in Definition 5.2 . Sometimes, they do.
Proposition 5.7 (Rewriting the energy). Let E H E_H E H be as introduced in Definition 5.2 and E ~ H \widetilde{E}_H E H be as in Definition 5.5 . Let ( u , θ , q ) ∈ L σ 2 (u,\theta,q)\in L^2_\sigma ( u , θ , q ) ∈ L σ 2 be corresponding coordinates ( p , M , σ , w ) (p,M,\sigma,w) ( p , M , σ , w ) as per Corollary 3.6. In general we have that
E H ( u , θ , q ) = 1 2 ∫ ∣ u h ∣ 2 + u 3 2 + ( 1 − H 2 ) ( θ − q H ) 2 + ( 1 + H 2 ) ( θ + q ) 2 . E_H(u,\theta,q)=\frac{1}{2}\int\lvert u_h\rvert^2+u_3^2+\left(1-\frac{H}{2}\right)(\theta-qH)^2+\left(1+\frac{H}{2}\right)(\theta+q)^2. E H ( u , θ , q ) = 2 1 ∫ ∣ u h ∣ 2 + u 3 2 + ( 1 − 2 H ) ( θ − q H ) 2 + ( 1 + 2 H ) ( θ + q ) 2 . Moreover, if q q q and its balanced component q B q_B q B share the same interface, i.e. if
H : = 1 ( q < 0 ) = 1 ( q B < 0 ) , H:=\mathbb{1}(q<0)=\mathbb{1}(q_B<0), H := 1 ( q < 0 ) = 1 ( q B < 0 ) , where note that 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) \mathbb{1}(q_B<0)=\mathbb{1}(M<\partial_3p) 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) , then
E H ( u , θ , q ) = E ~ H ( p , M , σ , w ) . E_H(u,\theta,q)=\widetilde{E}_H(p,M,\sigma,w). E H ( u , θ , q ) = E H ( p , M , σ , w ) .
Proof. The first identity follows from an immediate computation since H 2 = H H^2=H H 2 = H and hence
θ 2 + q 2 H = ( 1 − H 2 ) ( θ − q H ) 2 + H 2 ( θ + q ) 2 . \theta^2+q^2H=\left(1-\frac{H}{2}\right)(\theta-qH)^2+\frac{H}{2}(\theta+q)^2. θ 2 + q 2 H = ( 1 − 2 H ) ( θ − q H ) 2 + 2 H ( θ + q ) 2 .
To deduce the second identity we first note that Corollary 3.6 and Lemma A.7 tell us that, if H = H B : = 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) H=H_B:=\mathbb{1}(q_B<0)=\mathbb{1}(M<\partial_3p) H = H B := 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) , then
{ u h = ∇ h ⊥ p + σ h ⊥ , u 3 = w , θ − q H = ∂ 3 p + ( 1 + H ) σ 3 , and θ + q = M . \begin{cases}
u_h=\nabla_h^\perp p+\sigma_h^\perp,\\
u_3=w,\\
\theta-qH=\partial_3p+(1+H)\sigma_3,\ \text{and}\\
\theta+q=M.
\end{cases} ⎩ ⎨ ⎧ u h = ∇ h ⊥ p + σ h ⊥ , u 3 = w , θ − q H = ∂ 3 p + ( 1 + H ) σ 3 , and θ + q = M .
Therefore
∫ ∣ u h ∣ 2 + u 3 2 + ( 1 − H 2 ) ( θ − q H ) 2 + ( 1 + H 2 ) ( θ + q ) 2 = ∫ ∣ ∇ h p + σ h ∣ 2 + w 2 + ( 1 − H 2 ) ( ∂ 3 p + ( 1 + H ) σ 3 ) 2 + ( 1 + H 2 ) M 2 . \begin{aligned}
\int|u_h|^2+u_3^2+\left(1-\frac{H}{2}\right)(\theta-qH)^2+\left(1+\frac{H}{2}\right)(\theta+q)^2 \\
&=\int|\nabla_h p+\sigma_h|^2+w^2+\left(1-\frac{H}{2}\right)\left(\partial_3p+(1+H)\sigma_3\right)^2+\left(1+\frac{H}{2}\right)M^2.
\end{aligned} ∫ ∣ u h ∣ 2 + u 3 2 + ( 1 − 2 H ) ( θ − q H ) 2 + ( 1 + 2 H ) ( θ + q ) 2 = ∫ ∣ ∇ h p + σ h ∣ 2 + w 2 + ( 1 − 2 H ) ( ∂ 3 p + ( 1 + H ) σ 3 ) 2 + ( 1 + 2 H ) M 2 .
Finally, for A H A_H A H and A H − 1 A_H^{-1} A H − 1 as in Definition 5.5 we compute that, since A H A_H A H and A H − 1 A_H^{-1} A H − 1 are symmetric and since σ \sigma σ is divergence-free,
∫ ∣ ∇ h p + σ h ∣ 2 + ( 1 − H 2 ) ( ∂ 3 p + ( 1 + H ) σ 3 ) 2 = ∫ A H ( ∇ p + A H − 1 σ ) ⋅ ( ∇ p + A H − 1 σ ) = ∫ A H ∇ p ⋅ ∇ p + 2 ∫ ∇ p ⋅ σ ⏟ = 0 + ∫ A H − 1 σ ⋅ σ . \begin{aligned}
\int|\nabla_h p+\sigma_h|^2+\left(1-\frac{H}{2}\right)\left(\partial_3p+(1+H)\sigma_3\right)^2
=\int A_H(\nabla p+A_H^{-1}\sigma)\cdot(\nabla p+A_H^{-1}\sigma) \\
&=\int A_H\nabla p\cdot\nabla p+2\underbrace{\int\nabla p\cdot\sigma}_{=0}+\int A_H^{-1}\sigma\cdot\sigma.
\end{aligned} ∫ ∣ ∇ h p + σ h ∣ 2 + ( 1 − 2 H ) ( ∂ 3 p + ( 1 + H ) σ 3 ) 2 = ∫ A H ( ∇ p + A H − 1 σ ) ⋅ ( ∇ p + A H − 1 σ ) = ∫ A H ∇ p ⋅ ∇ p + 2 = 0 ∫ ∇ p ⋅ σ + ∫ A H − 1 σ ⋅ σ .
The second identity follows.
This is a good point to take a stock. Recall that we seek to define a metric pulled back from the measurement space M \mathcal{M} M onto the state space L σ 2 L^2_\sigma L σ 2 . Such a metric will automatically agree with the decomposition but, in order to make it agree with the energy, we must dimensionalize it appropriately. The identity recorded in Proposition 5.7 above tells us how to do this. The last remaining piece required to define our pullback metric is thus to identify how the measurements ( P V , M , j , w , and a ) (PV,M,j,w,\text{ and }a) ( P V , M , j , w , and a ) will determine the coordinates ( p , M , σ , and w ) (p,M,\sigma,\text{ and }w) ( p , M , σ , and w ) . Note that such a map is already implicit in Corollary 3.6, however we will do something slightly different: since the “fixed H H H ” energy already uses a fixed Heaviside, we will leverage this to define a map taking measurements to coordinates in a linear way. This linearity will make the ensuing pullback distance easier to work with.
Lemma 5.8 (Definition and invertibility of the fixed Heaviside parametrisation C H C_H C H ). Fix an indicator function H H H and define A H : = I − H 2 e 3 ⊗ e 3 A_H:=I-\frac{H}{2}e_3\otimes e_3 A H := I − 2 H e 3 ⊗ e 3 , such that A H − 1 = I + H e 3 ⊗ e 3 A_H^{-1}=I+He_3\otimes e_3 A H − 1 = I + H e 3 ⊗ e 3 (since ( 1 − H / 2 ) ( 1 + H ) ≡ 1 (1-H/2)(1+H)\equiv1 ( 1 − H /2 ) ( 1 + H ) ≡ 1 ). We define C H : M → C C_H:\mathcal{M}\to\mathcal{C} C H : M → C , where C \mathcal{C} C is the coordinate space defined in (21) . Given ( P V , M , j , w , a ) ∈ M (PV,M,j,w,a)\in\mathcal{M} ( P V , M , j , w , a ) ∈ M , let M : = M M:=M M := M , w : = w w:=w w := w , and let p p p and σ \sigma σ solve
∇ ⋅ ( A H ∇ p ) = P V − 1 2 ∂ 3 ( H M ) (36) \nabla\cdot(A_H\nabla p)=PV-\frac{1}{2}\partial_3(HM) \tag*{(36)} ∇ ⋅ ( A H ∇ p ) = P V − 2 1 ∂ 3 ( H M ) (36) and
∇ × ( A H − 1 σ ) = j + ( ∂ 3 w ) e 3 subject to ∇ ⋅ σ = 0 and ∫ A H − 1 σ = a . (37) \nabla\times(A_H^{-1}\sigma)=j+(\partial_3w)e_3\quad\text{subject to}\quad\nabla\cdot\sigma=0\quad\text{and}\quad\int A_H^{-1}\sigma=a. \tag*{(37)} ∇ × ( A H − 1 σ ) = j + ( ∂ 3 w ) e 3 subject to ∇ ⋅ σ = 0 and ∫ A H − 1 σ = a . (37) The map C H C_H C H is well-defined and invertible.
Note that if H = H B : = 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) H=H_B:=\mathbb{1}(q_B<0)=\mathbb{1}(M<\partial_3p) H = H B := 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) then C H C_H C H is precisely the map implicit in Corollary 3.6 which assigns the coordinates p p p , M M M , σ \sigma σ , and w w w to a state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) through its measurements P V PV P V , M M M , j j j , w w w , and a a a .
Proof of Lemma 5.8. It suffices to show that both (36) and (37) have a unique solution once P V PV P V , M M M , j j j , w w w , and a a a are fixed. The well-posedness of (36) follows from the fact that A H A_H A H is uniformly elliptic since A H ( x ) ≥ 1 2 I A_H(x)\geq\frac{1}{2}I A H ( x ) ≥ 2 1 I for all x x x . The well-posedness of (37) follows from the fact that A H − 1 A_H^{-1} A H − 1 is uniformly elliptic, since A H − 1 ( x ) ≥ I A_H^{-1}(x)\geq I A H − 1 ( x ) ≥ I for all x x x , and so Proposition A.8 applies to (37) .
We are now ready to define a dimensionalized version of the naive pullback metric of (35) .
Definition 5.10 (Dimensionalized Parseval distance). Fix an indicator function H H H . Define d H : L σ 2 × L σ 2 → [ 0 , ∞ ) d_H:\mathcal{L}^2_\sigma\times\mathcal{L}^2_\sigma\to[0,\infty) d H : L σ 2 × L σ 2 → [ 0 , ∞ ) as follows:
1 2 d H ( X 1 , X 2 ) 2 = E ~ H ( C H [ M ( X 1 ) − M ( X 2 ) ] ) . \frac{1}{2}d_H(\mathcal{X}_1,\mathcal{X}_2)^2=\widetilde{E}_H\left(C_H\left[\mathcal{M}(\mathcal{X}_1)-\mathcal{M}(\mathcal{X}_2)\right]\right). 2 1 d H ( X 1 , X 2 ) 2 = E H ( C H [ M ( X 1 ) − M ( X 2 ) ] ) . In other words, given X 1 , X 2 ∈ L σ 2 \mathcal{X}_1,\mathcal{X}_2\in\mathcal{L}^2_\sigma X 1 , X 2 ∈ L σ 2 , let
M : = M ( X 1 ) − M ( X 2 ) , w : = w ( X 1 ) − w ( X 2 ) , and a : = a ( X 1 ) − a ( X 2 ) , M:=\mathcal{M}(\mathcal{X}_1)-\mathcal{M}(\mathcal{X}_2),\quad w:=w(\mathcal{X}_1)-w(\mathcal{X}_2),\quad\text{and}\quad a:=a(\mathcal{X}_1)-a(\mathcal{X}_2), M := M ( X 1 ) − M ( X 2 ) , w := w ( X 1 ) − w ( X 2 ) , and a := a ( X 1 ) − a ( X 2 ) , let p p p solve, for P V : = P V ( X 1 ) − P V ( X 2 ) PV:=PV(\mathcal{X}_1)-PV(\mathcal{X}_2) P V := P V ( X 1 ) − P V ( X 2 ) ,
∇ ⋅ ( A H ∇ p ) = P V − 1 2 ∂ 3 ( H M ) \nabla\cdot(A_H\nabla p)=PV-\frac{1}{2}\partial_3(HM) ∇ ⋅ ( A H ∇ p ) = P V − 2 1 ∂ 3 ( H M ) and let σ \sigma σ solve, for j : = j ( X 1 ) − j ( X 2 ) j:=j(\mathcal{X}_1)-j(\mathcal{X}_2) j := j ( X 1 ) − j ( X 2 ) ,
∇ × ( A H − 1 σ ) = j + ( ∂ 3 w ) e 3 subject to ∫ A H − 1 σ = a and ∇ ⋅ σ = 0. \nabla\times(A_H^{-1}\sigma)=j+(\partial_3w)e_3\quad\text{subject to}\quad\int A_H^{-1}\sigma=a\quad\text{and}\quad\nabla\cdot\sigma=0. ∇ × ( A H − 1 σ ) = j + ( ∂ 3 w ) e 3 subject to ∫ A H − 1 σ = a and ∇ ⋅ σ = 0. Then
1 2 d H ( X 1 , X 2 ) 2 = ∫ Q H + ( p ) + ( 1 + H 2 ) M 2 + Q H − ( σ ) + w 2 \frac{1}{2}d_H(\mathcal{X}_1,\mathcal{X}_2)^2=\int Q_H^+(p)+\left(1+\frac{H}{2}\right)M^2+Q_H^-(\sigma)+w^2 2 1 d H ( X 1 , X 2 ) 2 = ∫ Q H + ( p ) + ( 1 + 2 H ) M 2 + Q H − ( σ ) + w 2 for Q H ± Q_H^\pm Q H ± as introduced in Definition 5.5 .
Proposition 5.11. For any indicator function H H H the function d H d_H d H defined in Definition 5.10 is a metric which agrees with the extraction of a balanced component. Moreover this distance is compatible with the energy when H = H 1 = H B , 1 = H 2 = H B , 2 H=H_1=H_{B,1}=H_2=H_{B,2} H = H 1 = H B , 1 = H 2 = H B , 2 , i.e. in that case
1 2 d H ( X 1 , X 2 ) 2 = E H ( X 1 − X 2 ) \frac{1}{2}d_H(\mathcal{X}_1,\mathcal{X}_2)^2=E_H(\mathcal{X}_1-\mathcal{X}_2) 2 1 d H ( X 1 , X 2 ) 2 = E H ( X 1 − X 2 ) = ∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( M 1 − M 2 ) 2 =\int|u_1-u_2|^2+(\theta_1-\theta_2)^2+(q_1-q_2)^2H+(M_1-M_2)^2 = ∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( M 1 − M 2 ) 2 where H = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) H=\mathbf{1}(q_1<0)=\mathbf{1}(q_2<0) H = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) .
Proof. First we show that d H d_H d H is a metric. Since E ~ H \widetilde{E}_H E H is positive-definite and viewed here as taking arguments in C \mathcal{C} C it is a quadratic form which induces an inner product, and hence a metric which we denote d E d_E d E , on C \mathcal{C} C . Since C H C_H C H is linear and since both C H C_H C H and M \mathcal{M} M are invertible (see Proposition 3.7 and Lemma 5.8 ) and hence injective it follows from Proposition A.13 that d H d_H d H is the pullback metric of d E d_E d E under C H ∘ M C_H\circ\mathcal{M} C H ∘ M . So d H d_H d H is indeed a metric on L σ 2 \mathcal{L}^2_\sigma L σ 2 .
Now we show that d H d_H d H conditionally agrees with the energy. Suppose that H = H 1 = H B , 1 = H 2 = H B , 2 H=H_1=H_{B,1}=H_2=H_{B,2} H = H 1 = H B , 1 = H 2 = H B , 2 . The argument is easier to follow if we rewrite d H d_H d H as follows, using the fact that C H C_H C H is linear:
1 2 d H ( X 1 , X 2 ) 2 = E ~ H [ C H ( M ( X 1 ) ) − C H ( M ( X 2 ) ) ] . \frac{1}{2}d_H(\mathcal{X}_1,\mathcal{X}_2)^2=\widetilde{E}_H\left[C_H(\mathcal{M}(\mathcal{X}_1))-C_H(\mathcal{M}(\mathcal{X}_2))\right]. 2 1 d H ( X 1 , X 2 ) 2 = E H [ C H ( M ( X 1 )) − C H ( M ( X 2 )) ] .
Let ( p 1 , M 1 , σ 1 , w 1 ) : = C H ( M ( X 1 ) ) (p_1,M_1,\sigma_1,w_1):=C_H(\mathcal{M}(\mathcal{X}_1)) ( p 1 , M 1 , σ 1 , w 1 ) := C H ( M ( X 1 )) . Since H = H B , 1 = 1 ( M 1 < ∂ 3 p 1 ) H=H_{B,1}=\mathbf{1}(M_1<\partial_3p_1) H = H B , 1 = 1 ( M 1 < ∂ 3 p 1 ) we see that
∇ ⋅ ( A H ∇ p 1 ) = P V 1 − 1 2 ∂ 3 ( H M 1 ) \nabla\cdot(A_H\nabla p_1)=PV_1-\frac{1}{2}\partial_3(HM_1) ∇ ⋅ ( A H ∇ p 1 ) = P V 1 − 2 1 ∂ 3 ( H M 1 )
⟺ Δ p 1 + 1 2 ∂ 3 min 0 ( M 1 − ∂ 3 p 1 ) = P V 1 . (38) \Longleftrightarrow\Delta p_1+\frac{1}{2}\partial_3\min_0(M_1-\partial_3p_1)=PV_1. \tag*{(38)} ⟺ Δ p 1 + 2 1 ∂ 3 0 min ( M 1 − ∂ 3 p 1 ) = P V 1 . (38)
Similarly, since H 1 = H B , 1 H_1=H_{B,1} H 1 = H B , 1 we have that
j 1 = ∂ 3 u h , 1 − ∇ h ⊥ ( θ 1 − min 0 q 1 ) = ∂ 3 u h , 1 − ∇ h ⊥ ( θ 1 − H B , 1 q 1 ) \begin{aligned}
j_1=\partial_3u_{h,1}-\nabla_h^\perp(\theta_1-\min_0q_1) \\
&=\partial_3u_{h,1}-\nabla_h^\perp(\theta_1-H_{B,1}q_1)
\end{aligned} j 1 = ∂ 3 u h , 1 − ∇ h ⊥ ( θ 1 − 0 min q 1 ) = ∂ 3 u h , 1 − ∇ h ⊥ ( θ 1 − H B , 1 q 1 )
and so σ 1 \sigma_1 σ 1 solves
∇ × ( A H − 1 σ 1 ) = ∂ 3 u h , 1 − ∇ h ⊥ ( θ 1 − H B , 1 q 1 ) + ( ∂ 3 w 1 ) e 3 . (39) \nabla\times\left(A_H^{-1}\sigma_1\right)=\partial_3u_{h,1}-\nabla_h^\perp\left(\theta_1-H_{B,1}q_1\right)+\left(\partial_3w_1\right)e_3. \tag*{(39)} ∇ × ( A H − 1 σ 1 ) = ∂ 3 u h , 1 − ∇ h ⊥ ( θ 1 − H B , 1 q 1 ) + ( ∂ 3 w 1 ) e 3 . (39)
We deduce from (38) and (39) that p 1 p_1 p 1 and σ 1 \sigma_1 σ 1 are precisely characterized as in Theorem 3.5 by
X 1 = Φ ( p 1 , M 1 ) + Ψ ( σ 1 , w 1 ) . \mathcal{X}_1=\Phi(p_1,M_1)+\Psi(\sigma_1,w_1). X 1 = Φ ( p 1 , M 1 ) + Ψ ( σ 1 , w 1 ) .
In other words: the approximate inversion C H ∘ M C_H\circ\mathcal{M} C H ∘ M is exact when H = H 1 = H B , 1 H=H_1=H_{B,1} H = H 1 = H B , 1 . Moreover Proposition 5.7 tells us that E ~ H \widetilde{E}_H E H agrees with E H E_H E H and with the nonlinear energy when H = H 1 = H B , 1 H=H_1=H_{B,1} H = H 1 = H B , 1 and so we conclude that indeed if H = H 1 = H B , 1 = H 2 = H B , 2 H=H_1=H_{B,1}=H_2=H_{B,2} H = H 1 = H B , 1 = H 2 = H B , 2 then
1 2 d H ( X 1 , X 2 ) 2 = E H ( X 1 − X 2 ) = ∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( M 1 − M 2 ) 2 , \begin{aligned}
\frac{1}{2}d_H(\mathcal{X}_1,\mathcal{X}_2)^2=E_H(\mathcal{X}_1-\mathcal{X}_2) \\
&=\int|u_1-u_2|^2+(\theta_1-\theta_2)^2+(q_1-q_2)^2H+(M_1-M_2)^2,
\end{aligned} 2 1 d H ( X 1 , X 2 ) 2 = E H ( X 1 − X 2 ) = ∫ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( M 1 − M 2 ) 2 ,
where H = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) H=\mathbf{1}(q_1<0)=\mathbf{1}(q_2<0) H = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) .
Finally we show that d H d_H d H agrees with the extraction of a balanced component. We fix X ∈ L σ 2 \mathcal{X}\in L_\sigma^2 X ∈ L σ 2 and let Y ∈ B \mathcal{Y}\in\mathcal{B} Y ∈ B . Since Y ∈ B \mathcal{Y}\in\mathcal{B} Y ∈ B we know from Proposition 3.1 that
j ( Y ) = 0 , w ( Y ) = 0 , and a ( Y ) = 0. j(\mathcal{Y})=0,\quad w(\mathcal{Y})=0,\quad\text{and}\quad a(\mathcal{Y})=0. j ( Y ) = 0 , w ( Y ) = 0 , and a ( Y ) = 0.
Let us define
M : = M ( X ) − M ( Y ) and P V : = P V ( X ) − P V ( Y ) , (40) M:=M(\mathcal{X})-M(\mathcal{Y})\quad\text{and}\quad PV:=PV(\mathcal{X})-PV(\mathcal{Y}), \tag*{(40)} M := M ( X ) − M ( Y ) and P V := P V ( X ) − P V ( Y ) , (40)
and let p p p be the solution of
∇ ⋅ ( A H ∇ p ) = P V − 1 2 ∂ 3 ( H M ) (41) \nabla\cdot(A_H\nabla p)=PV-\frac{1}{2}\partial_3(HM) \tag*{(41)} ∇ ⋅ ( A H ∇ p ) = P V − 2 1 ∂ 3 ( H M ) (41)
while σ \sigma σ is the solution of
∇ × ( A H − 1 σ ) = j ( X ) + [ ∂ 3 w ( X ) ] e 3 subject to ∫ A H − 1 σ = a ( X ) and ∇ ⋅ σ = 0. \nabla\times(A_H^{-1}\sigma)=j(\mathcal{X})+[\partial_3w(\mathcal{X})]e_3\quad\text{subject to}\quad\int A_H^{-1}\sigma=a(\mathcal{X})\quad\text{and}\quad\nabla\cdot\sigma=0. ∇ × ( A H − 1 σ ) = j ( X ) + [ ∂ 3 w ( X )] e 3 subject to ∫ A H − 1 σ = a ( X ) and ∇ ⋅ σ = 0.
Then we may write the distance between X \mathcal{X} X and Y \mathcal{Y} Y via
1 2 d H ( X , Y ) 2 = ∫ Q H + ( p ) + ( 1 + H 2 ) M 2 + Q H − ( σ ) + w 2 \frac{1}{2}d_H(\mathcal{X},\mathcal{Y})^2=\int Q_H^+(p)+\left(1+\frac{H}{2}\right)M^2+Q_H^-(\sigma)+w^2 2 1 d H ( X , Y ) 2 = ∫ Q H + ( p ) + ( 1 + 2 H ) M 2 + Q H − ( σ ) + w 2
for Q H ± Q_H^\pm Q H ± as in Definition 5.5. Crucially: σ \sigma σ and w w w only depend on X \mathcal{X} X ! So we may simplify the minimization problem as follows:
arg min Y ∈ B d H ( X , Y ) = arg min Y ∈ B ∫ Q H + ( p ) + ( 1 + H 2 ) M 2 . \underset{\mathcal{Y}\in\mathcal{B}}{\arg\min}\ d_H(\mathcal{X},\mathcal{Y})=\underset{\mathcal{Y}\in\mathcal{B}}{\arg\min}\ \int Q_H^+(p)+\left(1+\frac{H}{2}\right)M^2. Y ∈ B arg min d H ( X , Y ) = Y ∈ B arg min ∫ Q H + ( p ) + ( 1 + 2 H ) M 2 .
In particular, since Q H + ( v ) = A H v ⋅ v ≥ 1 2 ∣ v ∣ 2 Q_H^+(v)=A_Hv\cdot v\geq\frac{1}{2}|v|^2 Q H + ( v ) = A H v ⋅ v ≥ 2 1 ∣ v ∣ 2 , i.e. since Q H + Q_H^+ Q H + is positive-definite, the unique minimizer corresponds to
p = 0 and M = 0. p=0\quad\text{and}\quad M=0. p = 0 and M = 0.
Plugging this into (40) , and then into (41) , this means that M ( X ) = M ( Y ) M(\mathcal{X})=M(\mathcal{Y}) M ( X ) = M ( Y ) , and hence that P V ( X ) = P V ( Y ) PV(\mathcal{X})=PV(\mathcal{Y}) P V ( X ) = P V ( Y ) . In other words the unique minimizer is the balanced state Y \mathcal{Y} Y whose P V PV P V and M M M agree with X \mathcal{X} X . By Lemma 5.4, this means precisely that Y = X B \mathcal{Y}=\mathcal{X}_B Y = X B , as desired.
This concludes our first attempt at quantifying how far apart two states are in a way that agrees with
the extraction of a balanced component and
the conserved energy.
We see that the first of these two requirements was met by the distance introduced in Definition 5.10 above. However the second requirement was only met in the restricted setting where the fixed Heaviside agreed with the Heavisides arising from both states being compared.
In some sense, that only the first requirement is satisfied is not surprising. Indeed, the two distances introduced in this section are built on top of the measurements from the global change of coordinates of Proposition 3.7 and these measurements have been chosen precisely to be compatible with the decomposition. However, these measurements have, a priori, little to do with the conserved energy and so it is actually quite positive that the distance built in this way is in any way comparable to the energy.
This motivates an alternate approach: what if, instead of using decomposition–enable measurements as a starting point, we used the energy itself as a starting point to define a distance? This is the approach we take in the following section.
Energy-centric approach
Recall that we are after a way to measure how different two states are. Moreover we wish to do so in a manner which is compatible with both the balanced-unbalanced decomposition underpinning Theorem 3.5 and the conserved energy. In the previous section we approached this problem with the first requirement in mind, and eventually discovered that this approach failed to (generically) satisfy the second requirement. In this section we proceed the other way around: we enforce the second requirement by putting the conserved energy front and center in the construction of a new distance. The question will then be to identify whether or not this metric agrees with the decomposition.
Also, recall the difficulty that arises in trying to define a distance or metric or inner product that agrees with the energy in (32) : the nonlinear switch from the minimum function or Heaviside function. This nonlinear switch is activated based on the total water q q q , and it is not clear how to define the nonlinear switch when two different total water variables, q 1 q_1 q 1 and q 2 q_2 q 2 , are involved in quantifying the distance between two different states. This difficulty and others are discussed in more detail in the text preceding § 5.1 .
A key idea is helpful for overcoming the difficulties of the nonlinear switch: in defining a metric, begin by working pointwise in space . Then, depending on which phase that point belongs to, the conserved energy (32) determines a natural inner product to use.
More precisely, this section proceeds as follows. We begin by defining a metric pointwise which measures the discrepancy between two states X 1 ( x ) \mathcal{X}_1(x) X 1 ( x ) and X 2 ( x ) \mathcal{X}_2(x) X 2 ( x ) when evaluated at a specific point x x x of the domain T 3 \mathbb{T}^3 T 3 . This pointwise metric is defined to agree with the conserved energy density and so it is defined piecewise, taking different forms depending on the sign of q ( x ) q(x) q ( x ) . We then record the properties of this pointwise metric: it is well-defined, a metric, and agrees with the energy unconditionally. We also record how to compute this distance explicitly in a special case. We then make the key observation of this section: this pointwise metric does not agree with the extraction of a balanced component, i.e. it does not agree with the decomposition. Finally we conclude this section by defining the induced metric on the state space L σ 2 L^2_\sigma L σ 2 obtained by integrating the pointwise metric over the domain and explain why we do not expect this metric to agree with the decomposition either.
As mentioned above, a key idea in the definition of this metric is to work pointwise in space . In particular, depending on which phase that point belongs to, the conserved energy (32) determines a natural inner product to use. We introduce appropriate notation with this in mind.
Definition 5.12. For any S 1 = ( u 1 , θ 1 , q 1 ) ∈ R 5 S_1=(u_1,\theta_1,q_1)\in\mathbb{R}^5 S 1 = ( u 1 , θ 1 , q 1 ) ∈ R 5 and S 2 = ( u 2 , θ 2 , q 2 ) ∈ R 5 S_2=(u_2,\theta_2,q_2)\in\mathbb{R}^5 S 2 = ( u 2 , θ 2 , q 2 ) ∈ R 5 , where u 1 , u 2 ∈ R 3 u_1,u_2\in\mathbb{R}^3 u 1 , u 2 ∈ R 3 , we define the inner products
⟨ S 1 , S 2 ⟩ u : = u 1 ⋅ u 2 + θ 1 θ 2 + q 1 q 2 + ( θ 1 + q 1 ) ( θ 2 + q 2 ) \langle S_1,S_2\rangle_u := u_1\cdot u_2+\theta_1\theta_2+q_1q_2+(\theta_1+q_1)(\theta_2+q_2) ⟨ S 1 , S 2 ⟩ u := u 1 ⋅ u 2 + θ 1 θ 2 + q 1 q 2 + ( θ 1 + q 1 ) ( θ 2 + q 2 ) and
⟨ S 1 , S 2 ⟩ s : = u 1 ⋅ u 2 + θ 1 θ 2 + ( θ 1 + q 1 ) ( θ 2 + q 2 ) . \langle S_1,S_2\rangle_s := u_1\cdot u_2+\theta_1\theta_2+(\theta_1+q_1)(\theta_2+q_2). ⟨ S 1 , S 2 ⟩ s := u 1 ⋅ u 2 + θ 1 θ 2 + ( θ 1 + q 1 ) ( θ 2 + q 2 ) . We denote by ∥ ⋅ ∥ u \|\cdot\|_u ∥ ⋅ ∥ u and ∥ ⋅ ∥ s \|\cdot\|_s ∥ ⋅ ∥ s the corresponding norms.
The inner products introduced in Definition 5.12 are chosen precisely so that the ensuing norms ∥ ⋅ ∥ u \|\cdot\|_u ∥ ⋅ ∥ u and ∥ ⋅ ∥ s \|\cdot\|_s ∥ ⋅ ∥ s agree with the conserved energy density. Indeed, when q < 0 q<0 q < 0 we have that
∥ ( u , θ , q ) ∥ u 2 = ∣ u ∣ 2 + θ 2 + q 2 + ( θ + q ) 2 \|(u,\theta,q)\|_u^2=|u|^2+\theta^2+q^2+(\theta+q)^2 ∥ ( u , θ , q ) ∥ u 2 = ∣ u ∣ 2 + θ 2 + q 2 + ( θ + q ) 2
and when q ≥ 0 q\geq0 q ≥ 0 we have that
∥ ( u , θ , q ) ∥ s 2 = ∣ u ∣ 2 + θ 2 + ( θ + q ) 2 , \|(u,\theta,q)\|_s^2=|u|^2+\theta^2+(\theta+q)^2, ∥ ( u , θ , q ) ∥ s 2 = ∣ u ∣ 2 + θ 2 + ( θ + q ) 2 ,
such that indeed they agree with the conserved energy density as recorded in (32) .
Next we wish to define the pointwise metric. As alluded to at the start of this section, this metric will be defined piecewise (depending on the sign of q q q ) and so we introduce notation which describes each of these pieces.
Definition 5.13. We consider the following subsets of R 5 \mathbb{R}^5 R 5 .
U : = { ( u , θ , q ) ∈ R 5 : q < 0 } \mathcal{U}:=\{(u,\theta,q)\in\mathbb{R}^5:q<0\} U := {( u , θ , q ) ∈ R 5 : q < 0 } is the set of unsaturated states,
S : = { ( u , θ , q ) ∈ R 5 : q ≥ 0 } \mathcal{S}:=\{(u,\theta,q)\in\mathbb{R}^5:q\geq0\} S := {( u , θ , q ) ∈ R 5 : q ≥ 0 } is the set of saturated states, and
I : = { ( u , θ , q ) ∈ R 5 : q = 0 } \mathcal{I}:=\{(u,\theta,q)\in\mathbb{R}^5:q=0\} I := {( u , θ , q ) ∈ R 5 : q = 0 } is the state-space interface between unsaturated states and saturated states.
We have the necessary notation in hand to define the following Snell-type metric. We refer to this metric as being of Snell-type since shortest paths between states in different phases exhibit the piecewise affine behaviour observed by light passing across a boundary (see also Figure 2 ).
Figure 2. As introduced in Definition 5.14 we see that here that the “Snell-type” metric warrants its name. Shown here is the geodesic between two points in the ( θ , q ) (\theta,q) ( θ , q ) plane which live on different sides of the state-space interface { q = 0 } \{q=0\} { q = 0 } . As is familiar from Snell’s refraction law of optics, the geodesic intersects the interface at an angle. Note that we may use Lemma 5.17 in order to compute the intersection point and the value of θ ∗ \theta_\ast θ ∗ exactly.
Figure 1. The vectors { M ^ s , b ^ s } \{\widehat{M}_s,\widehat{b}_s\} { M s , b s } and { M ^ u , b ^ u } \{\widehat{M}_u,\widehat{b}_u\} { M u , b u } form orthonormal bases with respect to the inner products ⟨ ⋅ , ⋅ ⟩ s \langle\cdot,\cdot\rangle_s ⟨ ⋅ , ⋅ ⟩ s and ⟨ ⋅ , ⋅ ⟩ u \langle\cdot,\cdot\rangle_u ⟨ ⋅ , ⋅ ⟩ u , respectively, as they were introduced in Definition 5.12 , namely M ^ s = ( 0 , 1 ) \widehat{M}_s=(0,1) M s = ( 0 , 1 ) , b ^ s = ( 1 , − 1 ) \widehat{b}_s=(1,-1) b s = ( 1 , − 1 ) , M ^ u = 1 2 ( 1 , 1 ) \widehat{M}_u=\frac{1}{\sqrt{2}}(1,1) M u = 2 1 ( 1 , 1 ) , and b ^ u = 1 2 ( 1 , − 1 ) \widehat{b}_u=\frac{1}{\sqrt{2}}(1,-1) b u = 2 1 ( 1 , − 1 ) . Even though, especially in the saturated region { q > 0 } \{q>0\} { q > 0 } , labelling these vectors with the moist variable M M M and the buoyancy b b b may at first appear perplexing, this is justified by the fact that their corresponding dual vectors are precisely (multiples of) M M M and b b b . Indeed: ⟨ M ^ s , ( θ , q ) ⟩ s = θ + q = M \langle\widehat{M}_s,(\theta,q)\rangle_s=\theta+q=M ⟨ M s , ( θ , q ) ⟩ s = θ + q = M , ⟨ b ^ s , ( θ , q ) ⟩ s = θ = b s \langle\widehat{b}_s,(\theta,q)\rangle_s=\theta=b_s ⟨ b s , ( θ , q ) ⟩ s = θ = b s , ⟨ M ^ u , ( θ , q ) ⟩ u = 3 2 ( θ + q ) = 3 2 M \langle\widehat{M}_u,(\theta,q)\rangle_u=\sqrt{\frac{3}{2}}(\theta+q)=\sqrt{\frac{3}{2}}M ⟨ M u , ( θ , q ) ⟩ u = 2 3 ( θ + q ) = 2 3 M , and ⟨ b ^ u , ( θ , q ) ⟩ u = 1 2 θ − q = 1 2 b u \langle\widehat{b}_u,(\theta,q)\rangle_u=\frac{1}{\sqrt{2}}\theta-q=\frac{1}{\sqrt{2}}b_u ⟨ b u , ( θ , q ) ⟩ u = 2 1 θ − q = 2 1 b u .
Definition 5.14 (Snell-type metric, pointwise). We define d S n e l l : R 5 → R 5 → [ 0 , ∞ ) d_{\mathrm{Snell}} : \mathbb{R}^{5} \to\mathbb{R}^{5} \to[0,\infty) d Snell : R 5 → R 5 → [ 0 , ∞ ) as follows.
If S 1 , S 2 ∈ U S_1, S_2 \in\mathcal{U} S 1 , S 2 ∈ U then d S n e l l ( S 1 , S 2 ) : = ∥ S 1 − S 2 ∥ U d_{\mathrm{Snell}}(S_1,S_2) := \lVert S_1-S_2\rVert_{\mathcal{U}} d Snell ( S 1 , S 2 ) := ∥ S 1 − S 2 ∥ U .
If S 1 , S 2 ∈ S S_1, S_2 \in\mathcal{S} S 1 , S 2 ∈ S then d S n e l l ( S 1 , S 2 ) : = ∥ S 1 − S 2 ∥ S d_{\mathrm{Snell}}(S_1,S_2) := \lVert S_1-S_2\rVert_{\mathcal{S}} d Snell ( S 1 , S 2 ) := ∥ S 1 − S 2 ∥ S .
If S 1 ∈ S S_1 \in\mathcal{S} S 1 ∈ S and S 2 ∈ U S_2 \in\mathcal{U} S 2 ∈ U then
d S n e l l : = min I ∈ I ∥ S 1 − I ∥ S + ∥ S 2 − I ∥ U . d_{\mathrm{Snell}} := \min_{I\in\mathcal{I}} \lVert S_1-I\rVert_{\mathcal{S}}+\lVert S_2-I\rVert_{\mathcal{U}}. d Snell := I ∈ I min ∥ S 1 − I ∥ S + ∥ S 2 − I ∥ U . If S 1 ∈ U S_1 \in\mathcal{U} S 1 ∈ U and S 2 ∈ S S_2 \in\mathcal{S} S 2 ∈ S then d S n e l l ( S 1 , S 2 ) : = d S n e l l ( S 2 , S 1 ) d_{\mathrm{Snell}}(S_1,S_2) := d_{\mathrm{Snell}}(S_2,S_1) d Snell ( S 1 , S 2 ) := d Snell ( S 2 , S 1 ) .
First we verify that this distance function is well-defined.
Lemma 5.15 (The Snell-type distance is well-defined). For any S 1 ∈ S S_1 \in\mathcal{S} S 1 ∈ S and any S 2 ∈ U S_2 \in\mathcal{U} S 2 ∈ U there is a unique minimizer of
min I ∈ I ∥ S 1 − I ∥ S + ∥ S 2 − I ∥ U . \min_{I\in\mathcal{I}} \lVert S_1-I\rVert_{\mathcal{S}}+\lVert S_2-I\rVert_{\mathcal{U}}. I ∈ I min ∥ S 1 − I ∥ S + ∥ S 2 − I ∥ U .
Proof. To verify that d S n e l l d_{\mathrm{Snell}} d Snell is well-defined it suffices to verify that, for any S 1 ∈ S S_1 \in\mathcal{S} S 1 ∈ S and any S 2 ∈ U S_2 \in\mathcal{U} S 2 ∈ U , there exists a unique minimizer of
min I ∈ I ∥ S 1 − I ∥ S + ∥ S 2 − I ∥ U . (42) \min_{I\in\mathcal{I}} \lVert S_1-I\rVert_{\mathcal{S}}+\lVert S_2-I\rVert_{\mathcal{U}}. \tag*{(42)} I ∈ I min ∥ S 1 − I ∥ S + ∥ S 2 − I ∥ U . (42)
To do so, fix S ∈ R 5 ∖ I S \in\mathbb{R}^{5}\setminus\mathcal{I} S ∈ R 5 ∖ I and consider f : R 4 → R 5 f:\mathbb{R}^{4}\to\mathbb{R}^{5} f : R 4 → R 5 defined by, for every I = ( u h , w , θ , 0 ) ∈ I I=(u_h,w,\theta,0)\in\mathcal{I} I = ( u h , w , θ , 0 ) ∈ I ,
f ( u h , w , θ ) = ∥ S − I ∥ f(u_h,w,\theta)=\lVert S-I\rVert f ( u h , w , θ ) = ∥ S − I ∥
where ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ denotes ∥ ⋅ ∥ S \lVert\cdot\rVert_{\mathcal{S}} ∥ ⋅ ∥ S or ∥ ⋅ ∥ U \lVert\cdot\rVert_{\mathcal{U}} ∥ ⋅ ∥ U . To verify that (42) has a unique minimizer it suffices to show that f f f is strictly convex. Since f f f is smooth (as S S S is not in I \mathcal{I} I , and f f f and its derivatives would blow up precisely when S = I S=I S = I , but I ∈ I I\in\mathcal{I} I ∈ I ) it suffices to verify that
∇ 2 f > 0. \nabla^{2}f>0. ∇ 2 f > 0.
So we compute:
∂ i f = ∂ i ∥ S − I ∥ = ⟨ I − S , ∂ i I ⟩ ∥ S − I ∥ \partial_i f=\partial_i\lVert S-I\rVert=\frac{\langle I-S,\partial_i I\rangle}{\lVert S-I\rVert} ∂ i f = ∂ i ∥ S − I ∥ = ∥ S − I ∥ ⟨ I − S , ∂ i I ⟩
and hence
∂ i ∂ j f = ∂ j ( ⟨ I − S , ∂ i I ⟩ ∥ S − I ∥ ) = ⟨ ∂ i I , ∂ j I ⟩ ∥ S − I ∥ 2 − ⟨ ∂ i I , S − I ⟩ ⟨ ∂ j I , S − I ⟩ ∥ S − I ∥ 3 . \partial_i\partial_j f=\partial_j\left(\frac{\langle I-S,\partial_i I\rangle}{\lVert S-I\rVert}\right)=\frac{\langle\partial_i I,\partial_j I\rangle\lVert S-I\rVert^2-\langle\partial_i I,S-I\rangle\langle\partial_j I,S-I\rangle}{\lVert S-I\rVert^3}. ∂ i ∂ j f = ∂ j ( ∥ S − I ∥ ⟨ I − S , ∂ i I ⟩ ) = ∥ S − I ∥ 3 ⟨ ∂ i I , ∂ j I ⟩ ∥ S − I ∥ 2 − ⟨ ∂ i I , S − I ⟩ ⟨ ∂ j I , S − I ⟩ .
To make the remaining computation easier to follow, let us write
x i : = ∂ i I and z : = S − I . x^i:=\partial_i I\quad\text{and}\quad z:=S-I. x i := ∂ i I and z := S − I .
Then
∥ z ∥ 3 ∇ 2 f = ⟨ x i , x j ⟩ ∥ z ∥ 2 − ⟨ x i , z ⟩ ⟨ x j , z ⟩ = : A \lVert z\rVert^3\nabla^2 f=\langle x^i,x^j\rangle\lVert z\rVert^2-\langle x^i,z\rangle\langle x^j,z\rangle=:A ∥ z ∥ 3 ∇ 2 f = ⟨ x i , x j ⟩ ∥ z ∥ 2 − ⟨ x i , z ⟩ ⟨ x j , z ⟩ =: A
and so it suffices to show that A > 0 A>0 A > 0 (since ∥ z ∥ ≠ 0 \lVert z\rVert\ne0 ∥ z ∥ = 0 as long as S ∉ I S\notin\mathcal{I} S ∈ / I ). To simplify further, let us introduce
D : = diag ( ( ∥ x i ∥ ) 4 ) i = 1 4 . D:=\operatorname{diag}\left(\left(\lVert x^i\rVert\right)^4\right)_{i=1}^{4}. D := diag ( ( ∥ x i ∥ ) 4 ) i = 1 4 .
Then
A > 0 ⟺ D T A D > 0 A > 0 \Longleftrightarrow D^{T}AD > 0 A > 0 ⟺ D T A D > 0
where, since ( x i ) i = 1 4 (x^{i})_{i=1}^{4} ( x i ) i = 1 4 forms an orthogonal set with respect to both ∥ ⋅ ∥ s \lVert\cdot\rVert_{s} ∥ ⋅ ∥ s and ∥ ⋅ ∥ u \lVert\cdot\rVert_{u} ∥ ⋅ ∥ u , we have that D T A D = I ∥ z ∥ 2 − B D^{T}AD = I\lVert z\rVert^{2}-B D T A D = I ∥ z ∥ 2 − B for
B i j : = ⟨ x ^ i , z ⟩ ⟨ x ^ j , z ⟩ . B_{ij}:=\langle\hat{x}^{i},z\rangle\langle\hat{x}^{j},z\rangle. B ij := ⟨ x ^ i , z ⟩ ⟨ x ^ j , z ⟩ .
To conclude the proof it thus suffices to prove that B < ∥ z ∥ 2 I B < \lVert z\rVert^{2}I B < ∥ z ∥ 2 I . First we show that B ≤ ∥ z ∥ 2 I B \leq\lVert z\rVert^{2}I B ≤ ∥ z ∥ 2 I . Well, for any y ∈ R 4 y\in\mathbb{R}^{4} y ∈ R 4 ,
B y ⋅ y = y i ⟨ x ^ i , z ⟩ y j ⟨ x j , x ⟩ = y ⋅ ( ⟨ x ^ i , z ⟩ ) i = 1 4 ≤ ∣ y ∣ 2 ( ∑ i = 1 4 ⟨ x ^ i , z ⟩ 2 ) . By\cdot y=y_{i}\langle\hat{x}^{i},z\rangle y_{j}\langle x^{j},x\rangle=y\cdot\left(\langle\hat{x}^{i},z\rangle\right)_{i=1}^{4}\leq|y|^{2}\left(\sum_{i=1}^{4}\langle\hat{x}^{i},z\rangle^{2}\right). B y ⋅ y = y i ⟨ x ^ i , z ⟩ y j ⟨ x j , x ⟩ = y ⋅ ( ⟨ x ^ i , z ⟩ ) i = 1 4 ≤ ∣ y ∣ 2 ( i = 1 ∑ 4 ⟨ x ^ i , z ⟩ 2 ) .
Since ( x i ) i = 1 4 (x^{i})_{i=1}^{4} ( x i ) i = 1 4 is an orthonormal set, we have that
∑ i = 1 4 ⟨ x ^ i , z ⟩ 2 ≤ ∥ z ∥ 2 , \sum_{i=1}^{4}\langle\hat{x}^{i},z\rangle^{2}\leq\lVert z\rVert^{2}, i = 1 ∑ 4 ⟨ x ^ i , z ⟩ 2 ≤ ∥ z ∥ 2 ,
which verifies that B ≤ ∥ z ∥ 2 I B\leq\lVert z\rVert^{2}I B ≤ ∥ z ∥ 2 I . Finally we show that B < ∥ z ∥ 2 I B<\lVert z\rVert^{2}I B < ∥ z ∥ 2 I . This follows from the fact that ( x ^ i ) i = 1 4 (\hat{x}^{i})_{i=1}^{4} ( x ^ i ) i = 1 4 span I \mathcal{I} I , but not R 5 \mathbb{R}^{5} R 5 , and so
∑ i = 1 4 ⟨ x ^ i , z ⟩ 2 = ∥ proj I z ∥ 2 < ∥ z ∥ 2 \sum_{i=1}^{4}\langle\hat{x}^{i},z\rangle^{2}=\lVert\operatorname{proj}_{\mathcal{I}}z\rVert^{2}<\lVert z\rVert^{2} i = 1 ∑ 4 ⟨ x ^ i , z ⟩ 2 = ∥ proj I z ∥ 2 < ∥ z ∥ 2
since z = S − I ∉ I z=S-I\notin\mathcal{I} z = S − I ∈ / I since I ∈ I I\in\mathcal{I} I ∈ I but S ∉ I S\notin\mathcal{I} S ∈ / I . In other words: since S ∉ I S\notin\mathcal{I} S ∈ / I we have that B < ∥ z ∥ 2 I B<\lVert z\rVert^{2}I B < ∥ z ∥ 2 I , and so ∇ 2 f > 0 \nabla^{2}f>0 ∇ 2 f > 0 .
Now we verify that the distance function introduced in Definition 5.14 is indeed a metric. We will also show that this Snell-type distance also agrees with the energy, when they are comparable.
Lemma 5.16 (The Snell-type distance is a metric). As introduced in Definition 5.14 , d S n e l l d_{\mathrm{Snell}} d Snell is a metric on R 5 \mathbb{R}^{5} R 5 . Moreover it agrees with the energy density in the sense that, for any S 1 , S 2 ∈ R 5 S_{1},S_{2}\in\mathbb{R}^{5} S 1 , S 2 ∈ R 5 , if the signs of q 1 q_{1} q 1 and q 2 q_{2} q 2 agree (i.e. if the states S 1 S_{1} S 1 and S 2 S_{2} S 2 are in the same phase) then
d S n e l l ( S 1 , S 2 ) 2 = ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 h + ( ( θ 1 + q 1 ) − ( θ 2 − q 2 ) ) 2 d_{\mathrm{Snell}}(S_{1},S_{2})^{2}=|u_{1}-u_{2}|^{2}+(\theta_{1}-\theta_{2})^{2}+(q_{1}-q_{2})^{2}h+\big((\theta_{1}+q_{1})-(\theta_{2}-q_{2})\big)^{2} d Snell ( S 1 , S 2 ) 2 = ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 h + ( ( θ 1 + q 1 ) − ( θ 2 − q 2 ) ) 2 where
h = { 1 if q 1 , q 2 < 0 and 0 if q 1 , q 2 ≥ 0. h=\begin{cases}
1 & \text{if }q_{1},q_{2}<0\text{ and}\\
0 & \text{if }q_{1},q_{2}\geq0.
\end{cases} h = { 1 0 if q 1 , q 2 < 0 and if q 1 , q 2 ≥ 0. Proof. Positive-definiteness and symmetry are immediate so we only need to prove that the triangle inequality holds. Indeed: the agreement with the energy follows immediately from the definition of ∥ ⋅ ∥ u \lVert\cdot\rVert_{u} ∥ ⋅ ∥ u and ∥ ⋅ ∥ s \lVert\cdot\rVert_{s} ∥ ⋅ ∥ s in Definition 5.12 .
For S 1 , S 2 , S 3 ∈ R 5 S_{1},S_{2},S_{3}\in\mathbb{R}^{5} S 1 , S 2 , S 3 ∈ R 5 we therefore wish to show that
d S n e l l ( S 1 , S 2 ) ≤ d S n e l l ( S 1 , S 3 ) + d S n e l l ( S 3 , S 2 ) . d_{\mathrm{Snell}}(S_{1},S_{2})\leq d_{\mathrm{Snell}}(S_{1},S_{3})+d_{\mathrm{Snell}}(S_{3},S_{2}). d Snell ( S 1 , S 2 ) ≤ d Snell ( S 1 , S 3 ) + d Snell ( S 3 , S 2 ) . Clearly the triangle inequality holds when all three states belong to the same phase (either U \mathcal{U} U or S \mathcal{S} S ). So we need only consider the case where two of the states S 1 S_{1} S 1 , S 2 S_{2} S 2 , and S 3 S_{3} S 3 are in one phase and the last state is in the other. Since all that matters is whether or not the intermediate point S 3 S_{3} S 3 is alone in its phase, there are without loss of generality two cases to consider.
Case 1: S 1 , S 2 ∈ S S_{1},S_{2}\in\mathcal{S} S 1 , S 2 ∈ S and S 3 ∈ U S_{3}\in\mathcal{U} S 3 ∈ U . Let us denote by I 13 I_{13} I 13 the point intersecting the geodesic between S 1 S_{1} S 1 and S 3 S_{3} S 3 and the state-space interface I \mathcal{I} I , and similarly I 23 I_{23} I 23 for the point arising from the geodesic between S 2 S_{2} S 2 and S 3 S_{3} S 3 . We use the triangle inequality in U \mathcal{U} U , then the triangle inequality in S \mathcal{S} S :
d S n e l l ( S 1 , S 3 ) + d S n e l l ( S 3 , S 2 ) = d S n e l l ( S 1 , I 13 ) + d S n e l l ( I 13 , S 3 ) + d S n e l l ( S 3 , I 23 ) + d S n e l l ( I 23 , S 2 ) ≥ d S n e l l ( S 1 , I 13 ) + d S n e l l ( I 13 , I 23 ) + d S n e l l ( I 23 , S 2 ) ≥ d S n e l l ( S 1 , S 2 ) . \begin{aligned}
d_{\mathrm{Snell}}(S_{1},S_{3})+d_{\mathrm{Snell}}(S_{3},S_{2})
&=d_{\mathrm{Snell}}(S_{1},I_{13})+d_{\mathrm{Snell}}(I_{13},S_{3})+d_{\mathrm{Snell}}(S_{3},I_{23})+d_{\mathrm{Snell}}(I_{23},S_{2})\\
&\geq d_{\mathrm{Snell}}(S_{1},I_{13})+d_{\mathrm{Snell}}(I_{13},I_{23})+d_{\mathrm{Snell}}(I_{23},S_{2})\\
&\geq d_{\mathrm{Snell}}(S_{1},S_{2}).
\end{aligned} d Snell ( S 1 , S 3 ) + d Snell ( S 3 , S 2 ) = d Snell ( S 1 , I 13 ) + d Snell ( I 13 , S 3 ) + d Snell ( S 3 , I 23 ) + d Snell ( I 23 , S 2 ) ≥ d Snell ( S 1 , I 13 ) + d Snell ( I 13 , I 23 ) + d Snell ( I 23 , S 2 ) ≥ d Snell ( S 1 , S 2 ) .
Figure 3. A pictorial depiction of the two steps used in case 1 of the proof of Lemma 5.16 .
Figure 4. A pictorial depiction of the two steps used in case 2 of the proof of Lemma 5.16 .
Case 2: S 1 , S 3 ∈ S S_1, S_3 \in\mathcal{S} S 1 , S 3 ∈ S and S 2 ∈ U S_2 \in\mathcal{U} S 2 ∈ U . Let I 12 I_{12} I 12 and I 23 I_{23} I 23 be defined similarly to in case 1. We use the triangle inequality in S \mathcal{S} S and the definition of d S n e l l d_{\mathrm{Snell}} d Snell as a minimization problem when crossing the interface to deduce that
d S n e l l ( S 1 , S 3 ) + d S n e l l ( S 3 , S 2 ) = d S n e l l ( S 1 , S 3 ) + d S n e l l ( S 3 , I 23 ) + d S n e l l ( I 23 , S 2 ) ≥ d S n e l l ( S 1 , I 23 ) + d S n e l l ( I 23 , S 2 ) ≥ d S n e l l ( S 1 , S 2 ) . \begin{aligned}
d_{\mathrm{Snell}}(S_1,S_3)+d_{\mathrm{Snell}}(S_3,S_2)
&=d_{\mathrm{Snell}}(S_1,S_3)+d_{\mathrm{Snell}}(S_3,I_{23})+d_{\mathrm{Snell}}(I_{23},S_2)\\
&\ge d_{\mathrm{Snell}}(S_1,I_{23})+d_{\mathrm{Snell}}(I_{23},S_2)\\
&\ge d_{\mathrm{Snell}}(S_1,S_2).
\end{aligned} d Snell ( S 1 , S 3 ) + d Snell ( S 3 , S 2 ) = d Snell ( S 1 , S 3 ) + d Snell ( S 3 , I 23 ) + d Snell ( I 23 , S 2 ) ≥ d Snell ( S 1 , I 23 ) + d Snell ( I 23 , S 2 ) ≥ d Snell ( S 1 , S 2 ) .
We now seek to determine whether or not this metric, once ported over to measure the distance between spatially-varying states, agrees with the balanced-unbalanced decomposition of Theorem 3.5 . To do so we will need to understand this Snell-type metric in the pointwise case (which we have discussed so far) in more detail. In particular we will rely on the explicit computation of geodesic trajectories in the case where u 1 = u 2 = 0 u_1=u_2=0 u 1 = u 2 = 0 . In that case, the states live in the ( θ , q ) (\theta,q) ( θ , q ) –plane. We consider two states in that plane living on either side of the state-space interface I \mathcal{I} I , i.e. where one state is saturated and the other is not. The geodesic between these two points is then piecewise affine, changing direction when it reaches the interface. Crucially: in this simplified setting where u 1 = u 2 = 0 u_1=u_2=0 u 1 = u 2 = 0 we can compute explicitly where the geodesic intersects the interface.
Lemma 5.17 (Explicit computation of the intersection point). For any S 1 ∈ S S_1 \in\mathcal{S} S 1 ∈ S and any S 2 ∈ U S_2 \in\mathcal{U} S 2 ∈ U , if u 1 = u 2 = 0 u_1=u_2=0 u 1 = u 2 = 0 then, for I ( θ ) : = ( 0 , 0 , 0 , θ , 0 ) ∈ I ⊂ R 5 I(\theta):=(0,0,0,\theta,0)\in\mathcal{I}\subset\mathbb{R}^5 I ( θ ) := ( 0 , 0 , 0 , θ , 0 ) ∈ I ⊂ R 5 ,
arg min θ ∈ R ∥ S 1 − I ( θ ) ∥ s + ∥ S 2 − I ( θ ) ∥ u ∈ { α − δ β , α + δ β } \underset{\theta\in\mathbb{R}}{\arg\min}\ \lVert S_1-I(\theta)\rVert_s+\lVert S_2-I(\theta)\rVert_u \in\left\{\frac{\alpha-\sqrt{\delta}}{\beta},\frac{\alpha+\sqrt{\delta}}{\beta}\right\} θ ∈ R arg min ∥ S 1 − I ( θ ) ∥ s + ∥ S 2 − I ( θ ) ∥ u ∈ { β α − δ , β α + δ } for
α = 4 q 1 2 ( 2 θ 2 + q 2 ) − 6 q 2 2 ( 2 θ 1 + q 1 ) , \alpha=4q_1^2(2\theta_2+q_2)-6q_2^2(2\theta_1+q_1), α = 4 q 1 2 ( 2 θ 2 + q 2 ) − 6 q 2 2 ( 2 θ 1 + q 1 ) , β = 8 q 1 2 − 24 q 2 2 , and \beta=8q_1^2-24q_2^2,\quad\text{and} β = 8 q 1 2 − 24 q 2 2 , and δ = 12 q 1 2 q 2 2 ( 2 θ 1 + q 1 + 4 θ 2 + 2 q 2 ) 2 . \delta=12q_1^2q_2^2(2\theta_1+q_1+4\theta_2+2q_2)^2. δ = 12 q 1 2 q 2 2 ( 2 θ 1 + q 1 + 4 θ 2 + 2 q 2 ) 2 .
Figure 5. This shows an example where the two possible intersection points are computed from Lemma 5.17. In practice, it is very straightforward to identify the correct intersection point: the lengths of both paths may readily be computed, and the intersection point corresponding to the shortest path is then the correct one.
Proof. The key idea is that the first derivative condition for minimizers may be squared to produce a fourth-order polynomial equation. Solving this equation using symbolic computation software then produces the claim. More precisely, let us define f : R → R f:\mathbb{R}\to\mathbb{R} f : R → R via
f ( θ ) : = ∥ S 1 − I ( θ ) ∥ s + ∥ S 2 − I ( θ ) ∥ u . f(\theta):=\lVert S_1-I(\theta)\rVert_s+\lVert S_2-I(\theta)\rVert_u. f ( θ ) := ∥ S 1 − I ( θ ) ∥ s + ∥ S 2 − I ( θ ) ∥ u .
Then, proceeding as in the proof of Lemma 5.15 we see that
f ′ ( θ ) = ⟨ I − S 1 , ∂ θ I ⟩ s ∥ S 1 − I ∥ s + ⟨ I − S 2 , ∂ θ I ⟩ u ∥ S 2 − I ∥ u . f'(\theta)=\frac{\langle I-S_1,\partial_\theta I\rangle_s}{\lVert S_1-I\rVert_s}+\frac{\langle I-S_2,\partial_\theta I\rangle_u}{\lVert S_2-I\rVert_u}. f ′ ( θ ) = ∥ S 1 − I ∥ s ⟨ I − S 1 , ∂ θ I ⟩ s + ∥ S 2 − I ∥ u ⟨ I − S 2 , ∂ θ I ⟩ u .
and so, since ∂ θ I = e θ \partial_\theta I=e_\theta ∂ θ I = e θ , we see that if f ′ ( θ ) = 0 f'(\theta)=0 f ′ ( θ ) = 0 then
⟨ I − S 2 , e θ ⟩ u 2 ∥ S 1 − I ∥ s 2 = ⟨ I − S 1 , e θ ⟩ s 2 ∥ S 2 − I ∥ u 2 . (43) \langle I-S_2,e_\theta\rangle_u^2\lVert S_1-I\rVert_s^2=\langle I-S_1,e_\theta\rangle_s^2\lVert S_2-I\rVert_u^2. \tag*{(43)} ⟨ I − S 2 , e θ ⟩ u 2 ∥ S 1 − I ∥ s 2 = ⟨ I − S 1 , e θ ⟩ s 2 ∥ S 2 − I ∥ u 2 . (43)
Since
⟨ I − S i , e θ ⟩ u = ⟨ I − S i , e θ ⟩ s = 2 θ − 2 θ i − q i \langle I-S_i,e_\theta\rangle_u=\langle I-S_i,e_\theta\rangle_s=2\theta-2\theta_i-q_i ⟨ I − S i , e θ ⟩ u = ⟨ I − S i , e θ ⟩ s = 2 θ − 2 θ i − q i
while
∥ S 1 − I ∥ s 2 = ( θ − θ 1 ) 2 + ( θ − θ 1 − q 1 ) 2 \lVert S_1-I\rVert_s^2=(\theta-\theta_1)^2+(\theta-\theta_1-q_1)^2 ∥ S 1 − I ∥ s 2 = ( θ − θ 1 ) 2 + ( θ − θ 1 − q 1 ) 2
and
∥ S 2 − I ∥ u 2 = 1 2 ( θ − θ 2 + q 2 ) 2 + 3 2 ( θ − θ 2 − q 2 ) 2 , \lVert S_2-I\rVert_u^2=\frac{1}{2}(\theta-\theta_2+q_2)^2+\frac{3}{2}(\theta-\theta_2-q_2)^2, ∥ S 2 − I ∥ u 2 = 2 1 ( θ − θ 2 + q 2 ) 2 + 2 3 ( θ − θ 2 − q 2 ) 2 ,
where we have used the identity
θ 2 + q 2 h = ( 1 − h / 2 ) ( θ − q h ) 2 + ( h / 2 ) ( θ + q ) 2 , \theta^2+q^2h=(1-h/2)(\theta-qh)^2+(h/2)(\theta+q)^2, θ 2 + q 2 h = ( 1 − h /2 ) ( θ − q h ) 2 + ( h /2 ) ( θ + q ) 2 ,
the equation (43) comes down to a polynomial equation of order four for θ \theta θ . We may solve this equation explicitly using symbolic computation software, thus producing the claim.
We now discuss how to extract balanced components in this pointwise setting where the state space is R 5 \mathbb{R}^5 R 5 (and not L σ 2 L_\sigma^2 L σ 2 ). The system of interest is
{ d u h d t + 1 ε u h = 0 , d w d t + 1 ε ( θ − min 0 q ) = 0 , d θ d t + 1 ε w = 0 , and d q d t − 1 ε w = 0 , (44) \left\{
\begin{aligned}
\frac{d u_h}{d t}+\frac{1}{\varepsilon}u_h&=0,\\
\frac{d w}{d t}+\frac{1}{\varepsilon}(\theta-\min_0 q)&=0,\\
\frac{d\theta}{d t}+\frac{1}{\varepsilon}w&=0,\quad\text{and}\\
\frac{d q}{d t}-\frac{1}{\varepsilon}w&=0,
\end{aligned}
\right. \tag*{(44)} ⎩ ⎨ ⎧ d t d u h + ε 1 u h d t d w + ε 1 ( θ − 0 min q ) d t d θ + ε 1 w d t d q − ε 1 w = 0 , = 0 , = 0 , and = 0 , (44)
which fits into our generic framework as it can be written in the form, for S = ( u h , w , θ , q ) S=(u_h,w,\theta,q) S = ( u h , w , θ , q ) , as
d S d t + 1 ε N ( S ) = 0 for N ( S ) = ( u h θ − min 0 q w − w ) . \frac{dS}{dt}+\frac{1}{\varepsilon}N(S)=0\quad\text{for}\quad N(S)=
\begin{pmatrix}
u_h\\
\theta-\min_0 q\\
w\\
-w
\end{pmatrix}. d t d S + ε 1 N ( S ) = 0 for N ( S ) = u h θ − min 0 q w − w .
As detailed in § 3 , the first step in identifying an appropriate decomposition for this system is to study the kernel / zero set of N N N and the image of N N N . We see that the kernel of N N N may be characterized (akin to Proposition 3.1 ) as a zero set of appropriate measurements or may be explicitly parameterised:
N ( S ) = 0 ⟺ S ∈ ker N = : B N(S)=0 \Longleftrightarrow S\in\ker N=:B N ( S ) = 0 ⟺ S ∈ ker N =: B
⟺ u h = 0 , w = 0 , and b ( S ) : = θ − min 0 q = 0 \Longleftrightarrow u_h=0,\ w=0,\ \text{and }b(S):=\theta-\min_0 q=0 ⟺ u h = 0 , w = 0 , and b ( S ) := θ − 0 min q = 0
⟺ S = ( 0 0 1 2 min 0 M M − 1 2 min 0 M ) \Longleftrightarrow S=
\begin{pmatrix}
0\\
0\\
\frac{1}{2}\min_0 M\\
M-\frac{1}{2}\min_0 M
\end{pmatrix} ⟺ S = 0 0 2 1 min 0 M M − 2 1 min 0 M
while, akin to Proposition 3.3 , the image of N N N may be characterised as follows:
S ∈ im N = : W ⟺ M ( S ) : = θ + q = 0 ⟺ S = ( u h w η − η ) . S\in\operatorname{im}N=:W \Longleftrightarrow M(S):=\theta+q=0 \Longleftrightarrow S=
\begin{pmatrix}
u_h\\
w\\
\eta\\
-\eta
\end{pmatrix}. S ∈ im N =: W ⟺ M ( S ) := θ + q = 0 ⟺ S = u h w η − η .
This leads to the following decomposition of R 5 \mathbb{R}^5 R 5 :
S = ( u h w θ q ) = ( 0 0 1 2 min 0 M M − 1 2 min 0 M ) ⏟ ∈ ker N = B + ( u h w η − η ) ⏟ ∈ im N = W . (45) S=
\begin{pmatrix}
u_h\\
w\\
\theta\\
q
\end{pmatrix}
=
\underbrace{\begin{pmatrix}
0\\
0\\
\frac{1}{2}\min_0 M\\
M-\frac{1}{2}\min_0 M
\end{pmatrix}}_{\in\ker N=B}
+
\underbrace{\begin{pmatrix}
u_h\\
w\\
\eta\\
-\eta
\end{pmatrix}}_{\in\operatorname{im}N=W}. \tag*{(45)} S = u h w θ q = ∈ k e r N = B 0 0 2 1 min 0 M M − 2 1 min 0 M + ∈ im N = W u h w η − η . (45)
Now let us restrict our attention once again to the ( θ , q ) (\theta,q) ( θ , q ) –plane. We seek to determine whether or not the Snell-type metric agrees with the decomposition (45) . In the ( θ , q ) (\theta,q) ( θ , q ) –plane, the balanced component takes the form
( θ B q B ) = ( 1 2 min 0 M M − 1 2 min 0 M ) . \begin{pmatrix}
\theta_B\\
q_B
\end{pmatrix}
=
\begin{pmatrix}
\frac{1}{2}\min_0 M\\
M-\frac{1}{2}\min_0 M
\end{pmatrix}. ( θ B q B ) = ( 2 1 min 0 M M − 2 1 min 0 M ) .
It turns out that this decomposition fails to agree with the Snell-type energy.
Proposition 5.18 (The Snell-type distance does not agree with the pointwise decomposition). As introduced in Definition 5.14 , d S n e l l d_{\mathrm{Snell}} d Snell does not agree with the extraction of a balanced component, in the sense that there exists S ∈ R 5 S\in\mathbb{R}^5 S ∈ R 5 and T ∗ ∈ B T_*\in B T ∗ ∈ B such that
d S n e l l ( S , T ∗ ) < d S n e l l ( S , S B ) . d_{\mathrm{Snell}}(S,T_*)<d_{\mathrm{Snell}}(S,S_B). d Snell ( S , T ∗ ) < d Snell ( S , S B ) . In other words:
S B ≠ arg min T ∈ B d S n e l l ( S , T ) . S_B\ne\underset{T\in B}{\arg\min}\ d_{\mathrm{Snell}}(S,T). S B = T ∈ B arg min d Snell ( S , T ) .
Proof. Choose S = ( − 3 , 1 ) S=(-3,1) S = ( − 3 , 1 ) such that M = − 3 + 1 = − 2 M=-3+1=-2 M = − 3 + 1 = − 2 and so its balanced component is S B = ( − 1 , − 1 ) S_B=(-1,-1) S B = ( − 1 , − 1 ) and choose T ∗ = ( − 1.1 , − 1.1 ) T_*=(-1.1,-1.1) T ∗ = ( − 1.1 , − 1.1 ) . Note that S S S is saturated whereas S B S_B S B and T ∗ T_* T ∗ are unsaturated. And, of course, both S B S_B S B and T ∗ T_* T ∗ belong to the balanced set B = ker N B=\ker N B = ker N since they have vanishing M M M .
On the one hand, equipped with Lemma 5.17 we see that
d S n e l l ( S , S B ) = ∥ S − I 1 ∥ s + ∥ S B − I 1 ∥ u d_{\mathrm{Snell}}(S,S_B)=\lVert S-I_1\rVert_s+\lVert S_B-I_1\rVert_u d Snell ( S , S B ) = ∥ S − I 1 ∥ s + ∥ S B − I 1 ∥ u
where
I 1 : = I ( θ 1 ) for θ 1 = − 3 + 3 2 I_1:=I(\theta_1)\text{ for }\theta_1=-3+\frac{\sqrt{3}}{2} I 1 := I ( θ 1 ) for θ 1 = − 3 + 2 3
and hence
d S n e l l ( S , S B ) = 5 2 − 3 ( 1 + 3 ) > 2.394. d_{\mathrm{Snell}}(S,S_B)=\sqrt{\frac{5}{2}-\sqrt{3}(1+\sqrt{3})}>2.394. d Snell ( S , S B ) = 2 5 − 3 ( 1 + 3 ) > 2.394.
On the other hand, for I 2 : = I ( θ 2 ) I_2:=I(\theta_2) I 2 := I ( θ 2 ) where θ 2 = 11 526 ( − 135 + 17 3 ) \theta_2=\frac{11}{526}(-135+17\sqrt{3}) θ 2 = 526 11 ( − 135 + 17 3 ) ,
d S n e l l ( S , T ∗ ) = ∥ S − I 2 ∥ s + ∥ T ∗ − I 2 ∥ u = 101488 − 31790 3 ( 10 + 11 3 ) 2630 < 2.381. d_{\mathrm{Snell}}(S,T_*)=\lVert S-I_2\rVert_s+\lVert T_*-I_2\rVert_u=\frac{\sqrt{101488-31790\sqrt{3}(10+11\sqrt{3})}}{2630}<2.381. d Snell ( S , T ∗ ) = ∥ S − I 2 ∥ s + ∥ T ∗ − I 2 ∥ u = 2630 101488 − 31790 3 ( 10 + 11 3 ) < 2.381.
Since
d S n e l l ( S , T ∗ ) < d S n e l l ( S , S B ) d_{\mathrm{Snell}}(S,T_*)<d_{\mathrm{Snell}}(S,S_B) d Snell ( S , T ∗ ) < d Snell ( S , S B )
this proves the claim.
Figure 6. A graphical depiction of Proposition 5.18. Note that the location of the points is not exact: this is an artistic rendering of the setup of the proof of Proposition 5.18 used because it makes it easier to see what is going on.
We conclude this section by extending the Snell-type distance discussed so far pointwise to a distance between spatially-varying states.
Definition 5.19 (Snell-type metric, spatially-varying states). We define d S : L σ 2 × L σ 2 → [ 0 , ∞ ) d_S : L^2_\sigma\times L^2_\sigma\to[0,\infty) d S : L σ 2 × L σ 2 → [ 0 , ∞ ) as follows. For any X , Y ∈ L σ 2 \mathcal{X}, \mathcal{Y} \in L^2_\sigma X , Y ∈ L σ 2 ,
d S ( X , Y ) : = ∥ d S n e l l ( X , Y ) ∥ L 2 = ( ∫ d S n e l l 2 ( X ( x ) , Y ( x ) ) d x ) 1 / 2 . d_S(\mathcal{X},\mathcal{Y}) := \left\|d_{\mathrm{Snell}}(\mathcal{X},\mathcal{Y})\right\|_{L^2} = \left(\int d_{\mathrm{Snell}}^2(\mathcal{X}(x),\mathcal{Y}(x))\,dx\right)^{1/2}. d S ( X , Y ) := ∥ d Snell ( X , Y ) ∥ L 2 = ( ∫ d Snell 2 ( X ( x ) , Y ( x )) d x ) 1/2 .
This distance is indeed a metric, and it agrees with the conserved energy when comparable to it.
Proposition 5.20. As introduced in Definition 5.19 , d S d_S d S is a metric on L σ 2 L^2_\sigma L σ 2 which agrees with the energy in the sense that, if X 1 , X 2 ∈ L σ 2 \mathcal{X}_1,\mathcal{X}_2 \in L^2_\sigma X 1 , X 2 ∈ L σ 2 have the same phases, meaning that the signs of q 1 q_1 q 1 and q 2 q_2 q 2 agree everywhere, then
( d S ( X 1 , X 2 ) ) 2 = E H ( X 1 − X 2 ) \left(d_S(\mathcal{X}_1,\mathcal{X}_2)\right)^2 = E_H(\mathcal{X}_1-\mathcal{X}_2) ( d S ( X 1 , X 2 ) ) 2 = E H ( X 1 − X 2 ) = ∫ [ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 ] . = \int\left[|u_1-u_2|^2+(\theta_1-\theta_2)^2+(q_1-q_2)^2H+\left((\theta_1+q_1)-(\theta_2+q_2)\right)^2\right]. = ∫ [ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 ] . for H : = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) H:=\mathbb{1}(q_1<0)=\mathbb{1}(q_2<0) H := 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) .
Proof. First we verify that d S d_S d S is a metric. The positive-definiteness and symmetry follow immediately from the fact that d S n e l l d_{\mathrm{Snell}} d Snell is a metric, since the L 2 L^2 L 2 norm is also positive-definite.
The triangle inequality of d S d_S d S then follows from the triangle inequalities of d S n e l l d_{\mathrm{Snell}} d Snell and of L 2 L^2 L 2 and from the fact that s ↦ s s\mapsto\sqrt{s} s ↦ s and s ↦ s 2 s\mapsto s^2 s ↦ s 2 are increasing:
d S ( X 1 , X 2 ) = ( ∫ d S n e l l 2 ( X 1 , X 2 ) ) 1 / 2 ≤ ( ∫ ( d S n e l l ( X 1 , X 3 ) + d S n e l l ( X 3 , X 2 ) ) 2 ) 1 / 2 = ∥ d S n e l l ( X 1 , X 3 ) + d S n e l l ( X 3 , X 2 ) ∥ L 2 ≤ ∥ d S n e l l ( X 1 , X 3 ) ∥ L 2 + ∥ d S n e l l ( X 3 , X 2 ) ∥ L 2 = d S ( X 1 , X 3 ) + d S ( X 3 , X 2 ) . \begin{aligned}
d_S(\mathcal{X}_1,\mathcal{X}_2)=\left(\int d_{\mathrm{Snell}}^2(\mathcal{X}_1,\mathcal{X}_2)\right)^{1/2} \\
&\leq\left(\int\left(d_{\mathrm{Snell}}(\mathcal{X}_1,\mathcal{X}_3)+d_{\mathrm{Snell}}(\mathcal{X}_3,\mathcal{X}_2)\right)^2\right)^{1/2} \\
&=\left\|d_{\mathrm{Snell}}(\mathcal{X}_1,\mathcal{X}_3)+d_{\mathrm{Snell}}(\mathcal{X}_3,\mathcal{X}_2)\right\|_{L^2} \\
&\leq\left\|d_{\mathrm{Snell}}(\mathcal{X}_1,\mathcal{X}_3)\right\|_{L^2}+\left\|d_{\mathrm{Snell}}(\mathcal{X}_3,\mathcal{X}_2)\right\|_{L^2} \\
&=d_S(\mathcal{X}_1,\mathcal{X}_3)+d_S(\mathcal{X}_3,\mathcal{X}_2).
\end{aligned} d S ( X 1 , X 2 ) = ( ∫ d Snell 2 ( X 1 , X 2 ) ) 1/2 ≤ ( ∫ ( d Snell ( X 1 , X 3 ) + d Snell ( X 3 , X 2 ) ) 2 ) 1/2 = ∥ d Snell ( X 1 , X 3 ) + d Snell ( X 3 , X 2 ) ∥ L 2 ≤ ∥ d Snell ( X 1 , X 3 ) ∥ L 2 + ∥ d Snell ( X 3 , X 2 ) ∥ L 2 = d S ( X 1 , X 3 ) + d S ( X 3 , X 2 ) .
Finally we prove that d S d_S d S agrees with the energy, when comparable. So suppose that q 1 ( x ) q_1(x) q 1 ( x ) and q 2 ( x ) q_2(x) q 2 ( x ) have the same phases, i.e. their signs agree everywhere in the spatial domain. Then, by virtue of Lemma 5.16 , we have that
d S 2 ( X 1 , X 2 ) = ∫ d S n e l l 2 ( X 1 ( x ) , X 2 ( x ) ) d x = ∫ [ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 ] = E H ( X 1 − X 2 ) \begin{aligned}
d_S^2(\mathcal{X}_1,\mathcal{X}_2)=\int d_{\mathrm{Snell}}^2(\mathcal{X}_1(x),\mathcal{X}_2(x))\,\mathrm{d}x \\
&=\int\left[|u_1-u_2|^2+(\theta_1-\theta_2)^2+(q_1-q_2)^2H+\left((\theta_1+q_1)-(\theta_2+q_2)\right)^2\right] \\
&=E_H(\mathcal{X}_1-\mathcal{X}_2)
\end{aligned} d S 2 ( X 1 , X 2 ) = ∫ d Snell 2 ( X 1 ( x ) , X 2 ( x )) d x = ∫ [ ∣ u 1 − u 2 ∣ 2 + ( θ 1 − θ 2 ) 2 + ( q 1 − q 2 ) 2 H + ( ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ) 2 ] = E H ( X 1 − X 2 )
for H : = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) H:=\mathbb{1}(q_1<0)=\mathbb{1}(q_2<0) H := 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) .
Remark 5.21. Finally we note that, although this distance agrees with the energy, the evidence from the restricted case of spatially constant cases strongly suggests that this Snell-type distance does not agree with the extraction of a balanced component, especially in light of Lemma A.23 which tells us that spatially constant solutions are honest-to-goodness solutions of the moist Boussinesq system considered here.
While this Snell-type distance may not agree with the balanced–unbalanced decomposition, it could be useful for other applications such as data-driven methods that identify nonlinear features in dynamical systems and utilize a metric, distance function, and/or kernel function [19 , 44 , 9 , 32 , 42 ] .
Metric-less approach
To make the approach of this section easy to follow we must be clear about what the last two sections, § 5.1 and § 5.2 , achieved.
In § 5.1 we devised a metric which agreed with the decomposition, but which failed to agree with the energy (it agreed only conditionally ).
In § 5.2 we devised a metric which agreed with the energy, but which failed to agree with the decomposition.
In this section we will see that it is possible to agree with both the decomposition and the energy, at a price. That price is that we lose the triangle inequality. Instead of a metric, we will therefore be dealing with a statistical divergence.
Definition 5.22 (Statistical divergence). Let H \mathcal{H} H be a Hilbert space. A statistical divergence on H \mathcal{H} H is a map D : H × H → R D:\mathcal{H}\times\mathcal{H}\to\mathbb{R} D : H × H → R satisfying the following properties.
Non-negative: D ( x , y ) ≥ 0 D(x,y)\ge0 D ( x , y ) ≥ 0 for every x , y ∈ H x,y\in\mathcal{H} x , y ∈ H .
Definite: D ( x , y ) = 0 D(x,y)=0 D ( x , y ) = 0 if and only if x = y x=y x = y .
Limiting quadratic: for every x ∈ H x\in\mathcal{H} x ∈ H there is a positive-definite quadratic form Q x Q_x Q x such that, for every y ∈ H y\in\mathcal{H} y ∈ H ,
1 ε 2 D ( x , x + ε y ) → Q x ( y ) \frac{1}{\varepsilon^2}D(x,x+\varepsilon y)\to Q_x(y) ε 2 1 D ( x , x + ε y ) → Q x ( y ) as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . Since this quadratic form gives rise to a positive-definite inner product, Q x Q_x Q x is often referred to as the Riemann metric induced by the divergence .
The prototypical example of a statistical divergence is the relative entropy, or Kullback–Leibler divergence, from information theory [24 , 13 , 22 ] . Another reminiscent notion from the PDE literature is the relative energy [28 , 38 , 40 , 8 ] .
The statistical divergence here is built in a simple way: we use the metric d H d_H d H introduced in § 5.1 , but now let H H H vary depending on the states whose divergence is being measured.
Definition 5.23. For any X 1 , X 2 ∈ L σ 2 \mathcal{X}_1,\mathcal{X}_2\in L^2_\sigma X 1 , X 2 ∈ L σ 2 we define
D ( X 1 ; X 2 ) : = 1 2 d H B 1 ( X 1 , X 2 ) 2 D(\mathcal{X}_1;\mathcal{X}_2):=\frac{1}{2}d_{H_{B_1}}(\mathcal{X}_1,\mathcal{X}_2)^2 D ( X 1 ; X 2 ) := 2 1 d H B 1 ( X 1 , X 2 ) 2 where d H d_H d H is as in Definition 5.10 , where H B 1 : = 1 ( q B 1 < 0 ) H_{B_1}:=\mathbb{1}(q_{B_1}<0) H B 1 := 1 ( q B 1 < 0 ) denotes the balance interface of the first state X 1 \mathcal{X}_1 X 1 .
Remark 5.24. We detail how to compute the statistical divergence D D D introduced in Definition 5.23 (similarly to how we detailed the computation of the metric d H d_H d H in Definition 5.10 ). So let X 1 = ( u 1 , θ 1 , q 1 ) \mathcal{X}_1=(u_1,\theta_1,q_1) X 1 = ( u 1 , θ 1 , q 1 ) and X 2 = ( u 2 , θ 2 , q 2 ) \mathcal{X}_2=(u_2,\theta_2,q_2) X 2 = ( u 2 , θ 2 , q 2 ) be states in L σ 2 \mathbb{L}^2_\sigma L σ 2 .
First we compute X B , 1 \mathcal{X}_{B,1} X B , 1 in order to determine the indicator H B 1 H_{B_1} H B 1 . This is done by setting
M 1 : = M ( X 1 ) = θ 1 + q 1 and P V 1 = P V ( X 1 ) = ∇ h ⊥ ⋅ u h , 1 + ∂ 3 θ 1 M_1:=M(\mathcal{X}_1)=\theta_1+q_1\quad\text{and}\quad PV_1=PV(\mathcal{X}_1)=\nabla_h^\perp\cdot u_{h,1}+\partial_3\theta_1 M 1 := M ( X 1 ) = θ 1 + q 1 and P V 1 = P V ( X 1 ) = ∇ h ⊥ ⋅ u h , 1 + ∂ 3 θ 1
and finding the solution p 1 ∈ H ˙ 1 p_1\in\dot{H}^1 p 1 ∈ H ˙ 1 of
Δ p 1 + 1 2 ∂ 3 min 0 ( M 1 − ∂ 3 p 1 ) = P V 1 . \Delta p_1+\frac{1}{2}\partial_3\min_0\left(M_1-\partial_3p_1\right)=PV_1. Δ p 1 + 2 1 ∂ 3 0 min ( M 1 − ∂ 3 p 1 ) = P V 1 .
The phases of the balanced component of the first state are then encoded in
H B 1 : = 1 ( q B , 1 < 0 ) = 1 ( ∂ 3 p 1 < M 1 ) . H_{B_1}:=\mathbb{1}(q_{B,1}<0)=\mathbb{1}(\partial_3p_1<M_1). H B 1 := 1 ( q B , 1 < 0 ) = 1 ( ∂ 3 p 1 < M 1 ) .
Second we proceed as in Definition 5.10 and compute d H B 1 ( X 1 , X 2 ) d_{H_{B_1}}(\mathcal{X}_1,\mathcal{X}_2) d H B 1 ( X 1 , X 2 ) . This means that we let
P V : = P V ( X 1 ) − P V ( X 2 ) , M : = M ( X 1 ) − M ( X 2 ) , PV:=PV(\mathcal{X}_1)-PV(\mathcal{X}_2),\quad M:=M(\mathcal{X}_1)-M(\mathcal{X}_2), P V := P V ( X 1 ) − P V ( X 2 ) , M := M ( X 1 ) − M ( X 2 ) ,
j : = j ( X 1 ) − j ( X 2 ) , w : = w ( X 1 ) − w ( X 2 ) , and a : = a ( X 1 ) − a ( X 2 ) , j:=j(\mathcal{X}_1)-j(\mathcal{X}_2),\quad w:=w(\mathcal{X}_1)-w(\mathcal{X}_2),\quad\text{and}\quad a:=a(\mathcal{X}_1)-a(\mathcal{X}_2), j := j ( X 1 ) − j ( X 2 ) , w := w ( X 1 ) − w ( X 2 ) , and a := a ( X 1 ) − a ( X 2 ) ,
and that we let p ∈ H ˙ 1 p\in\dot{H}^1 p ∈ H ˙ 1 and σ ∈ ( L 2 ) σ 3 \sigma\in(L^2)^3_\sigma σ ∈ ( L 2 ) σ 3 solve
∇ ⋅ ( A H B 1 ∇ p ) = P V − 1 2 ∂ 3 ( H M ) \nabla\cdot(A_{H_{B_1}}\nabla p)=PV-\frac{1}{2}\partial_3(HM) ∇ ⋅ ( A H B 1 ∇ p ) = P V − 2 1 ∂ 3 ( H M )
and
∇ × ( A H B 1 − 1 σ ) = j + ( ∂ 3 w ) e 3 subject to ∫ A H B 1 − 1 σ = a and ∇ ⋅ σ = 0 \nabla\times\left(A_{H_{B_1}}^{-1}\sigma\right)=j+(\partial_3w)e_3\quad\text{subject to}\quad\int A_{H_{B_1}}^{-1}\sigma=a\quad\text{and}\quad\nabla\cdot\sigma=0 ∇ × ( A H B 1 − 1 σ ) = j + ( ∂ 3 w ) e 3 subject to ∫ A H B 1 − 1 σ = a and ∇ ⋅ σ = 0
for A A A and A − 1 A^{-1} A − 1 as in Definition 5.5 .
So finally we have that
D ( X 1 , X 2 ) = 1 2 d H B 1 ( X 1 , X 2 ) 2 = ∫ Q H + ( p ) + ( 1 + H 2 ) M 2 + Q H − ( σ ) + w 2 D(\mathcal{X}_1,\mathcal{X}_2)=\frac{1}{2}d_{H_{B_1}}(\mathcal{X}_1,\mathcal{X}_2)^2=\int Q^+_H(p)+\left(1+\frac{H}{2}\right)M^2+Q^-_H(\sigma)+w^2 D ( X 1 , X 2 ) = 2 1 d H B 1 ( X 1 , X 2 ) 2 = ∫ Q H + ( p ) + ( 1 + 2 H ) M 2 + Q H − ( σ ) + w 2
for Q H ± Q_H^\pm Q H ± as introduced in Definition 5.5 . We note that, in the more concise notation of Definition 5.10 we may write this as
D ( X 1 , X 2 ) = E ~ H B 1 ( C H B 1 [ M ( X 1 ) − M ( X 2 ) ] ) D(\mathcal{X}_1,\mathcal{X}_2)=\widetilde{E}_{H_{B_1}}\left(C_{H_{B_1}}\left[M(\mathcal{X}_1)-M(\mathcal{X}_2)\right]\right) D ( X 1 , X 2 ) = E H B 1 ( C H B 1 [ M ( X 1 ) − M ( X 2 ) ] )
Remark 5.25. We saw in Remark 5.24 that we may write the statistical divergence in terms of C H C_H C H . This map is an analog of C C C , which is the map from measurement space to coordinate space which associates to a state’s measurements its corresponding coordinates (see Remark 3.9). The analogous map C H C_H C H proceeds similarly, except that now the three transformations (between state space, measurement space, and coordinate space) are linear . The linearity comes from treating a particular indicator function H H H as fixed.
More precisely, given an indicator function H H H , we have the following.
S H ( p , M , σ , w ) = ( u , θ , q ) ⟺ { u = ∇ h ⊥ p + σ h ⊥ , θ = ∂ 3 p + 1 2 H ( M − ∂ 3 p ) + σ 3 q = M − ∂ 3 p − 1 2 H ( M − ∂ 3 p ) − σ 3 , S_H(p,M,\sigma,w)=(u,\theta,q)\Longleftrightarrow
\begin{cases}
u=\nabla_h^\perp p+\sigma_h^\perp,\\
\theta=\partial_3p+\frac{1}{2}H(M-\partial_3p)+\sigma_3\\
q=M-\partial_3p-\frac{1}{2}H(M-\partial_3p)-\sigma_3,
\end{cases} S H ( p , M , σ , w ) = ( u , θ , q ) ⟺ ⎩ ⎨ ⎧ u = ∇ h ⊥ p + σ h ⊥ , θ = ∂ 3 p + 2 1 H ( M − ∂ 3 p ) + σ 3 q = M − ∂ 3 p − 2 1 H ( M − ∂ 3 p ) − σ 3 ,
where recall that C \mathcal{C} C defined in Remark 3.9 is the coordinate space.
M H ( u , θ , q ) = ( P V , M , j H , w , a H ) M_H(u,\theta,q)=(PV,M,j_H,w,a_H) M H ( u , θ , q ) = ( P V , M , j H , w , a H )
where
j H ( u , θ , q ) = ∂ 3 u h − ∇ h ⊥ ( θ − q H ) and a H = − ∫ u h ⊥ + ∫ ( θ − q H ) e 3 . j_H(u,\theta,q)=\partial_3u_h-\nabla_h^\perp(\theta-qH)\quad\text{and}\quad a_H=-\int u_h^\perp+\int(\theta-qH)e_3. j H ( u , θ , q ) = ∂ 3 u h − ∇ h ⊥ ( θ − q H ) and a H = − ∫ u h ⊥ + ∫ ( θ − q H ) e 3 .
This means that only j j j and a a a are modified into j H j_H j H and a H a_H a H , respectively, but the remaining measurements are unchanged. This is because they are the only nonlinear measurements in the first place since they are the only measurements involving the buoyancy.
These maps will all be used below to compute the Riemann metric induced by the statistical divergence introduced in Definition 5.23 .
Proposition 5.26. The map D : L σ 2 × L σ 2 → [ 0 , ∞ ) D : \mathcal{L}^2_\sigma\times\mathcal{L}^2_\sigma\to[0,\infty) D : L σ 2 × L σ 2 → [ 0 , ∞ ) introduced in Definition 5.23 is a statistical divergence in the sense of Definition 5.22 . In particular, for any X , Y ∈ L σ 2 \mathcal{X},\mathcal{Y} \in\mathcal{L}^2_\sigma X , Y ∈ L σ 2 ,
1 ε 2 D ( X ; X + ε Y ) → E B ( [ Y ] X ) \frac{1}{\varepsilon^2}D(\mathcal{X};\mathcal{X}+\varepsilon\mathcal{Y}) \to E_B([\mathcal{Y}]_{\mathcal{X}}) ε 2 1 D ( X ; X + ε Y ) → E B ([ Y ] X ) where [ Y ] X : = ( S B ∘ C B ∘ M H ) ( Y ) [\mathcal{Y}]_{\mathcal{X}} := (S_B \circ C_B \circ M_H)(\mathcal{Y}) [ Y ] X := ( S B ∘ C B ∘ M H ) ( Y ) is known as the X \mathcal{X} X -representative of Y \mathcal{Y} Y (such that if H X = H Y H_{\mathcal{X}}=H_{\mathcal{Y}} H X = H Y and H B , X = H B , Y H_{B,\mathcal{X}}=H_{B,\mathcal{Y}} H B , X = H B , Y then [ Y ] X = Y [\mathcal{Y}]_{\mathcal{X}}=\mathcal{Y} [ Y ] X = Y ). Moreover
this statistical divergence agrees with the conserved energy in the sense that, if H 1 = H 2 = H B , 1 = H B , 2 H_1=H_2=H_{B,1}=H_{B,2} H 1 = H 2 = H B , 1 = H B , 2 then
D ( X 1 ; X 2 ) = E H ( X 1 − X 2 ) D(\mathcal{X}_1;\mathcal{X}_2)=E_H(\mathcal{X}_1-\mathcal{X}_2) D ( X 1 ; X 2 ) = E H ( X 1 − X 2 ) where H = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) H=\mathbf{1}(q_1<0)=\mathbf{1}(q_2<0) H = 1 ( q 1 < 0 ) = 1 ( q 2 < 0 ) and
this statistical divergence agrees with the extraction of a balanced component in the sense that
X B = arg min Y ∈ B D ( X ; Y ) . \mathcal{X}_B=\underset{\mathcal{Y}\in\mathcal{B}}{\arg\min}\ D(\mathcal{X};\mathcal{Y}). X B = Y ∈ B arg min D ( X ; Y ) .
Proof. The fact that D D D is non-negative and definite follows immediately from the positive-definiteness of d H d_H d H , which is established in Proposition 5.11 . The fact that D D D agrees with the energy and is compatible with the extraction of a balanced component also follows from the proof of Proposition 5.11 . So finally we turn our attention to the Riemann metric induced by this statistical divergence. Let X , Y ∈ L σ 2 \mathcal{X},\mathcal{Y}\in\mathcal{L}^2_\sigma X , Y ∈ L σ 2 , which we write as X = ( u , θ , q ) \mathcal{X}=(u,\theta,q) X = ( u , θ , q ) and Y = ( v , φ , r ) \mathcal{Y}=(v,\varphi,r) Y = ( v , φ , r ) , and let ε > 0 \varepsilon>0 ε > 0 .
We note that
j ( X + ε Y ) − j ( X ) = ∂ 3 ( u h + ε v h ) − ∇ h ⊥ ( θ + ε φ ) + ∇ h ( min 0 ( q + ε r ) ) − ∂ 3 u h + ∇ h ⊥ θ − ∇ h min 0 q = ε ∂ 3 v h − ε ∇ h ⊥ φ + ∇ h [ min 0 ( q + ε r ) − min 0 q ] = ε ∂ 3 v h − ε ∇ h ⊥ φ + ∇ h [ ε r 1 ( q + ε r < 0 ) + R ε ( q ; r ) ] = ε j H ( X + ε Y ) ( Y ) + ∇ h R ε ( q ; r ) \begin{aligned}
j(\mathcal{X}+\varepsilon\mathcal{Y})-j(\mathcal{X})
&=\partial_3(u_h+\varepsilon v_h)-\nabla_h^\perp(\theta+\varepsilon\varphi)+\nabla_h\bigl(\min_0(q+\varepsilon r)\bigr)\\
&\quad-\partial_3u_h+\nabla_h^\perp\theta-\nabla_h\min_0q\\
&=\varepsilon\partial_3v_h-\varepsilon\nabla_h^\perp\varphi+\nabla_h\bigl[\min_0(q+\varepsilon r)-\min_0q\bigr]\\
&=\varepsilon\partial_3v_h-\varepsilon\nabla_h^\perp\varphi+\nabla_h\bigl[\varepsilon r\mathbf{1}(q+\varepsilon r<0)+R_\varepsilon(q;r)\bigr]\\
&=\varepsilon j_{H(\mathcal{X}+\varepsilon\mathcal{Y})}(\mathcal{Y})+\nabla_h R_\varepsilon(q;r)
\end{aligned} j ( X + ε Y ) − j ( X ) = ∂ 3 ( u h + ε v h ) − ∇ h ⊥ ( θ + εφ ) + ∇ h ( 0 min ( q + ε r ) ) − ∂ 3 u h + ∇ h ⊥ θ − ∇ h 0 min q = ε ∂ 3 v h − ε ∇ h ⊥ φ + ∇ h [ 0 min ( q + ε r ) − 0 min q ] = ε ∂ 3 v h − ε ∇ h ⊥ φ + ∇ h [ ε r 1 ( q + ε r < 0 ) + R ε ( q ; r ) ] = ε j H ( X + ε Y ) ( Y ) + ∇ h R ε ( q ; r )
for
R ε ( q ; r ) = q [ 1 ( q + ε r < 0 ) − 1 ( q < 0 ) ] = ε ⋅ q ε [ 1 ( q ε + r < 0 ) − 1 ( q ε < 0 ) ] \begin{aligned}
R_\varepsilon(q;r)&=q\bigl[\mathbf{1}(q+\varepsilon r<0)-\mathbf{1}(q<0)\bigr]\\
&=\varepsilon\cdot\frac{q}{\varepsilon}\left[\mathbf{1}\left(\frac{q}{\varepsilon}+r<0\right)-\mathbf{1}\left(\frac{q}{\varepsilon}<0\right)\right]
\end{aligned} R ε ( q ; r ) = q [ 1 ( q + ε r < 0 ) − 1 ( q < 0 ) ] = ε ⋅ ε q [ 1 ( ε q + r < 0 ) − 1 ( ε q < 0 ) ]
where δ 0 \delta_0 δ 0 denotes the Dirac delta distribution at zero such that, since q ( δ 0 ∘ q ) = 0 q(\delta_0\circ q)=0 q ( δ 0 ∘ q ) = 0 , 1 ε R ε ( q ; r ) → 0 \frac{1}{\varepsilon}R_\varepsilon(q;r)\to0 ε 1 R ε ( q ; r ) → 0 as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 .
Therefore, since all other measurements in M \mathcal{M} M besides j j j are linear,
M ( X + ε Y ) − M ( X ) = ( ε P V ( Y ) ε M ( Y ) ε j H ( X + ε Y ) ( Y ) + ∇ h ⊥ R ε ( q ; r ) ε w ( Y ) ε a ( Y ) ) = ε ( P V ( Y ) M ( Y ) j H ( X + ε Y ) ( Y ) w ( Y ) a ( Y ) ) + ε ( 0 0 ∇ h ⊥ 1 ε R ε 0 0 ) = : ε M H ε ( Y ) + ε R , \begin{aligned}
\mathcal{M}(\mathcal{X}+\varepsilon\mathcal{Y})-\mathcal{M}(\mathcal{X})
&=
\begin{pmatrix}
\varepsilon PV(\mathcal{Y})\\
\varepsilon M(\mathcal{Y})\\
\varepsilon j_{H(\mathcal{X}+\varepsilon\mathcal{Y})}(\mathcal{Y})+\nabla_h^\perp R_\varepsilon(q;r)\\
\varepsilon w(\mathcal{Y})\\
\varepsilon a(\mathcal{Y})
\end{pmatrix}\\
&=\varepsilon
\begin{pmatrix}
PV(\mathcal{Y})\\
M(\mathcal{Y})\\
j_{H(\mathcal{X}+\varepsilon\mathcal{Y})}(\mathcal{Y})\\
w(\mathcal{Y})\\
a(\mathcal{Y})
\end{pmatrix}
+\varepsilon
\begin{pmatrix}
0\\
0\\
\nabla_h^\perp\frac{1}{\varepsilon}R_\varepsilon\\
0\\
0
\end{pmatrix}\\
&=:\varepsilon M_{H_\varepsilon}(\mathcal{Y})+\varepsilon\mathcal{R},
\end{aligned} M ( X + ε Y ) − M ( X ) = εP V ( Y ) εM ( Y ) ε j H ( X + ε Y ) ( Y ) + ∇ h ⊥ R ε ( q ; r ) εw ( Y ) ε a ( Y ) = ε P V ( Y ) M ( Y ) j H ( X + ε Y ) ( Y ) w ( Y ) a ( Y ) + ε 0 0 ∇ h ⊥ ε 1 R ε 0 0 =: ε M H ε ( Y ) + ε R ,
where as noted above R → 0 \mathcal{R}\to0 R → 0 when ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 .
So finally we have that
D ( X ; X + ε Y ) = E B ( C B [ M ( X + ε Y ) − M ( X ) ] ) = E B ( C B [ ε M H ε ( Y ) + ε R ] ) = ε 2 E B ( [ Y ] X + ε Y + C B R ) \begin{aligned}
D(\mathcal{X};\mathcal{X}+\varepsilon\mathcal{Y})
&=E_B\bigl(C_B[\mathcal{M}(\mathcal{X}+\varepsilon\mathcal{Y})-\mathcal{M}(\mathcal{X})]\bigr)\\
&=E_B\bigl(C_B[\varepsilon M_{H_\varepsilon}(\mathcal{Y})+\varepsilon\mathcal{R}]\bigr)\\
&=\varepsilon^2E_B\bigl([\mathcal{Y}]_{\mathcal{X}+\varepsilon\mathcal{Y}}+C_B\mathcal{R}\bigr)
\end{aligned} D ( X ; X + ε Y ) = E B ( C B [ M ( X + ε Y ) − M ( X )] ) = E B ( C B [ ε M H ε ( Y ) + ε R ] ) = ε 2 E B ( [ Y ] X + ε Y + C B R )
such that indeed, since H ( X + ε Y ) → H ( X ) = H H(\mathcal{X}+\varepsilon\mathcal{Y})\to H(\mathcal{X})=H H ( X + ε Y ) → H ( X ) = H ,
1 ε 2 D ( X ; X + ε Y ) → E B ( [ Y ] X ) , \frac{1}{\varepsilon^2}D(\mathcal{X};\mathcal{X}+\varepsilon\mathcal{Y})\to E_B([\mathcal{Y}]_{\mathcal{X}}), ε 2 1 D ( X ; X + ε Y ) → E B ([ Y ] X ) ,
as claimed.
Summary
The takeaway message from this section is the following: if we wish to quantify the discrepancy between two states in L σ 2 L_{\sigma}^{2} L σ 2 in a manner which is both consistent with the decomposition and with the conserved energy then we must give up on the triangle inequality. As § 5.1 and § 5.2 show, we are unable to obtain a metric which has both of these properties. We therefore identify, in § 5.3 , a statistical divergence which fails to satisfy the triangle inequality but which is consistent with both the decomposition and the conserved energy. This is summarized in the table below.
Table 3.
Iterative methods for the decomposition and their geometry
In this section we discuss how to compute the decomposition of a state into its slow and fast components in practice . The only computationally non-trivial step in that process is to extract the balanced component, since this requires the inversion of the nonlinear elliptic PDE
Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) = P V \Delta p + \frac{1}{2}\partial_{3}\min_{0}\left(M-\partial_{3}p\right)=PV Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) = P V
where P V PV P V and M M M are given and we solve for p p p .
Our ability to solve this nonlinear elliptic PDE comes from identifying a variational form of this equation. That is the observation underpinning [89 ] . We recall this variational formulation here.
Theorem 6.1 (Variational formulation of P V PV P V -and-M M M inversion). For any M ∈ L 2 M \in L^{2} M ∈ L 2 , P V ∈ H − 1 PV \in H^{-1} P V ∈ H − 1 , and p ∈ H ˚ 1 p \in\mathring{H}^{1} p ∈ H ˚ 1 , p p p is a global minimizer of E v a r E_{\mathrm{var}} E var , defined as
E v a r : = ∫ T 3 1 2 ∣ ∇ p ∣ 2 − 1 4 min 0 ( M − ∂ 3 p ) 2 + ⟨ P V , p ⟩ H − 1 × H ˚ 1 , E_{\mathrm{var}}:=\int_{\mathbb{T}^{3}}\frac{1}{2}|\nabla p|^{2}-\frac{1}{4}\min_{0}\left(M-\partial_{3}p\right)^{2}+\langle PV,p\rangle_{H^{-1}\times\mathring{H}^{1}}, E var := ∫ T 3 2 1 ∣∇ p ∣ 2 − 4 1 0 min ( M − ∂ 3 p ) 2 + ⟨ P V , p ⟩ H − 1 × H ˚ 1 , if and only if p p p is an H 1 H^{1} H 1 -weak solution of
Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) = P V \Delta p+\frac{1}{2}\partial_{3}\min_{0}\left(M-\partial_{3}p\right)=PV Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) = P V in the sense that
∫ T 3 ∇ p ⋅ ∇ ϕ + 1 2 min 0 ( M − ∂ 3 p ) ∂ 3 ϕ = − ⟨ P V , ϕ ⟩ H − 1 × H ˚ 1 for every ϕ ∈ H ˚ 1 . (46) \int_{\mathbb{T}^{3}}\nabla p\cdot\nabla\phi+\frac{1}{2}\min_{0}\left(M-\partial_{3}p\right)\partial_{3}\phi=-\langle PV,\phi\rangle_{H^{-1}\times\mathring{H}^{1}}\text{ for every }\phi\in\mathring{H}^{1}. \tag*{(46)} ∫ T 3 ∇ p ⋅ ∇ ϕ + 2 1 0 min ( M − ∂ 3 p ) ∂ 3 ϕ = − ⟨ P V , ϕ ⟩ H − 1 × H ˚ 1 for every ϕ ∈ H ˚ 1 . (46) Proof. See Lemma 4.9 in [89 ] .
With this variational formulation of P V PV P V -and-M M M inversion in hand, we can now leverage iterative methods from optimization in order to find computational methods. We discuss such methods in § 6.1 below, obtaining convergence guarantees. In § 6.2 we then relate one of these iterative methods (arising from Newton’s method) to the geometry of the balanced set B \mathcal{B} B , obtaining a geometric interpretation of this descent method.
Convergence guarantees
In this section we introduce and compare several gradient descent methods that can be employed to minimise the variational energy introduced in Theorem 6.1 , and thus find solutions of P V PV P V -and-M M M inversion. We will discuss three descent methods: H 1 H^{1} H 1 -gradient descent, L 2 L^{2} L 2 -gradient descent, and Newton descent (which, as we will discuss, can be interpreted locally as a gradient descent), recording corresponding convergence rates for each method. Ultimately, the convergence rates of interest are those obtained for backtracking line search (see Definition 6.18) which are recorded in Theorem 6.20 but we also record analogous convergence rates in Theorem 6.16 for exact line search. This is done to make the rates obtained in the main result of Theorem 6.20 easier to contextualise, by comparing them with what can be obtained in the idealized scenario of exact line search.
We then conclude this section with a consistency result that guarantees that minimizers of the variational energy restricted to an (appropriate) finite-dimensional subspace of H ˙ 1 \dot{H}^{1} H ˙ 1 do converge to the true minimizer of the energy. Put together, the convergence rates of Theorem 6.20 and the consistency result of Theorem 6.22 tell us that we have a ready-to-implement method for finding approximate minimizers which are guaranteed to converge to the true minimizer. Before we get started we recall results from [89 ] that will be used in this section.
Theorem 6.2 (Properties of the variational energy). The variational energy E v a r E_{\mathrm{var}} E var defined in Theorem 6.1 has the following properties.
E v a r E_{\mathrm{var}} E var is 1 2 \frac{1}{2} 2 1 -strongly convex, weakly lower semi-continuous over H ˙ 1 \dot{H}^{1} H ˙ 1 , and coercive over H ˙ 1 \dot{H}^{1} H ˙ 1 in the sense that
E v a r ( p ) ≥ 1 16 ∥ p ∥ H ˙ 1 2 − 3 4 ∥ M ∥ L 2 2 − 4 ∥ P V ∥ H − 1 2 E_{\mathrm{var}}(p) \ge\frac{1}{16}\lVert p\rVert_{\dot{H}^{1}}^{2}-\frac{3}{4}\lVert M\rVert_{L^{2}}^{2}-4\lVert PV\rVert_{H^{-1}}^{2} E var ( p ) ≥ 16 1 ∥ p ∥ H ˙ 1 2 − 4 3 ∥ M ∥ L 2 2 − 4 ∥ P V ∥ H − 1 2 for every p ∈ H ˙ 1 p \in\dot{H}^{1} p ∈ H ˙ 1 .
We may write
E v a r ( p ) = ∫ T 3 e M ( x , ∇ p ( x ) ) d x + ⟨ P V , p ⟩ (47) E_{\mathrm{var}}(p)=\int_{\mathbb{T}^{3}}e_{M}(x,\nabla p(x))\,dx+\langle PV,p\rangle \tag*{(47)} E var ( p ) = ∫ T 3 e M ( x , ∇ p ( x )) d x + ⟨ P V , p ⟩ (47) for e M ( x , u ) : = 1 2 ∣ u ∣ 2 − 1 4 min ( M ( x ) − u 3 , 0 ) 2 e_{M}(x,u):=\frac{1}{2}\lvert u\rvert^{2}-\frac{1}{4}\min\left(M(x)-u_{3},0\right)^{2} e M ( x , u ) := 2 1 ∣ u ∣ 2 − 4 1 min ( M ( x ) − u 3 , 0 ) 2 .
E v a r E_{\mathrm{var}} E var is Gâteaux differentiable on H ˙ 1 \dot{H}^{1} H ˙ 1 with Gâteaux derivative at any p ∈ H ˙ 1 p\in\dot{H}^{1} p ∈ H ˙ 1 , denoted by D E v a r ( p ) DE_{\mathrm{var}}(p) D E var ( p ) , given by
D E v a r ( p ) ϕ = ∫ T 3 ∇ p ⋅ ∇ ϕ + 1 2 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ + ⟨ P V , ϕ ⟩ (48) DE_{\mathrm{var}}(p)\phi=\int_{\mathbb{T}^{3}}\nabla p\cdot\nabla\phi+\frac{1}{2}\min(M-\partial_{3}p,0)\partial_{3}\phi+\langle PV,\phi\rangle \tag*{(48)} D E var ( p ) ϕ = ∫ T 3 ∇ p ⋅ ∇ ϕ + 2 1 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ + ⟨ P V , ϕ ⟩ (48) for every ϕ ∈ H ˙ 1 \phi\in\dot{H}^{1} ϕ ∈ H ˙ 1 .
The Gâteaux derivative E v a r E_{\mathrm{var}} E var is Lipschitz, with specifically the estimate ∥ D E v a r ( p 1 ) − D E v a r ( p 2 ) ∥ H − 1 ≤ 3 2 ∥ p 1 − p 2 ∥ H ˙ 1 \lVert DE_{\mathrm{var}}(p_{1})-DE_{\mathrm{var}}(p_{2})\rVert_{H^{-1}}\le\frac{3}{2}\lVert p_{1}-p_{2}\rVert_{\dot{H}^{1}} ∥ D E var ( p 1 ) − D E var ( p 2 ) ∥ H − 1 ≤ 2 3 ∥ p 1 − p 2 ∥ H ˙ 1 for every p 1 , p 2 ∈ H ˙ 1 p_{1},p_{2}\in\dot{H}^{1} p 1 , p 2 ∈ H ˙ 1 , and so E v a r E_{\mathrm{var}} E var is Fréchet differentiable.
Suppose that E v a r E_{\mathrm{var}} E var has a minimizer p ∗ ∈ H ˙ 1 p^{*}\in\dot{H}^{1} p ∗ ∈ H ˙ 1 . For any p ∈ H ˙ 1 p\in\dot{H}^{1} p ∈ H ˙ 1 the following estimate holds:
∥ p − p ∗ ∥ H ˙ 1 2 ≤ 4 ( E v a r ( p ) − E v a r ( p ∗ ) ) = 4 ( E v a r ( p ) − min E v a r ) . \lVert p-p^{*}\rVert_{\dot{H}^{1}}^{2}\le4\left(E_{\mathrm{var}}(p)-E_{\mathrm{var}}(p^{*})\right)=4\left(E_{\mathrm{var}}(p)-\min E_{\mathrm{var}}\right). ∥ p − p ∗ ∥ H ˙ 1 2 ≤ 4 ( E var ( p ) − E var ( p ∗ ) ) = 4 ( E var ( p ) − min E var ) . Proof. See Proposition 4.4, Lemma 4.6, Corollary 4.7, and Lemma 4.8 of [89 ] .
In particular [89 ] establishes the well-posedness of nonlinear P V PV P V -and-M M M inversion.
Theorem 6.3 (Existence and uniqueness). For every M ∈ L 2 M\in L^{2} M ∈ L 2 , and P V ∈ H − 1 PV\in H^{-1} P V ∈ H − 1 the variational energy E v a r E_{\mathrm{var}} E var introduced in Theorem 6.1 has a unique global minimizer in H ˙ 1 \dot{H}^{1} H ˙ 1 which is also the unique weak solution of (46) . Moreover we have the following H 2 H^{2} H 2 regularity result ([89 ] proves higher-order regularity, but this H 2 H^{2} H 2 is all that is required below). Let M ∈ H 1 M\in H^{1} M ∈ H 1 and P V ∈ L 2 PV\in L^{2} P V ∈ L 2 . Any H ˙ 1 \dot{H}^{1} H ˙ 1 -weak solution p p p of (46) belongs to H 2 H^{2} H 2 and satisfies ∥ p ∥ H ˙ 2 ≤ ∥ M ∥ H ˙ 1 + 2 ∥ P V ∥ L 2 \lVert p\rVert_{\dot{H}^{2}}\le\lVert M\rVert_{\dot{H}^{1}}+2\lVert PV\rVert_{L^{2}} ∥ p ∥ H ˙ 2 ≤ ∥ M ∥ H ˙ 1 + 2 ∥ P V ∥ L 2 .
Proof. This is Theorem 5.1 and Lemma 5.2 in [89 ] .
Now to kick things off in earnest we define the gradient of a functional, taking particular care to highlight the role payed by the choice of the underlying inner product.
Definition 6.4 (Gradient). Consider a functional F : B → R \mathcal{F}:B\to\mathbb{R} F : B → R on a Banach space B B B equipped with an inner product ( ⋅ , ⋅ ) (\cdot,\cdot) ( ⋅ , ⋅ ) which is not necessarily complete and suppose that there exists a map D : B → B D:B\to B D : B → B such that D F ( u ) ϕ = ( D ( u ) , ϕ ) D\mathcal{F}(u)\phi=(D(u),\phi) D F ( u ) ϕ = ( D ( u ) , ϕ ) for every u , ϕ ∈ B u,\phi\in B u , ϕ ∈ B . We call D D D the gradient of F \mathcal{F} F with respect to the inner product ( ⋅ , ⋅ ) (\cdot,\cdot) ( ⋅ , ⋅ ) . Moreover, if ( ⋅ , ⋅ ) (\cdot,\cdot) ( ⋅ , ⋅ ) is known as the “X X X -inner product” then we write D = ∇ X F D=\nabla_{X}\mathcal{F} D = ∇ X F and call it the X X X -gradient of F \mathcal{F} F .
With Definition 6.4 in hand we may now turn our attention towards two gradients of the variational energy that we will use throughout: the H ˙ 1 \dot{H}^{1} H ˙ 1 and L 2 L^{2} L 2 -gradients. We record expressions for both of these below.
Lemma 6.5 (H ˙ 1 \dot{H}^{1} H ˙ 1 and L 2 L^{2} L 2 -gradients for our variational energy). Consider the variational energy E v a r E_{\mathrm{var}} E var introduced in Theorem 6.1 . The H ˙ 1 \dot{H}^{1} H ˙ 1 -gradient of E v a r E_{\mathrm{var}} E var is given by
( ∇ H ˙ 1 E v a r ) ( p ) = ( − Δ ) − 1 [ − Δ p − 1 2 ∂ 3 2 ( min ( M − ∂ 3 p , 0 ) ) + P V ] , \left(\nabla_{\dot{H}^{1}}E_{\mathrm{var}}\right)(p)=(-\Delta)^{-1}\left[-\Delta p-\frac{1}{2}\partial_{3}^{2}\left(\min(M-\partial_{3}p,0)\right)+PV\right], ( ∇ H ˙ 1 E var ) ( p ) = ( − Δ ) − 1 [ − Δ p − 2 1 ∂ 3 2 ( min ( M − ∂ 3 p , 0 ) ) + P V ] , meaning that u = ( ∇ H ˙ 1 E v a r ) ( p ) u=\left(\nabla_{\dot{H}^{1}}E_{\mathrm{var}}\right)(p) u = ( ∇ H ˙ 1 E var ) ( p ) is the unique weak solution of
− Δ u = − Δ p − 1 2 ∂ 3 2 ( min ( M − ∂ 3 p , 0 ) ) + P V (49) -\Delta u=-\Delta p-\frac{1}{2}\partial_{3}^{2}\left(\min(M-\partial_{3}p,0)\right)+PV \tag*{(49)} − Δ u = − Δ p − 2 1 ∂ 3 2 ( min ( M − ∂ 3 p , 0 ) ) + P V (49)
satisfying f T 3 u = 0 f_{\mathbb{T}^{3}}u=0 f T 3 u = 0 , i.e. u ∈ H ˚ 1 u\in\mathring{H}^{1} u ∈ H ˚ 1 such that, for every ϕ ∈ H ˚ 1 \phi\in\mathring{H}^{1} ϕ ∈ H ˚ 1 ,
∫ T 3 ∇ u ⋅ ∇ ϕ = ∫ T 3 ∇ p ⋅ ∇ ϕ + 1 2 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ + ⟨ P V , ϕ ⟩ (50) \int_{\mathbb{T}^{3}}\nabla u\cdot\nabla\phi=\int_{\mathbb{T}^{3}}\nabla p\cdot\nabla\phi+\frac{1}{2}\min(M-\partial_{3}p,0)\,\partial_{3}\phi+\langle PV,\phi\rangle \tag*{(50)} ∫ T 3 ∇ u ⋅ ∇ ϕ = ∫ T 3 ∇ p ⋅ ∇ ϕ + 2 1 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ + ⟨ P V , ϕ ⟩ (50)
Moreover, if p ∈ H 2 p\in H^{2} p ∈ H 2 , M ∈ H 1 M\in H^{1} M ∈ H 1 , and P V ∈ L 2 PV\in L^{2} P V ∈ L 2 then the L 2 L^{2} L 2 -gradient of E v a r E_{\mathrm{var}} E var is given by
( ∇ L 2 E v a r ) ( p ) = − Δ p − 1 2 ∂ 3 ( min ( M − ∂ 3 p , 0 ) ) + P V . (51) (\nabla_{L^{2}}E_{\mathrm{var}})(p)=-\Delta p-\frac{1}{2}\partial_{3}\bigl(\min(M-\partial_{3}p,0)\bigr)+PV. \tag*{(51)} ( ∇ L 2 E var ) ( p ) = − Δ p − 2 1 ∂ 3 ( min ( M − ∂ 3 p , 0 ) ) + P V . (51)
Proof. Item 3 of Theorem 6.2 tells us that u = ∇ H ˚ 1 E v a r ( p ) u=\nabla_{\mathring{H}^{1}}E_{\mathrm{var}}(p) u = ∇ H ˚ 1 E var ( p ) satisfies
∫ T 3 ∇ u ⋅ ∇ ϕ = ( u , ϕ ) H ˚ 1 = ( ∇ H ˚ 1 E v a r ( p ) , ϕ ) H ˚ 1 = D E v a r ( p ) ϕ = ∫ T 3 ∇ p ⋅ ∇ ϕ + 1 2 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ + ⟨ P V , ϕ ⟩ , \begin{aligned}
\int_{\mathbb{T}^{3}}\nabla u\cdot\nabla\phi=(u,\phi)_{\mathring{H}^{1}}=(\nabla_{\mathring{H}^{1}}E_{\mathrm{var}}(p),\phi)_{\mathring{H}^{1}}=DE_{\mathrm{var}}(p)\phi\\
&=\int_{\mathbb{T}^{3}}\nabla p\cdot\nabla\phi+\frac{1}{2}\min(M-\partial_{3}p,0)\,\partial_{3}\phi+\langle PV,\phi\rangle,
\end{aligned} ∫ T 3 ∇ u ⋅ ∇ ϕ = ( u , ϕ ) H ˚ 1 = ( ∇ H ˚ 1 E var ( p ) , ϕ ) H ˚ 1 = D E var ( p ) ϕ = ∫ T 3 ∇ p ⋅ ∇ ϕ + 2 1 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ + ⟨ P V , ϕ ⟩ ,
which is precisely (50) , the weak formulation of (49) . In order to establish (51) it suffices to integrate by parts the expression for the Gâteaux derivative of E v a r E_{\mathrm{var}} E var obtained in item 3 of Theorem 6.2 . This yields
( ∇ L 2 E v a r ( p ) , ϕ ) L 2 = D E v a r ( p ) ϕ = ∫ T 3 [ − Δ p − 1 2 ∂ 3 ( min ( M − ∂ 3 p , 0 ) ) + P V ] ϕ , (\nabla_{L^{2}}E_{\mathrm{var}}(p),\phi)_{L^{2}}=DE_{\mathrm{var}}(p)\phi=\int_{\mathbb{T}^{3}}\left[-\Delta p-\frac{1}{2}\partial_{3}\bigl(\min(M-\partial_{3}p,0)\bigr)+PV\right]\phi, ( ∇ L 2 E var ( p ) , ϕ ) L 2 = D E var ( p ) ϕ = ∫ T 3 [ − Δ p − 2 1 ∂ 3 ( min ( M − ∂ 3 p , 0 ) ) + P V ] ϕ ,
from which the claim follows.
Remark 6.6 (Comparing the L 2 L^{2} L 2 and H ˚ 1 \mathring{H}^{1} H ˚ 1 gradients). It is not a coincidence that in Lemma 6.5 above the L 2 L^{2} L 2 and H ˚ 1 \mathring{H}^{1} H ˚ 1 gradients are related via ∇ H ˚ 1 E v a r = ( − Δ ) − 1 ( ∇ L 2 E v a r ) \nabla_{\mathring{H}^{1}}E_{\mathrm{var}}=(-\Delta)^{-1}(\nabla_{L^{2}}E_{\mathrm{var}}) ∇ H ˚ 1 E var = ( − Δ ) − 1 ( ∇ L 2 E var ) . This identity actually follows from the following fact. For any Hilbert space H ⊂ L 2 H\subset L^{2} H ⊂ L 2 and any Fréchet differentiable F : L 2 → R F:L^{2}\to\mathbb{R} F : L 2 → R , if we denote by R H : R ∗ → R ∗ R_{H}:R^{*}\to R^{*} R H : R ∗ → R ∗ the Riesz representation map on H H H then ∇ H F = R H ( ∇ L 2 F ) \nabla_{H}F=R_{H}(\nabla_{L^{2}}F) ∇ H F = R H ( ∇ L 2 F ) . Indeed, for any u , ϕ ∈ H u,\phi\in H u , ϕ ∈ H : D F ( u ) ϕ = ( ∇ L 2 F ( u ) , ϕ ) L 2 = ⟨ ∇ L 2 F ( u ) , ϕ ⟩ H ∗ × H = ( R H ( ∇ L 2 F ) ( u ) , ϕ ) H DF(u)\phi=(\nabla_{L^{2}}F(u),\phi)_{L^{2}}=\langle\nabla_{L^{2}}F(u),\phi\rangle_{H^{*}\times H}=(R_{H}(\nabla_{L^{2}}F)(u),\phi)_{H} D F ( u ) ϕ = ( ∇ L 2 F ( u ) , ϕ ) L 2 = ⟨ ∇ L 2 F ( u ) , ϕ ⟩ H ∗ × H = ( R H ( ∇ L 2 F ) ( u ) , ϕ ) H .
In our case we have that R H ˚ 1 = ( − Δ ) − 1 R_{\mathring{H}^{1}}=(-\Delta)^{-1} R H ˚ 1 = ( − Δ ) − 1 in the sense that, for any f ∈ H − 1 f\in H^{-1} f ∈ H − 1 , u = R H ˚ 1 F u=R_{\mathring{H}^{1}}F u = R H ˚ 1 F if and only if u u u is the H ˚ 1 \mathring{H}^{1} H ˚ 1 -weak solution of − Δ u = f -\Delta u=f − Δ u = f in T 3 \mathbb{T}^{3} T 3 . The general fact above thus tells us that ∇ H ˚ 1 E v a r = R H ˚ 1 ( ∇ L 2 E v a r ) = ( − Δ ) − 1 ( ∇ L 2 E v a r ) \nabla_{\mathring{H}^{1}}E_{\mathrm{var}}=R_{\mathring{H}^{1}}(\nabla_{L^{2}}E_{\mathrm{var}})=(-\Delta)^{-1}(\nabla_{L^{2}}E_{\mathrm{var}}) ∇ H ˚ 1 E var = R H ˚ 1 ( ∇ L 2 E var ) = ( − Δ ) − 1 ( ∇ L 2 E var ) , as observed in Lemma 6.5 .
Besides gradient descent methods using the H ˚ 1 \mathring{H}^{1} H ˚ 1 and L 2 L^{2} L 2 gradients, the third descent method we will discuss is Newton descent . This comes from using, locally, the inner product generated by the Hessian of the variational energy to define a gradient direction. This idea is now made precise.
Definition 6.7 (Newton descent). Let F : H → R F:H\to\mathbb{R} F : H → R be a twice Fréchet differentiable map over a Hilbert space H H H . Suppose there exists C > 0 C>0 C > 0 such that, for every u ∈ H u\in H u ∈ H , D u 2 F ( ⋅ , ⋅ ) D^{2}_{u}F(\cdot,\cdot) D u 2 F ( ⋅ , ⋅ ) induces an inner product, written ( ⋅ , ⋅ ) D u 2 F (\cdot,\cdot)_{D^{2}_{u}F} ( ⋅ , ⋅ ) D u 2 F and called the Hessian inner product , such that 1 C ∥ ϕ ∥ H ≤ ∥ ϕ ∥ D u 2 F ≤ C ∥ ϕ ∥ H \frac{1}{C}\|\phi\|_{H}\leq\|\phi\|_{D^{2}_{u}F}\leq C\|\phi\|_{H} C 1 ∥ ϕ ∥ H ≤ ∥ ϕ ∥ D u 2 F ≤ C ∥ ϕ ∥ H for every ϕ ∈ H \phi\in H ϕ ∈ H (i.e. ∥ ⋅ ∥ D u 2 F \|\cdot\|_{D^{2}_{u}F} ∥ ⋅ ∥ D u 2 F is uniformly comparable to ∥ ⋅ ∥ H \|\cdot\|_{H} ∥ ⋅ ∥ H ). The negative of the gradient of F F F with respect to the Hessian inner product is called the Newton descent direction .
Having defined Newton descent we now record its form for our variational energy of interest.
Proposition 6.8 (Newton descent for our variational energy). Let E v a r E_{\mathrm{var}} E var be the variational energy introduced in Theorem 6.1 . The Hessian inner product induced by E v a r E_{\mathrm{var}} E var is, for any p ∈ H ˚ 1 p\in\mathring{H}^{1} p ∈ H ˚ 1 and M ∈ L 2 M\in L^{2} M ∈ L 2 ,
( ϕ , ψ ) D p 2 E v a r = ∫ T 3 ∇ ϕ ⋅ ∇ ψ − 1 2 1 ( M < ∂ 3 p ) ( ∂ 3 ϕ ) ( ∂ 3 ψ ) (52) (\phi,\psi)_{D^{2}_{p}E_{\mathrm{var}}}=\int_{\mathbb{T}^{3}}\nabla\phi\cdot\nabla\psi-\frac{1}{2}\mathbb{1}(M<\partial_{3}p)(\partial_{3}\phi)(\partial_{3}\psi) \tag*{(52)} ( ϕ , ψ ) D p 2 E var = ∫ T 3 ∇ ϕ ⋅ ∇ ψ − 2 1 1 ( M < ∂ 3 p ) ( ∂ 3 ϕ ) ( ∂ 3 ψ ) (52)
and so the Newton descent direction π \pi π is the H ˚ 1 \mathring{H}^{1} H ˚ 1 -weak solution of
− ∇ ⋅ [ ( I − 1 2 1 ( M < ∂ 3 p ) e 3 ⊗ e 3 ) ∇ π ] = Δ p + 1 2 ∂ 3 ( min ( M − ∂ 3 p , 0 ) ) − P V , (53) -\nabla\cdot\left[\left(I-\frac{1}{2}\mathbb{1}(M<\partial_{3}p)e_{3}\otimes e_{3}\right)\nabla\pi\right]=\Delta p+\frac{1}{2}\partial_{3}\bigl(\min(M-\partial_{3}p,0)\bigr)-PV, \tag*{(53)} − ∇ ⋅ [ ( I − 2 1 1 ( M < ∂ 3 p ) e 3 ⊗ e 3 ) ∇ π ] = Δ p + 2 1 ∂ 3 ( min ( M − ∂ 3 p , 0 ) ) − P V , (53)
for P V ∈ H − 1 PV\in H^{-1} P V ∈ H − 1 , i.e., for all ϕ ∈ H ˚ 1 \phi\in\mathring{H}^{1} ϕ ∈ H ˚ 1 ,
∫ T 3 ∇ π ⋅ ∇ ϕ − 1 2 1 ( M < ∂ 3 p ) ( ∂ 3 π ) ∂ 3 ϕ = ∫ T 3 − ∇ p ⋅ ∇ ϕ − 1 2 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ − ⟨ P V , ϕ ⟩ . (54) \begin{aligned}
\int_{\mathbb{T}^{3}}\nabla\pi\cdot\nabla\phi-\frac{1}{2}\mathbb{1}(M<\partial_{3}p)(\partial_{3}\pi)\partial_{3}\phi\\
&=\int_{\mathbb{T}^{3}}-\nabla p\cdot\nabla\phi-\frac{1}{2}\min(M-\partial_{3}p,0)\,\partial_{3}\phi-\langle PV,\phi\rangle. \tag*{(54)}
\end{aligned} ∫ T 3 ∇ π ⋅ ∇ ϕ − 2 1 1 ( M < ∂ 3 p ) ( ∂ 3 π ) ∂ 3 ϕ = ∫ T 3 − ∇ p ⋅ ∇ ϕ − 2 1 min ( M − ∂ 3 p , 0 ) ∂ 3 ϕ − ⟨ P V , ϕ ⟩ . (54)
Proof. Note that ( ϕ , ψ ) D 2 E v a r = D 2 E v a r ( ϕ , ψ ) (\phi,\psi)_{D^2 E_{\mathrm{var}}}=D^2E_{\mathrm{var}}(\phi,\psi) ( ϕ , ψ ) D 2 E var = D 2 E var ( ϕ , ψ ) and so (52) follows from differentiating the expression for D E v a r DE_{\mathrm{var}} D E var recorded in item 1 of Theorem 6.2 , since min ( ⋅ , 0 ) \min(\cdot,0) min ( ⋅ , 0 ) is Lipschitz with derivative min ( ⋅ , 0 ) ′ = 1 ( ⋅ < 0 ) \min(\cdot,0)'=\mathbb{1}(\cdot<0) min ( ⋅ , 0 ) ′ = 1 ( ⋅ < 0 ) . Since the Newton descent direction π \pi π is characterised by the fact that ( π , ϕ ) D 2 E v a r = − D E v a r ( p ) ϕ (\pi,\phi)_{D^2 E_{\mathrm{var}}}=-DE_{\mathrm{var}}(p)\phi ( π , ϕ ) D 2 E var = − D E var ( p ) ϕ for every ϕ ∈ H ˚ 1 \phi\in\mathring{H}^1 ϕ ∈ H ˚ 1 , (54) then follows immediately from (52) and the expression for D E v a r DE_{\mathrm{var}} D E var of item 3 of Theorem 6.2 . To see that (53) is indeed the strong form of (54) we integrate by parts.
− ∇ ⋅ ( A H k ∇ π ) = ∇ ⋅ ( A H k ∇ p k ) + 1 2 ∂ 3 ( H k M ) − P V -\nabla\cdot(A_{H_k}\nabla\pi)=\nabla\cdot(A_{H_k}\nabla p_k)+\frac{1}{2}\partial_3(H_kM)-PV − ∇ ⋅ ( A H k ∇ π ) = ∇ ⋅ ( A H k ∇ p k ) + 2 1 ∂ 3 ( H k M ) − P V
We may rearrange this equation and deduce that the new iterate p k + 1 = p k + π p_{k+1}=p_k+\pi p k + 1 = p k + π solves
∇ ⋅ ( A H k ∇ p k + 1 ) = P V − 1 2 ∂ 3 ( H k M ) . \nabla\cdot(A_{H_k}\nabla p_{k+1})=PV-\frac{1}{2}\partial_3(H_kM). ∇ ⋅ ( A H k ∇ p k + 1 ) = P V − 2 1 ∂ 3 ( H k M ) .
In other words: the next iterate p k + 1 p_{k+1} p k + 1 is obtained by solving the linearised P V PV P V -and-M M M inversion where the Heaviside characterising the linearisation comes from the previous step. Crucially: this is precisely the method that was used in previous works [48 , 96 ] to numerically solve the nonlinear P V PV P V -and-M M M inversion. This means that the convergence rates presented here for Newton descent are applicable to previous works. Moreover this means that the convergence guarantees for Newton descent are also the first convergence guarantees of this numerical method which is already used in practice.
The remainder of this section is devoted to the obtention of convergence rates for each of the three descent methods discussed above. There are two key assumptions that must be verified about our variational energy in order for such convergence rates to be obtained. The first comes from convexity and takes the form of a quadratic lower bound on the variational energy. The second, which takes the form of a quadratic upper bound on the variational energy, is that to which we turn our attention now. The basic ideas are that (1) we will introduce a simpler energy density for which, due to its piecewise quadratic nature, we can easily find such a quadratic upper bound, and (2) we will show that this simpler energy density can be turned into our true variational energy density up to some affine transformations. The result will then follow from the fact that such affine transformations will only interact with these quadratic upper bounds in predictable ways. In order to carry out this argument it is particularly convenient to introduce the notion of a quadratic remainder , which we define below.
Definition 6.10 (Quadratic remainder). Consider a continuously differentiable function f : R d → R f:\mathbb{R}^d\to\mathbb{R} f : R d → R . For any x ∈ R d x\in\mathbb{R}^d x ∈ R d we call function T x f : R d → R T_x^f:\mathbb{R}^d\to\mathbb{R} T x f : R d → R defined as T x f ( y ) : = f ( y ) − f ( x ) − D f ( x ) ( y − x ) T_x^f(y):=f(y)-f(x)-Df(x)(y-x) T x f ( y ) := f ( y ) − f ( x ) − D f ( x ) ( y − x ) for every y ∈ R d y\in\mathbb{R}^d y ∈ R d the quadratic remainder of f f f at x x x .
We now show that the aforementioned “simpler” energy density – introduced in the result below – does indeed have a quadratic upper bound.
Lemma 6.11 (Quadratic upper bound for a piecewise quadratic function). Consider f : R 3 → R f:\mathbb{R}^3\to\mathbb{R} f : R 3 → R defined by f ( x ) = 1 2 ∣ x ∣ 2 − 1 4 min ( x 3 , 0 ) 2 f(x)=\frac{1}{2}|x|^2-\frac{1}{4}\min(x_3,0)^2 f ( x ) = 2 1 ∣ x ∣ 2 − 4 1 min ( x 3 , 0 ) 2 for all x ∈ R 3 x\in\mathbb{R}^3 x ∈ R 3 . For every x ∈ R 3 x\in\mathbb{R}^3 x ∈ R 3 , f ( y ) ≤ f ( x ) + D f ( x ) ( y − x ) + 1 2 ∣ y − x ∣ 2 f(y)\le f(x)+Df(x)(y-x)+\frac{1}{2}|y-x|^2 f ( y ) ≤ f ( x ) + D f ( x ) ( y − x ) + 2 1 ∣ y − x ∣ 2 .
Proof. To make the computations easier we introduce the following function: A ( x ) : = I − 1 2 1 ( x 3 < 0 ) e 3 ⊗ e 3 A(x):=I-\frac{1}{2}\mathbb{1}(x_3<0)e_3\otimes e_3 A ( x ) := I − 2 1 1 ( x 3 < 0 ) e 3 ⊗ e 3 , such that f ( x ) = 1 2 A ( x ) x ⋅ x f(x)=\frac{1}{2}A(x)x\cdot x f ( x ) = 2 1 A ( x ) x ⋅ x , and hence D f ( x ) = A ( x ) x Df(x)=A(x)x D f ( x ) = A ( x ) x . In particular, for every x x x , the largest eigenvalue of A ( x ) A(x) A ( x ) is equal to 1, and so A ( x ) ξ ⋅ ξ ≤ ∣ ξ ∣ 2 A(x)\xi\cdot\xi\le|\xi|^2 A ( x ) ξ ⋅ ξ ≤ ∣ ξ ∣ 2 for all ξ ∈ R 3 \xi\in\mathbb{R}^3 ξ ∈ R 3 .
In terms of the quadratic remainder T x f T_x^f T x f , we may write the target inequality as T x f ( y ) ≤ 1 2 ∣ y − x ∣ 2 T_x^f(y)\le\frac{1}{2}|y-x|^2 T x f ( y ) ≤ 2 1 ∣ y − x ∣ 2 , so we begin by using the fact that A A A is pointwise symmetric to rewrite T x f T_x^f T x f in a more convenient way. We obtain
T x f ( y ) = 1 2 [ A ( y ) − A ( x ) ] y ⋅ y + 1 2 A ( x ) ( y − x ) ⋅ ( y − x ) . T_x^f(y)=\frac{1}{2}[A(y)-A(x)]y\cdot y+\frac{1}{2}A(x)(y-x)\cdot(y-x). T x f ( y ) = 2 1 [ A ( y ) − A ( x )] y ⋅ y + 2 1 A ( x ) ( y − x ) ⋅ ( y − x ) .
Since A ( y ) − A ( x ) = 1 2 [ 1 ( x 3 < 0 ≤ y 3 ) − 1 ( y 3 < 0 ≤ x 3 ) ] e 3 ⊗ e 3 A(y)-A(x)=\frac{1}{2}\left[\mathbb{1}(x_{3}<0\le y_{3})-\mathbb{1}(y_{3}<0\le x_{3})\right]e_{3}\otimes e_{3} A ( y ) − A ( x ) = 2 1 [ 1 ( x 3 < 0 ≤ y 3 ) − 1 ( y 3 < 0 ≤ x 3 ) ] e 3 ⊗ e 3 we compute that
T x f ( y ) = { − y 3 2 4 + 1 2 A ( x ) ( y − x ) ⋅ ( y − x ) if y 3 < 0 ≤ x 3 , x 3 2 4 + 1 2 A ( x ) ( y − x ) ⋅ ( y − x ) if x 3 < 0 ≤ y 3 , and 1 2 A ( x ) ( y − x ) ⋅ ( y − x ) otherwise. T_{x}^{f}(y)=
\begin{cases}
-\dfrac{y_{3}^{2}}{4}+\dfrac{1}{2}A(x)(y-x)\cdot(y-x) & \text{if } y_{3}<0\le x_{3},\\
\dfrac{x_{3}^{2}}{4}+\dfrac{1}{2}A(x)(y-x)\cdot(y-x) & \text{if } x_{3}<0\le y_{3},\text{ and}\\
\dfrac{1}{2}A(x)(y-x)\cdot(y-x) & \text{otherwise.}
\end{cases} T x f ( y ) = ⎩ ⎨ ⎧ − 4 y 3 2 + 2 1 A ( x ) ( y − x ) ⋅ ( y − x ) 4 x 3 2 + 2 1 A ( x ) ( y − x ) ⋅ ( y − x ) 2 1 A ( x ) ( y − x ) ⋅ ( y − x ) if y 3 < 0 ≤ x 3 , if x 3 < 0 ≤ y 3 , and otherwise.
Only one of these three cases is troublesome since, if y 3 < 0 ≤ x 3 y_{3}<0\le x_{3} y 3 < 0 ≤ x 3 or if x 3 x_{3} x 3 and y 3 y_{3} y 3 have the same sign then T x f ( y ) ≤ 1 2 A ( x ) ( y − x ) ⋅ ( y − x ) ≤ 1 2 ∣ y − x ∣ 2 T_{x}^{f}(y)\le\frac{1}{2}A(x)(y-x)\cdot(y-x)\le\frac{1}{2}|y-x|^{2} T x f ( y ) ≤ 2 1 A ( x ) ( y − x ) ⋅ ( y − x ) ≤ 2 1 ∣ y − x ∣ 2 . In the troublesome case, i.e. if x 3 < 0 ≤ y 3 x_{3}<0\le y_{3} x 3 < 0 ≤ y 3 , then we need to be more careful. In this case writing z h = ( z 1 , z 2 ) z_{h}=(z_{1},z_{2}) z h = ( z 1 , z 2 ) for any x ∈ R 3 x\in\mathbb{R}^{3} x ∈ R 3 tells us that
T x f ( y ) = 1 2 A ( x ) ( y − x ) ⋅ ( y − x ) + 1 4 x 3 2 = 1 2 ∣ x h − y h ∣ 2 + 1 4 ( x 3 − y 3 ) 2 + 1 4 x 3 2 . T_{x}^{f}(y)=\frac{1}{2}A(x)(y-x)\cdot(y-x)+\frac{1}{4}x_{3}^{2}=\frac{1}{2}|x_{h}-y_{h}|^{2}+\frac{1}{4}(x_{3}-y_{3})^{2}+\frac{1}{4}x_{3}^{2}. T x f ( y ) = 2 1 A ( x ) ( y − x ) ⋅ ( y − x ) + 4 1 x 3 2 = 2 1 ∣ x h − y h ∣ 2 + 4 1 ( x 3 − y 3 ) 2 + 4 1 x 3 2 .
Crucially: if x 3 < 0 ≤ y 3 x_{3}<0\le y_{3} x 3 < 0 ≤ y 3 then x 3 y 3 ≥ 0 x_{3}y_{3}\ge0 x 3 y 3 ≥ 0 and hence 1 4 ( x 3 − y 3 ) 2 + 1 4 x 3 2 ≤ 1 2 ( x 3 − y 3 ) 2 \frac{1}{4}(x_{3}-y_{3})^{2}+\frac{1}{4}x_{3}^{2}\le\frac{1}{2}(x_{3}-y_{3})^{2} 4 1 ( x 3 − y 3 ) 2 + 4 1 x 3 2 ≤ 2 1 ( x 3 − y 3 ) 2 , which proves that T x f ( y ) ≤ 1 2 ∣ x − y ∣ 2 T_{x}^{f}(y)\le\frac{1}{2}|x-y|^{2} T x f ( y ) ≤ 2 1 ∣ x − y ∣ 2 even in this case, concluding the proof.
With Lemma 6.11 in hand we may now use the fact that quadratic upper bounds on quadratic remainders behave nicely under transformations (see Lemmas A.16 –A.18 ) to deduce that the true variational energy density of interest also has a quadratic upper bound.
Lemma 6.12 (Quadratic upper bound for our variational energy density). Consider e : R 3 → R e:\mathbb{R}^{3}\to\mathbb{R} e : R 3 → R defined by e ( x ) : = 1 2 ∣ x ∣ 2 − 1 4 min ( r − x 3 , 0 ) 2 e(x):=\frac{1}{2}|x|^{2}-\frac{1}{4}\min(r-x_{3},0)^{2} e ( x ) := 2 1 ∣ x ∣ 2 − 4 1 min ( r − x 3 , 0 ) 2 , where r r r is fixed. For every x , y ∈ R 3 x,y\in\mathbb{R}^{3} x , y ∈ R 3 , e ( y ) ≤ e ( x ) + D e ( x ) ( y − x ) + 1 2 ∣ x − y ∣ 2 e(y)\le e(x)+De(x)(y-x)+\frac{1}{2}|x-y|^{2} e ( y ) ≤ e ( x ) + D e ( x ) ( y − x ) + 2 1 ∣ x − y ∣ 2 .
Proof. We note that, for f f f as in Lemma 6.11 , f ( r u 3 − x ) = e ( x ) + a ( x ) f(ru_{3}-x)=e(x)+a(x) f ( r u 3 − x ) = e ( x ) + a ( x ) for an affine function a ( x ) = − r x 3 + 1 2 r 2 a(x)=-rx_{3}+\frac{1}{2}r^{2} a ( x ) = − r x 3 + 2 1 r 2 . Introducing the affine transformation Φ ( x ) = r e 3 − x \Phi(x)=re_{3}-x Φ ( x ) = r e 3 − x this means that e = f ∘ Φ − a e=f\circ\Phi-a e = f ∘ Φ − a . Therefore we may use Lemmas A.16 , A.17 , and A.18 to see that the quadratic remainder of e e e is given by T x e = T x f ∘ Φ − T x a = T Φ ( x ) f ∘ Φ T_{x}^{e}=T_{x}^{f\circ\Phi}-T_{x}^{a}=T_{\Phi(x)}^{f}\circ\Phi T x e = T x f ∘ Φ − T x a = T Φ ( x ) f ∘ Φ . In particular we deduce from Lemma 6.11 that T x e ( y ) ≤ 1 2 ∣ Φ ( y ) − Φ ( x ) ∣ 2 = 1 2 ∣ y − x ∣ 2 T_{x}^{e}(y)\le\frac{1}{2}|\Phi(y)-\Phi(x)|^{2}=\frac{1}{2}|y-x|^{2} T x e ( y ) ≤ 2 1 ∣Φ ( y ) − Φ ( x ) ∣ 2 = 2 1 ∣ y − x ∣ 2 , which is precisely the desired inequality.
Our search for a quadratic upper bound on the variational energy now concludes as we leverage the quadratic upper bound on the variational energy density obtained in Lemma 6.12 to deduce that the variational energy itself does indeed have a comparable quadratic upper bound.
Lemma 6.13 (Quadratic upper model for our variational energy). Let E v a r E_{\mathrm{var}} E var be the variational energy introduced in Theorem 6.1 . For every p , r ∈ H ˚ 1 p,r\in\mathring{H}^{1} p , r ∈ H ˚ 1 , the following inequality holds: E v a r ( r ) ≤ E v a r ( p ) + D E v a r ( p ) ( r − p ) + 1 2 ∥ r − p ∥ H ˚ 1 2 E_{\mathrm{var}}(r)\le E_{\mathrm{var}}(p)+DE_{\mathrm{var}}(p)(r-p)+\frac{1}{2}\|r-p\|_{\mathring{H}^{1}}^{2} E var ( r ) ≤ E var ( p ) + D E var ( p ) ( r − p ) + 2 1 ∥ r − p ∥ H ˚ 1 2 .
Proof. We write E v a r E_{\mathrm{var}} E var as in (47) , abusively writing e M ( ∇ p ) e_{M}(\nabla p) e M ( ∇ p ) to mean e M ( x , ∇ p ( x ) ) e_{M}(x,\nabla p(x)) e M ( x , ∇ p ( x )) in the sequel, and note that item 3 of Theorem 6.2 tells us that
D E v a r ( p ) ϕ = ∫ T 3 ∇ u e M ( ∇ p ) ⋅ ∇ ϕ + ⟨ P V , ϕ ⟩ . (55) DE_{\mathrm{var}}(p)\phi=\int_{\mathbb{T}^{3}}\nabla_{u}e_{M}(\nabla p)\cdot\nabla\phi+\langle PV,\phi\rangle. \tag*{(55)} D E var ( p ) ϕ = ∫ T 3 ∇ u e M ( ∇ p ) ⋅ ∇ ϕ + ⟨ P V , ϕ ⟩ . (55)
Similarly we may reinterpret Lemma 6.12 as telling us that
e M ( ∇ r ) ≤ e M ( ∇ p ) + ∇ u e M ( ∇ p ) ⋅ ∇ ( r − p ) + 1 2 ∣ ∇ ( r − p ) ∣ 2 . (56) e_{M}(\nabla r)\le e_{M}(\nabla p)+\nabla_{u}e_{M}(\nabla p)\cdot\nabla(r-p)+\frac{1}{2}|\nabla(r-p)|^{2}. \tag*{(56)} e M ( ∇ r ) ≤ e M ( ∇ p ) + ∇ u e M ( ∇ p ) ⋅ ∇ ( r − p ) + 2 1 ∣∇ ( r − p ) ∣ 2 . (56)
Putting (47) , (55) , and (56) together then yields the claim.
We now have all of the ingredients we need to prove the convergence rates for our three descent methods in the context of exact line search. First we state a generic version of that result in Proposition 6.14 before specialising it to our energy afterwards in Theorem 6.16 . Note that the result below is classical – see for example [15 ] . We record it here, along with its proof, in order to ensure that good care is taken of tracking the constants arising from comparable norms involved in the descent method. This is particularly important for us since ultimately it will lead to significant differences between the convergence rate of the L 2 L^{2} L 2 -gradient descent compared with the other descent methods.
Proposition 6.14 (Key inequality for the convergence rate of gradient descent with exact line search in Hilbert spaces). Let ( ⋅ , ⋅ ) X (\cdot,\cdot)_X ( ⋅ , ⋅ ) X be a complete inner product on a Hilbert space H H H for which there exists a constant C > 0 C>0 C > 0 such that
∥ u ∥ H ≤ C ∥ u ∥ X for every u ∈ H . (57) \|u\|_H \le C\|u\|_X \quad\text{for every }u\in H. \tag*{(57)} ∥ u ∥ H ≤ C ∥ u ∥ X for every u ∈ H . (57) Let F : H → R \mathcal{F}:H\to\mathbb{R} F : H → R be a continuously Fréchet differentiable functional satisfying, for some constants m , M > 0 m,M>0 m , M > 0 ,
m 2 ∥ v − u ∥ H 2 ≤ F ( v ) − F ( u ) − D F ( u ) ( v − u ) ≤ M 2 ∥ v − u ∥ H 2 (58) \frac{m}{2}\|v-u\|_H^2 \le\mathcal{F}(v)-\mathcal{F}(u)-D\mathcal{F}(u)(v-u)\le\frac{M}{2}\|v-u\|_H^2 \tag*{(58)} 2 m ∥ v − u ∥ H 2 ≤ F ( v ) − F ( u ) − D F ( u ) ( v − u ) ≤ 2 M ∥ v − u ∥ H 2 (58) for every u , v ∈ H u,v\in H u , v ∈ H , and let u ∗ u^* u ∗ denote the unique minimizer of F \mathcal{F} F . Then
F ( u − 1 C 2 M ∇ X F ( u ) ) ≤ F ( u ) − 1 2 C 2 M ∥ ∇ X F ( u ) ∥ X 2 (59) \mathcal{F}\left(u-\frac{1}{C^2M}\nabla_X\mathcal{F}(u)\right)\le\mathcal{F}(u)-\frac{1}{2C^2M}\|\nabla_X\mathcal{F}(u)\|_X^2 \tag*{(59)} F ( u − C 2 M 1 ∇ X F ( u ) ) ≤ F ( u ) − 2 C 2 M 1 ∥ ∇ X F ( u ) ∥ X 2 (59) and
F ( u ) − F ( u ∗ ) ≤ 1 2 m ∥ ∇ H F ( u ) ∥ H 2 (60) \mathcal{F}(u)-\mathcal{F}(u^*)\le\frac{1}{2m}\|\nabla_H\mathcal{F}(u)\|_H^2 \tag*{(60)} F ( u ) − F ( u ∗ ) ≤ 2 m 1 ∥ ∇ H F ( u ) ∥ H 2 (60) such that
F ( u − 1 C 2 M ∇ X F ( u ) ) − F ( u ∗ ) ≤ ( 1 − m C 4 M ) ( F ( u ) − F ( u ∗ ) ) . (61) \mathcal{F}\left(u-\frac{1}{C^2M}\nabla_X\mathcal{F}(u)\right)-\mathcal{F}(u^*)\le\left(1-\frac{m}{C^4M}\right)\left(\mathcal{F}(u)-\mathcal{F}(u^*)\right). \tag*{(61)} F ( u − C 2 M 1 ∇ X F ( u ) ) − F ( u ∗ ) ≤ ( 1 − C 4 M m ) ( F ( u ) − F ( u ∗ ) ) . (61)
Remark 6.15 (On the importance of the key inequality). The so-called “key inequality” (61) warrants such a lofty name since the actual convergence rate follows from that inequality immediately. Indeed, if we denote let x 0 ∈ T 3 x_0\in T^3 x 0 ∈ T 3 and denote by x k x_k x k the iterates obtained via gradient descent with exact line search, i.e.
x k : = arg min t ∈ R F ( x k − 1 − t ∇ F ( x k − 1 ) ) , x_k:=\underset{t\in\mathbb{R}}{\arg\min}\mathcal{F}(x_{k-1}-t\nabla\mathcal{F}(x_{k-1})), x k := t ∈ R arg min F ( x k − 1 − t ∇ F ( x k − 1 )) ,
then inducting on the key inequality tells us that
F ( x k − 1 ) − F ( x ∗ ) ≤ ( 1 − m C 4 M ) k ( F ( x 0 ) − F ( x ∗ ) ) , \mathcal{F}(x_{k-1})-\mathcal{F}(x^*)\le\left(1-\frac{m}{C^4M}\right)^k\left(\mathcal{F}(x_0)-\mathcal{F}(x^*)\right), F ( x k − 1 ) − F ( x ∗ ) ≤ ( 1 − C 4 M m ) k ( F ( x 0 ) − F ( x ∗ ) ) ,
which guarantees that indeed F ( x k ) \mathcal{F}(x_k) F ( x k ) converges to F ( x 0 ) \mathcal{F}(x_0) F ( x 0 ) .
Proof of Proposition 6.14. Plugging v = u − t ∇ X F ( u ) v=u-t\nabla_X\mathcal{F}(u) v = u − t ∇ X F ( u ) into the second inequality in (58) tells us that
F ( u − t ∇ X F ( u ) ) ≤ F ( u ) − t ∥ ∇ X F ( u ) ∥ X 2 + M t 2 2 ∥ ∇ X F ( u ) ∥ H 2 ≤ F ( u ) + ( C 2 M t 2 2 − t ) ∥ ∇ X F ( u ) ∥ X 2 . \begin{aligned}
\mathcal{F}(u-t\nabla_X\mathcal{F}(u))\le\mathcal{F}(u)-t\|\nabla_X\mathcal{F}(u)\|_X^2+\frac{Mt^2}{2}\|\nabla_X\mathcal{F}(u)\|_H^2 \\
&\le\mathcal{F}(u)+\left(\frac{C^2Mt^2}{2}-t\right)\|\nabla_X\mathcal{F}(u)\|_X^2.
\end{aligned} F ( u − t ∇ X F ( u )) ≤ F ( u ) − t ∥ ∇ X F ( u ) ∥ X 2 + 2 M t 2 ∥ ∇ X F ( u ) ∥ H 2 ≤ F ( u ) + ( 2 C 2 M t 2 − t ) ∥ ∇ X F ( u ) ∥ X 2 .
Since the quadratic q ( t ) : = C 2 M t 2 2 − t q(t):=\frac{C^2Mt^2}{2}-t q ( t ) := 2 C 2 M t 2 − t is minimized at t ∗ = 1 C 2 M t^*=\frac{1}{C^2M} t ∗ = C 2 M 1 , where it attains the value q ( t ∗ ) = − 1 2 C 2 M q(t^*)=-\frac{1}{2C^2M} q ( t ∗ ) = − 2 C 2 M 1 , we deduce (59) from the inequality above.
To establish (60) we use the first inequality in (58) . We begin by writing D F ( u ) ( v − u ) = ( ∇ H F ( u ) , v − u ) H D\mathcal{F}(u)(v-u)=(\nabla_H\mathcal{F}(u),v-u)_H D F ( u ) ( v − u ) = ( ∇ H F ( u ) , v − u ) H , which tells us that with respect to v v v the right-hand side of that inequality is minimized at v ∗ = u − 1 m ∇ H F ( u ) v^*=u-\frac{1}{m}\nabla_H\mathcal{F}(u) v ∗ = u − m 1 ∇ H F ( u ) . Therefore, for any u , v ∈ H u,v\in H u , v ∈ H ,
F ( v ) ≥ [ F ( u ) + ( ∇ H F ( u ) , v − u ) H + m 2 ∥ v − u ∥ H 2 ] v = v ∗ = F ( u ) − 1 2 m ∥ ∇ H F ( u ) ∥ H 2 . \mathcal{F}(v)\ge\left[\mathcal{F}(u)+(\nabla_H\mathcal{F}(u),v-u)_H+\frac{m}{2}\|v-u\|_H^2\right]_{v=v^*}=\mathcal{F}(u)-\frac{1}{2m}\|\nabla_H\mathcal{F}(u)\|_H^2. F ( v ) ≥ [ F ( u ) + ( ∇ H F ( u ) , v − u ) H + 2 m ∥ v − u ∥ H 2 ] v = v ∗ = F ( u ) − 2 m 1 ∥ ∇ H F ( u ) ∥ H 2 .
Setting v = u ∗ v=u^* v = u ∗ produces (60) .
Finally we turn our attention to the key inequality (61) . (57) and (60) tell us that ∥ ∇ X F ( u ) ∥ X 2 ≥ 2 m C 2 ( F ( u ) − F ( u ∗ ) ) \|\nabla_X\mathcal{F}(u)\|_X^2\ge\frac{2m}{C^2}(\mathcal{F}(u)-\mathcal{F}(u^*)) ∥ ∇ X F ( u ) ∥ X 2 ≥ C 2 2 m ( F ( u ) − F ( u ∗ )) and so combining this inequality with (59) yields (61) .
We now specialise the generic result proved above to the case of our variational energy.
Theorem 6.16 (Convergence rates for our variational energy with exact line search). Let E v a r E_{\mathrm{var}} E var be the variational energy introduced in Theorem 6.1 , let V V V be a non-empty subspace of H ˙ 1 \dot{H}^1 H ˙ 1 (possibly equal to H ˙ 1 \dot{H}^1 H ˙ 1 ), and let p ∗ p^* p ∗ be the unique minimizer of E v a r E_{\mathrm{var}} E var in V V V .
Let p ∈ V p\in V p ∈ V and let π = − ∇ H ˙ 1 E v a r ∣ V ( p ) \pi=-\nabla_{\dot{H}^1}E_{\mathrm{var}}|_V(p) π = − ∇ H ˙ 1 E var ∣ V ( p ) denote the H ˙ 1 \dot{H}^1 H ˙ 1 -gradient descent direction for E v a r ∣ V E_{\mathrm{var}}|_V E var ∣ V at p p p . Then
E v a r ( p + π ) − E v a r ( p ∗ ) ≤ 1 2 ( E v a r ( p ) − E v a r ( p ∗ ) ) . E_{\mathrm{var}}(p+\pi)-E_{\mathrm{var}}(p^*)\le\frac{1}{2}\left(E_{\mathrm{var}}(p)-E_{\mathrm{var}}(p^*)\right). E var ( p + π ) − E var ( p ∗ ) ≤ 2 1 ( E var ( p ) − E var ( p ∗ ) ) . Let p ∈ V p\in V p ∈ V and let π \pi π denote the Newton descent direction for E v a r ∣ V E_{\mathrm{var}}|_V E var ∣ V at p p p . Then E v a r ( p + 1 4 π ) − E v a r ( p ∗ ) ≤ 31 32 ( E v a r ( p ) − E v a r ( p ∗ ) ) E_{\mathrm{var}}(p+\frac{1}{4}\pi)-E_{\mathrm{var}}(p^*)\le\frac{31}{32}\left(E_{\mathrm{var}}(p)-E_{\mathrm{var}}(p^*)\right) E var ( p + 4 1 π ) − E var ( p ∗ ) ≤ 32 31 ( E var ( p ) − E var ( p ∗ ) ) .
Suppose that V ⊆ H 2 V \subseteq H^2 V ⊆ H 2 , suppose that there exist a constant C n > 0 C_n > 0 C n > 0 such that ∥ p ∥ H ˙ 1 ≤ C n ∥ p ∥ L 2 \|p\|_{\dot{H}^1} \le C_n\|p\|_{L^2} ∥ p ∥ H ˙ 1 ≤ C n ∥ p ∥ L 2 for every p ∈ V p \in V p ∈ V , and suppose that M ∈ H 1 M \in H^1 M ∈ H 1 and P V ∈ L 2 P V \in L^2 P V ∈ L 2 . Let p ∈ V p \in V p ∈ V and let π = − ∇ L 2 E v a r ∣ V ( p ) \pi= -\nabla_{L^2} E_{\mathrm{var}}|_V(p) π = − ∇ L 2 E var ∣ V ( p ) denote the L 2 L^2 L 2 -gradient descent direction for E v a r ∣ V E_{\mathrm{var}}|_V E var ∣ V at p p p . Then E v a r ( p + 1 C n 2 π ) − E v a r ( p ∗ ) ≤ ( 1 − 1 2 C n 4 ) ( E v a r ( p ) − E v a r ( p ∗ ) ) E_{\mathrm{var}}\left(p+\frac{1}{C_n^2}\pi\right)-E_{\mathrm{var}}(p^*) \le\left(1-\frac{1}{2C_n^4}\right)\left(E_{\mathrm{var}}(p)-E_{\mathrm{var}}(p^*)\right) E var ( p + C n 2 1 π ) − E var ( p ∗ ) ≤ ( 1 − 2 C n 4 1 ) ( E var ( p ) − E var ( p ∗ ) ) .
Proof. Since all of these descent directions are gradient descents, we need to verify that the hypotheses of Proposition 6.14 are verified. Item 4 of Theorem 6.2 guarantees that E v a r E_{\mathrm{var}} E var is continuously Fréchet differentiable, item 1 of Theorem 6.2 , Lemma 6.13 , and Lemma A.15 show that (58) is verified with m = 1 2 m=\frac{1}{2} m = 2 1 and M = 1 M=1 M = 1 , and Theorem 6.3 verifies that E v a r E_{\mathrm{var}} E var does indeed have a unique minimizer. The only hypothesis left to check is (57) , and the value of the constant C C C varies from case to case since the inner product ( ⋅ , ⋅ ) X (\cdot,\cdot)_X ( ⋅ , ⋅ ) X used to define the gradient differs in each case.
In this case ( ⋅ , ⋅ ) X = ( ⋅ , ⋅ ) H ˙ 1 (\cdot,\cdot)_X=(\cdot,\cdot)_{\dot{H}^1} ( ⋅ , ⋅ ) X = ( ⋅ , ⋅ ) H ˙ 1 and so C = 1 C=1 C = 1 .
In this case we may use the Hessian inner product recorded in Proposition 6.8 to see that ∥ φ ∥ D p 2 E v a r 2 = ∫ T 3 ∣ ∇ φ ∣ 2 − 1 2 1 ( M < ∂ 3 p ) ≥ 1 2 ∥ φ ∥ H ˙ 1 2 \|\varphi\|^2_{D^2_pE_{\mathrm{var}}}=\int_{\mathbb{T}^3}|\nabla\varphi|^2-\frac{1}{2}\mathbf{1}_{(M<\partial_3p)}\ge\frac{1}{2}\|\varphi\|^2_{\dot{H}^1} ∥ φ ∥ D p 2 E var 2 = ∫ T 3 ∣∇ φ ∣ 2 − 2 1 1 ( M < ∂ 3 p ) ≥ 2 1 ∥ φ ∥ H ˙ 1 2 , i.e. C = 2 C=2 C = 2 .
In this case C = C n C=C_n C = C n , by assumption.
Remark 6.17 (On the convergence rate of L 2 L^2 L 2 -gradient descent). We comment on the constant C n C_n C n which appears in the third item of Theorem 6.16 above. Note that in practice such a constant exists since we work with V V V a finite-dimensional subspace of H ˙ 1 \dot{H}^1 H ˙ 1 , in which case all norms on V V V are equivalent. Moreover, say that in such an instance the dimension of V V V is given by n n n (typically commensurate with the number of elements in a finite-element scheme). Then, as n n n approaches infinity, the constant C n C_n C n would necessarily grow unboundedly, which would in turn mean that 1 − 1 2 C n 4 ↑ 1 1-\frac{1}{2C_n^4}\uparrow1 1 − 2 C n 4 1 ↑ 1 . This means that in the case of L 2 L^2 L 2 -gradient descent the convergence rate remains linear but the constant involved worsens (i.e. approaches one) as the dimension of the subspace we work with increases (e.g. as the mesh is refined). Note that this remark also apply when backtracking is used instead of exact line search – c.f. Theorem 6.20 below.
We now turn our attention towards the second, and arguably more important, set of convergence rates recorded in this section: those arising from backtracking line search. For we define backtracking line search below. Recall that, by contrast with exact line search discussed so far, backtracking line search can be implemented immediately in practical situations (which is typically not the case for exact line search) while retaining enough favourable structure such that provable convergence rates are still in play with backtracking line search.
Definition 6.18 (Backtracking line search). Consider a functional F : B → R \mathcal{F}:B\to\mathbb{R} F : B → R on a Banach space B B B , let x ∈ B x\in B x ∈ B be a current iterate, let Δ x ∈ B \Delta x\in B Δ x ∈ B be a search direction, and consider two parameters α ∈ ( 0 , 1 / 2 ) \alpha\in(0,1/2) α ∈ ( 0 , 1/2 ) and β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) . Backtracking line search proceeds as follows: initiate the step size t 0 = 1 t_0=1 t 0 = 1 and, for i ≥ 0 i\ge0 i ≥ 0 , if
f ( x + t i Δ x ) > f ( x ) + α t i D f ( x ) Δ x f(x+t_i\Delta x)>f(x)+\alpha t_i Df(x)\Delta x f ( x + t i Δ x ) > f ( x ) + α t i D f ( x ) Δ x then set t i + 1 : = β t i t_{i+1}:=\beta t_i t i + 1 := β t i , while if
f ( x + t i Δ x ) ≤ f ( x ) + α t i D f ( x ) Δ x (62) f(x+t_i\Delta x)\le f(x)+\alpha t_i Df(x)\Delta x \tag*{(62)} f ( x + t i Δ x ) ≤ f ( x ) + α t i D f ( x ) Δ x (62) then set the step size t ∗ : = t i t_*:=t_i t ∗ := t i and terminate.
As we did above in the context of exact line search when specialising from the generic result of Proposition 6.14 to Theorem 6.16 which applies to the variational energy of interest, we first record a generic result for the convergence rate of gradient descent methods with backtracking line search. As with Proposition 6.14 above, such a result is classical (see [15 ] ) and its proof is provided here in order to carefully track the constants that appear due to the presence of comparable norms.
Proposition 6.19 (Key inequality for the convergence rate of gradient descent with exact line search in Hilbert spaces). Consider the same notation and assumptions as Proposition 6.14 and let α ∈ ( 0 , 1 / 2 ) \alpha\in(0,1/2) α ∈ ( 0 , 1/2 ) and β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) . For any u ∈ H u\in H u ∈ H and any t ∈ [ 0 , 1 C 2 M ] t\in\left[0,\frac{1}{C^2M}\right] t ∈ [ 0 , C 2 M 1 ] ,
F ( u − t ∇ X F ( u ) ) ≤ F ( u ) − α t ∥ ∇ X F ( u ) ∥ X 2 (63) \mathcal{F}(u-t\nabla_X\mathcal{F}(u))\le\mathcal{F}(u)-\alpha t\|\nabla_X\mathcal{F}(u)\|_X^2 \tag*{(63)} F ( u − t ∇ X F ( u )) ≤ F ( u ) − α t ∥ ∇ X F ( u ) ∥ X 2 (63) i.e. the backtracking exit condition (62) holds. Moreover, for any u ∈ H u\in H u ∈ H if t ∗ t_* t ∗ denotes the step size obtained via backtracking line search with parameters α \alpha α and β \beta β and search direction − ∇ X F ( u ) -\nabla_X\mathcal{F}(u) − ∇ X F ( u ) then
t ∗ ≥ min ( 1 , β C 2 M ) (64) t_* \ge\min\left(1,\frac{\beta}{C^2M}\right) \tag*{(64)} t ∗ ≥ min ( 1 , C 2 M β ) (64)
and
F ( u − t ∗ ∇ X F ( u ) ) − F ( u ∗ ) ≤ ( 1 − 2 α m C 2 min ( 1 , β C 2 M ) ) ( F ( u ) − F ( u ∗ ) ) . \mathcal{F}\left(u-t_{*}\nabla_X\mathcal{F}(u)\right)-\mathcal{F}(u^*)\leq\left(1-\frac{2\alpha m}{C^2}\min\left(1,\frac{\beta}{C^2M}\right)\right)\left(\mathcal{F}(u)-\mathcal{F}(u^*)\right). F ( u − t ∗ ∇ X F ( u ) ) − F ( u ∗ ) ≤ ( 1 − C 2 2 α m min ( 1 , C 2 M β ) ) ( F ( u ) − F ( u ∗ ) ) .
Proof. Proceeding as in the proof of Proposition 6.14 we deduce from (58) that
F ( u − t ∇ X F ( u ) ) ≤ F ( u ) + ( C 2 M t 2 2 − t ) ∥ ∇ X F ( u ) ∥ X 2 . \mathcal{F}\left(u-t\nabla_X\mathcal{F}(u)\right)\leq\mathcal{F}(u)+\left(\frac{C^2Mt^2}{2}-t\right)\lVert\nabla_X\mathcal{F}(u)\rVert_X^2. F ( u − t ∇ X F ( u ) ) ≤ F ( u ) + ( 2 C 2 M t 2 − t ) ∥ ∇ X F ( u ) ∥ X 2 .
In particular we note that if t ∈ [ 0 , 1 C 2 M ] t\in\left[0,\frac{1}{C^2M}\right] t ∈ [ 0 , C 2 M 1 ] then C 2 M t 2 2 − t ≤ − t 2 < − α t \frac{C^2Mt^2}{2}-t\leq-\frac{t}{2}<-\alpha t 2 C 2 M t 2 − t ≤ − 2 t < − α t , and so indeed (63) holds. As a consequence this means that either t ∗ = t 0 = 1 t_*=t_0=1 t ∗ = t 0 = 1 or t ∗ = β t i t_*=\beta t_i t ∗ = β t i for some t i t_i t i which did not satisfy the backtracking exit condition, and so which must satisfy t i > 1 C 2 M t_i>\frac{1}{C^2M} t i > C 2 M 1 . Putting these two options together establishes (64) . Finally we note that t ∗ t_* t ∗ must satisfy the backtracking exit condition (63) and so combining that inequality with (58) , (60) , and (64) allows us to obtain the main inequality. □
We now specialise Proposition 6.19 above to our variational energy.
Theorem 6.20 (Convergence rates for our variational energy with backtracking line search). Let E v a r E_{\mathrm{var}} E var be the variational energy introduced in Theorem 6.1, let V V V be a non-empty subspace of H ˚ 1 \mathring{H}^1 H ˚ 1 (possibly equal to H ˚ 1 \mathring{H}^1 H ˚ 1 ), let p ∗ p^* p ∗ denote the unique minimizer of E v a r E_{\mathrm{var}} E var in V V V , and let α ∈ ( 0 , 1 / 2 ) \alpha\in(0,1/2) α ∈ ( 0 , 1/2 ) and β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) . In each case below let t ∗ t_* t ∗ denote the step size obtained via backtracking line search with direction π \pi π (chosen differently in each case) and parameters α \alpha α and β \beta β .
Let p ∈ V p\in V p ∈ V and let π = − ∇ H ˚ 1 E v a r ∣ V ( p ) \pi=-\nabla_{\mathring{H}^1}E_{\mathrm{var}}|_V(p) π = − ∇ H ˚ 1 E var ∣ V ( p ) denote the H ˚ 1 \mathring{H}^1 H ˚ 1 -gradient descent direction for E v a r ∣ V E_{\mathrm{var}}|_V E var ∣ V at p p p . Then
t ∗ ≥ β and E v a r ( p + t ∗ π ) − E v a r ( p ∗ ) ≤ ( 1 − α β ) ( E v a r ( p ) − E v a r ( p ∗ ) ) . t_*\geq\beta\quad\text{and}\quad E_{\mathrm{var}}(p+t_*\pi)-E_{\mathrm{var}}(p^*)\leq(1-\alpha\beta)\left(E_{\mathrm{var}}(p)-E_{\mathrm{var}}(p^*)\right). t ∗ ≥ β and E var ( p + t ∗ π ) − E var ( p ∗ ) ≤ ( 1 − α β ) ( E var ( p ) − E var ( p ∗ ) ) . Let p ∈ V p\in V p ∈ V and let π \pi π denote the Newton descent direction for E v a r ∣ V E_{\mathrm{var}}|_V E var ∣ V at p p p . Then
t ∗ ≥ β 4 and E v a r ( p + t ∗ π ) − E v a r ( p ∗ ) ≤ ( 1 − α β 16 ) ( E v a r ( p ) − E v a r ( p ∗ ) ) . t_*\geq\frac{\beta}{4}\quad\text{and}\quad E_{\mathrm{var}}(p+t_*\pi)-E_{\mathrm{var}}(p^*)\leq\left(1-\frac{\alpha\beta}{16}\right)\left(E_{\mathrm{var}}(p)-E_{\mathrm{var}}(p^*)\right). t ∗ ≥ 4 β and E var ( p + t ∗ π ) − E var ( p ∗ ) ≤ ( 1 − 16 α β ) ( E var ( p ) − E var ( p ∗ ) ) . Suppose that V ⊆ H 2 V\subseteq H^2 V ⊆ H 2 , suppose that there exist a constant C n > 0 C_n>0 C n > 0 such that ∥ p ∥ H ˚ 1 ≤ C n ∥ p ∥ L 2 \lVert p\rVert_{\mathring{H}^1}\leq C_n\lVert p\rVert_{L^2} ∥ p ∥ H ˚ 1 ≤ C n ∥ p ∥ L 2 for every p ∈ V p\in V p ∈ V , and suppose that M ∈ H 1 M\in H^1 M ∈ H 1 and P V ∈ L 2 PV\in L^2 P V ∈ L 2 . Let p ∈ V p\in V p ∈ V and let π = − ∇ L 2 E v a r ∣ V ( p ) \pi=-\nabla_{L^2}E_{\mathrm{var}}|_V(p) π = − ∇ L 2 E var ∣ V ( p ) denote the L 2 L^2 L 2 -gradient descent direction for E v a r ∣ V E_{\mathrm{var}}|_V E var ∣ V at p p p . Then t ∗ ≥ min ( 1 , β C n 2 ) t_*\geq\min\left(1,\frac{\beta}{C_n^2}\right) t ∗ ≥ min ( 1 , C n 2 β ) and
E v a r ( p + t ∗ π ) − E v a r ( p ∗ ) ≤ ( 1 − α C n 2 min ( 1 , β C n 2 ) ) ( E v a r ( p ) − E v a r ( p ∗ ) ) . E_{\mathrm{var}}(p+t_*\pi)-E_{\mathrm{var}}(p^*)\leq\left(1-\frac{\alpha}{C_n^2}\min\left(1,\frac{\beta}{C_n^2}\right)\right)\left(E_{\mathrm{var}}(p)-E_{\mathrm{var}}(p^*)\right). E var ( p + t ∗ π ) − E var ( p ∗ ) ≤ ( 1 − C n 2 α min ( 1 , C n 2 β ) ) ( E var ( p ) − E var ( p ∗ ) ) .
Proof. This follows from Proposition 6.19 just as Theorem 6.16 follows from Proposition 6.14, noting that the appropriate values of the constant C C C in each case were identified in the proof of Theorem 6.16. □
With the convergence rates in hand we change gear and turn our attention towards a consistency result. We seek to prove that minimizers of the variational energy restricted to (appropriate) finite-dimensional subspaces of H ˚ 1 \mathring{H}^1 H ˚ 1 do converge to the true minimizer. In order to do so we first prove the following elementary estimate.
Lemma 6.21 (Explicit local modulus of continuity of the variational energy). Let E v a r E_{\mathrm{var}} E var be the variational energy introduced in Theorem 6.1. For any p 1 , p 2 ∈ H ˚ 1 p_1,p_2\in\mathring{H}^1 p 1 , p 2 ∈ H ˚ 1 the following estimate holds.
E v a r ( p 1 ) − E v a r ( p 2 ) ≤ ( 1 2 ∥ M ∥ L 2 + 3 4 ∥ ∇ p 1 ∥ L 2 + 3 4 ∥ ∇ p 2 ∥ L 2 ) ∥ ∇ ( p 1 − p 2 ) ∥ L 2 . (65) E_{\mathrm{var}}(p_1)-E_{\mathrm{var}}(p_2)\leq\left(\frac{1}{2}\lVert M\rVert_{L^2}+\frac{3}{4}\lVert\nabla p_1\rVert_{L^2}+\frac{3}{4}\lVert\nabla p_2\rVert_{L^2}\right)\lVert\nabla(p_1-p_2)\rVert_{L^2}. \tag*{(65)} E var ( p 1 ) − E var ( p 2 ) ≤ ( 2 1 ∥ M ∥ L 2 + 4 3 ∥ ∇ p 1 ∥ L 2 + 4 3 ∥ ∇ p 2 ∥ L 2 ) ∥ ∇ ( p 1 − p 2 ) ∥ L 2 . (65)
Proof. We write g 0 ( s ) : = min ( s , 0 ) g_0(s):=\min(s,0) g 0 ( s ) := min ( s , 0 ) such that
E v a r ( p 1 ) − E v a r ( p 2 ) = ∫ T 1 2 ( ∣ ∇ p 1 ∣ 2 − ∣ ∇ p 2 ∣ 2 ) − 1 4 [ g 0 ( M − ∂ 3 p 1 ) − g 0 ( M − ∂ 3 p 2 ) ] . E_{\mathrm{var}}(p_1)-E_{\mathrm{var}}(p_2)=\int_T\frac{1}{2}\left(\lvert\nabla p_1\rvert^2-\lvert\nabla p_2\rvert^2\right)-\frac{1}{4}\left[g_0(M-\partial_3p_1)-g_0(M-\partial_3p_2)\right]. E var ( p 1 ) − E var ( p 2 ) = ∫ T 2 1 ( ∣ ∇ p 1 ∣ 2 − ∣ ∇ p 2 ∣ 2 ) − 4 1 [ g 0 ( M − ∂ 3 p 1 ) − g 0 ( M − ∂ 3 p 2 ) ] .
We then estimate
∣ ∇ p 1 ∣ 2 − ∣ ∇ p 2 ∣ 2 = ∣ ∇ ( p 1 + p 2 ) ∣ ∣ ∇ ( p 1 − p 2 ) ∣ ≤ ( ∣ ∇ p 1 ∣ + ∣ ∇ p 2 ∣ ) ∣ ∇ ( p 1 − p 2 ) ∣ \left\lvert\nabla p_1\right\rvert^2-\left\lvert\nabla p_2\right\rvert^2=\lvert\nabla(p_1+p_2)\rvert\lvert\nabla(p_1-p_2)\rvert\leq\left(\lvert\nabla p_1\rvert+\lvert\nabla p_2\rvert\right)\lvert\nabla(p_1-p_2)\rvert ∣ ∇ p 1 ∣ 2 − ∣ ∇ p 2 ∣ 2 = ∣ ∇ ( p 1 + p 2 )∣ ∣ ∇ ( p 1 − p 2 )∣ ≤ ( ∣ ∇ p 1 ∣ + ∣ ∇ p 2 ∣ ) ∣ ∇ ( p 1 − p 2 )∣
and, since g 0 g_0 g 0 is 1-Lipschitz, we may proceed in the same way to estimate
∣ g 0 ( M − ∂ 3 p 1 ) 2 − g 0 ( M − ∂ 3 p 2 ) 2 ∣ ≤ ( 2 ∣ M ∣ + ∣ ∂ 3 p 1 ∣ + ∣ ∂ 3 p 2 ∣ ) ∣ ∂ 3 ( p 1 − p 2 ) ∣ . \left\lvert g_0(M-\partial_3p_1)^2-g_0(M-\partial_3p_2)^2\right\rvert\leq\left(2\lvert M\rvert+\lvert\partial_3p_1\rvert+\lvert\partial_3p_2\rvert\right)\lvert\partial_3(p_1-p_2)\rvert. g 0 ( M − ∂ 3 p 1 ) 2 − g 0 ( M − ∂ 3 p 2 ) 2 ≤ ( 2 ∣ M ∣ + ∣ ∂ 3 p 1 ∣ + ∣ ∂ 3 p 2 ∣ ) ∣ ∂ 3 ( p 1 − p 2 )∣ .
Putting it all together produces the desired inequality. □
We are now equipped to prove the last main result of this section, which pertains to numerical consistency.
Theorem 6.22 (Convergence of the minimizers of the approximate variational energy to the minimizer of the true variational energy). Let V n ⊂ H ˙ 1 V_n \subset\dot{H}^{1} V n ⊂ H ˙ 1 be a closed subspace of H ˙ 1 \dot{H}^{1} H ˙ 1 , define E n : = E v a r ∣ V n E_n := E_{\mathrm{var}}|_{V_n} E n := E var ∣ V n , and let p ∗ : = arg min E v a r p^{*} := \operatorname{arg\,min} E_{\mathrm{var}} p ∗ := arg min E var and p n ∗ : = arg min E n p_n^{*} := \operatorname{arg\,min} E_n p n ∗ := arg min E n . For any p n ε , p n ∈ V n p_n^{\varepsilon}, p_n \in V_n p n ε , p n ∈ V n ,
∥ p n ε − p ∗ ∥ H ˙ 1 ≤ 4 [ E n ( p n ε ) − min E n ] + ( 8 ∥ M ∥ L 2 + 12 ∥ P V ∥ H − 1 + 3 ∥ p n − p ∗ ∥ H ˙ 1 ) ∥ p n − p ∗ ∥ H ˙ 1 . (66) \lVert p_n^{\varepsilon} - p^{*} \rVert_{\dot{H}^{1}} \leq4\left[E_n(p_n^{\varepsilon})-\min E_n\right]+\left(8\lVert M\rVert_{L^2}+12\lVert PV\rVert_{H^{-1}}+3\lVert p_n-p^{*}\rVert_{\dot{H}^{1}}\right)\lVert p_n-p^{*}\rVert_{\dot{H}^{1}}. \tag*{(66)} ∥ p n ε − p ∗ ∥ H ˙ 1 ≤ 4 [ E n ( p n ε ) − min E n ] + ( 8 ∥ M ∥ L 2 + 12 ∥ P V ∥ H − 1 + 3 ∥ p n − p ∗ ∥ H ˙ 1 ) ∥ p n − p ∗ ∥ H ˙ 1 . (66) In particular, if p n ε p_n^{\varepsilon} p n ε and p n p_n p n are chosen to satisfy, for some α , C > 0 \alpha, C>0 α , C > 0 ,
E n ( p n ε ) − min E n ≤ ε and ∥ p n − p ∗ ∥ H ˙ 1 ≤ C n − α for all n (67) E_n(p_n^{\varepsilon})-\min E_n\leq\varepsilon\quad\text{and}\quad\lVert p_n-p^{*}\rVert_{\dot{H}^{1}}\leq Cn^{-\alpha}\text{ for all }n \tag*{(67)} E n ( p n ε ) − min E n ≤ ε and ∥ p n − p ∗ ∥ H ˙ 1 ≤ C n − α for all n (67) then, for n n n sufficiently large to ensure C n − α ≤ 1 Cn^{-\alpha}\leq1 C n − α ≤ 1 ,
∥ p n ε − p ∗ ∥ H ˙ 1 ≲ ε + ( 1 + ∥ M ∥ L 2 + ∥ P V ∥ H − 1 ) n − α , (68) \lVert p_n^{\varepsilon}-p^{*}\rVert_{\dot{H}^{1}}\lesssim\varepsilon+\left(1+\lVert M\rVert_{L^2}+\lVert PV\rVert_{H^{-1}}\right)n^{-\alpha}, \tag*{(68)} ∥ p n ε − p ∗ ∥ H ˙ 1 ≲ ε + ( 1 + ∥ M ∥ L 2 + ∥ P V ∥ H − 1 ) n − α , (68) i.e. ε \varepsilon ε -minimizers of the discretised variational energy E n E_n E n can be made arbitrarily close to the minimizer of the true variational energy E v a r E_{\mathrm{var}} E var .
Proof. First note that the p n ∗ p_n^{*} p n ∗ ’s are well-defined since the properties of E v a r E_{\mathrm{var}} E var parlay over to properties of E n E_n E n and the proof of Theorem 6.3 can be reproduced for E n E_n E n .
Since (68) follows from (66) under the additional assumptions of (67) it suffices to establish (66) . First we note that, since p ∗ p^{*} p ∗ and p n ∗ p_n^{*} p n ∗ are minimizers of E v a r E_{\mathrm{var}} E var and E n E_n E n , respectively, item 5 of Theorem 6.2 (which holds mutatis mutandis for E n E_n E n ) tells us that
∥ p n ε − p ∗ ∥ H ˙ 1 ≤ ∥ p n ε − p n ∗ ∥ H ˙ 1 + ∥ p n ∗ − p ∗ ∥ H ˙ 1 \lVert p_n^{\varepsilon}-p^{*}\rVert_{\dot{H}^{1}}\leq\lVert p_n^{\varepsilon}-p_n^{*}\rVert_{\dot{H}^{1}}+\lVert p_n^{*}-p^{*}\rVert_{\dot{H}^{1}} ∥ p n ε − p ∗ ∥ H ˙ 1 ≤ ∥ p n ε − p n ∗ ∥ H ˙ 1 + ∥ p n ∗ − p ∗ ∥ H ˙ 1
≤ 4 [ E n ( p n ε ) − E n ( p n ∗ ) ] + 4 [ E v a r ( p n ∗ ) − E v a r ( p ∗ ) ] . (69) \leq4\left[E_n(p_n^{\varepsilon})-E_n(p_n^{*})\right]+4\left[E_{\mathrm{var}}(p_n^{*})-E_{\mathrm{var}}(p^{*})\right]. \tag*{(69)} ≤ 4 [ E n ( p n ε ) − E n ( p n ∗ ) ] + 4 [ E var ( p n ∗ ) − E var ( p ∗ ) ] . (69)
By virtue of Theorem 6.3 and (66) we may estimate the second term above via
E v a r ( p n ∗ ) − E v a r ( p ∗ ) = E n ( p n ∗ ) − E v a r ( p ∗ ) = min E n − E v a r ( p ∗ ) ≤ E v a r ( p n ) − E v a r ( p ∗ ) ≤ ( 1 2 ∥ M ∥ L 2 + 3 4 ( ∥ p n ∥ H ˙ 1 + ∥ p ∗ ∥ H ˙ 1 ) ) ∥ p n − p ∗ ∥ H ˙ 1 ≤ ( 2 ∥ M ∥ L 2 + 3 ∥ P V ∥ H − 1 + 3 4 ∥ p n − p ∗ ∥ H ˙ 1 ) ∥ p n − p ∗ ∥ H ˙ 1 . (70) \begin{aligned}
E_{\mathrm{var}}(p_n^{*})-E_{\mathrm{var}}(p^{*})=E_n(p_n^{*})-E_{\mathrm{var}}(p^{*})=\min E_n-E_{\mathrm{var}}(p^{*})\leq E_{\mathrm{var}}(p_n)-E_{\mathrm{var}}(p^{*}) \\
&\leq\left(\frac{1}{2}\lVert M\rVert_{L^2}+\frac{3}{4}\left(\lVert p_n\rVert_{\dot{H}^{1}}+\lVert p^{*}\rVert_{\dot{H}^{1}}\right)\right)\lVert p_n-p^{*}\rVert_{\dot{H}^{1}} \\
&\leq\left(2\lVert M\rVert_{L^2}+3\lVert PV\rVert_{H^{-1}}+\frac{3}{4}\lVert p_n-p^{*}\rVert_{\dot{H}^{1}}\right)\lVert p_n-p^{*}\rVert_{\dot{H}^{1}}. \tag*{(70)}
\end{aligned} E var ( p n ∗ ) − E var ( p ∗ ) = E n ( p n ∗ ) − E var ( p ∗ ) = min E n − E var ( p ∗ ) ≤ E var ( p n ) − E var ( p ∗ ) ≤ ( 2 1 ∥ M ∥ L 2 + 4 3 ( ∥ p n ∥ H ˙ 1 + ∥ p ∗ ∥ H ˙ 1 ) ) ∥ p n − p ∗ ∥ H ˙ 1 ≤ ( 2 ∥ M ∥ L 2 + 3 ∥ P V ∥ H − 1 + 4 3 ∥ p n − p ∗ ∥ H ˙ 1 ) ∥ p n − p ∗ ∥ H ˙ 1 . (70)
Putting (69) and (70) together then produces (66) .
Geometry
In this section we do two things.
Foreshadowed by Proposition 3.1 we see that p p p and M M M may be viewed as coordinates along the balanced set B \mathcal{B} B . Moreover P V PV P V and M M M then arise naturally as dual coordinates (with respect to some natural inner product arising from the energy).
Equipped with this geometric language we interpret the Newton descent iteration of § 6.1 above in geometric terms. We explain how the Newton iteration can be viewed as a flow along the balanced set B \mathcal{B} B . This flow follows the coordinate direction along M M M while driving P V PV P V towards its true value.
We begin by recalling a definition from § 3 .
Definition 6.23 (Global chart for the balanced set). For H : = H ˙ 1 × L 2 \mathcal{H}:=\dot{H}^{1}\times L^{2} H := H ˙ 1 × L 2 we define Φ : H → B \Phi:\mathcal{H}\to\mathcal{B} Φ : H → B via
Φ ( p , M ) = ( ∇ h ⊥ p ∂ 3 p + 1 2 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 1 2 min 0 ( M − ∂ 3 p ) ) . \Phi(p,M)=
\begin{pmatrix}
\nabla_h^{\perp}p\\
\partial_3p+\frac{1}{2}\min_0(M-\partial_3p)\\
M-\partial_3p-\frac{1}{2}\min_0(M-\partial_3p)
\end{pmatrix}. Φ ( p , M ) = ∇ h ⊥ p ∂ 3 p + 2 1 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 2 1 min 0 ( M − ∂ 3 p ) .
We now collect some more properties of this map.
Proposition 6.24. Φ \Phi Φ is bi -Lipschitz, i.e. Lipschitz with Lipschitz inverse .
In particular, since the parametrisation Φ \Phi Φ of B \mathcal{B} B is bi-Lipschitz it follows that the balanced set is an honest-to-goodness manifold. This is depicted in Figure 7 .
Figure 7. A depiction of the global chart for the balanced set introduced in Definition 6.23 .
Corollary 6.25. The balanced set B \mathcal{B} B is a Lipschitz-regular Hilbert manifold, modeled after the Hilbert space H \mathbb{H} H , with global chart Φ \Phi Φ .
Proof of Corollary 6.25 . This follows from immediately from Proposition 6.24 .
Proof of Proposition 6.24 . First we note that the inverse of Φ \Phi Φ is given by
Φ − 1 ( u , θ , q ) = ( Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − min 0 q ) ] θ + q ) . \Phi^{-1}(u,\theta,q)=\begin{pmatrix}\Delta^{-1}\left[\nabla_h^\perp\cdot u_h+\partial_3(\theta-\min_0 q)\right]\\ \theta+q\end{pmatrix}. Φ − 1 ( u , θ , q ) = ( Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − min 0 q ) ] θ + q ) .
Indeed, suppose that ( u , θ , q ) = Φ ( p , M ) (u,\theta,q)=\Phi(p,M) ( u , θ , q ) = Φ ( p , M ) . We see that
sign q = sign ( M − ∂ 3 p ) \operatorname{sign}q=\operatorname{sign}(M-\partial_3p) sign q = sign ( M − ∂ 3 p )
and so Lemmas A.6 and A.5 tell us that
( u , θ , q ) = Φ ( p , M ) ⟺ { u = ∇ h ⊥ p θ = ∂ 3 p + 1 2 min 0 ( M − ∂ 3 p ) q = M − ∂ 3 p − 1 2 min 0 ( M − ∂ 3 p ) ⟺ { u = ∇ h ⊥ p θ + q = M θ − min 0 q = ∂ 3 p ⟺ { − u h ⊥ + ( θ − min 0 q ) e 3 = ∇ p u 3 = 0 θ + q = M ⟺ { p = Δ − 1 ∇ ⋅ [ − u h ⊥ + ( θ − min 0 q ) e 3 ] u 3 = 0 M = θ + q ⟺ { p = Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − min 0 q ) ] M = θ + q u 3 = 0 ⟺ ( p , M ) = Φ − 1 ( u , θ , q ) . (u,\theta,q)=\Phi(p,M)\Longleftrightarrow
\begin{cases}
u=\nabla_h^\perp p\\
\theta=\partial_3p+\frac{1}{2}\min_0(M-\partial_3p)\\
q=M-\partial_3p-\frac{1}{2}\min_0(M-\partial_3p)
\end{cases}
\Longleftrightarrow
\begin{cases}
u=\nabla_h^\perp p\\
\theta+q=M\\
\theta-\min_0q=\partial_3p
\end{cases}
\Longleftrightarrow
\begin{cases}
-u_h^\perp+(\theta-\min_0q)e_3=\nabla p\\
u_3=0\\
\theta+q=M
\end{cases}
\Longleftrightarrow
\begin{cases}
p=\Delta^{-1}\nabla\cdot\left[-u_h^\perp+(\theta-\min_0q)e_3\right]\\
u_3=0\\
M=\theta+q
\end{cases}
\Longleftrightarrow
\begin{cases}
p=\Delta^{-1}\left[\nabla_h^\perp\cdot u_h+\partial_3(\theta-\min_0q)\right]\\
M=\theta+q\\
u_3=0
\end{cases}
\Longleftrightarrow(p,M)=\Phi^{-1}(u,\theta,q). ( u , θ , q ) = Φ ( p , M ) ⟺ ⎩ ⎨ ⎧ u = ∇ h ⊥ p θ = ∂ 3 p + 2 1 min 0 ( M − ∂ 3 p ) q = M − ∂ 3 p − 2 1 min 0 ( M − ∂ 3 p ) ⟺ ⎩ ⎨ ⎧ u = ∇ h ⊥ p θ + q = M θ − min 0 q = ∂ 3 p ⟺ ⎩ ⎨ ⎧ − u h ⊥ + ( θ − min 0 q ) e 3 = ∇ p u 3 = 0 θ + q = M ⟺ ⎩ ⎨ ⎧ p = Δ − 1 ∇ ⋅ [ − u h ⊥ + ( θ − min 0 q ) e 3 ] u 3 = 0 M = θ + q ⟺ ⎩ ⎨ ⎧ p = Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − min 0 q ) ] M = θ + q u 3 = 0 ⟺ ( p , M ) = Φ − 1 ( u , θ , q ) .
The fact that both Φ \Phi Φ and Φ − 1 \Phi^{-1} Φ − 1 are Lipschitz then follows from the fact that s ↦ min 0 s s\mapsto\min_0s s ↦ min 0 s is Lipschitz, that Δ : H 1 → H − 1 \Delta:H^1\to H^{-1} Δ : H 1 → H − 1 is an isometry, and from Lemma A.14 . Indeed, for Φ \Phi Φ we have that
∥ Φ ( p 1 , M 1 ) − Φ ( p 2 , M 2 ) ∥ L 2 2 ≲ ∥ ∇ p 1 − ∇ p 2 ∥ L 2 2 + ∥ M 1 − M 2 ∥ L 2 2 = ∥ ( p 1 , M 1 ) − ( p 2 , M 2 ) ∥ H ˙ 1 × L 2 2 \|\Phi(p_1,M_1)-\Phi(p_2,M_2)\|_{L^2}^2\lesssim\|\nabla p_1-\nabla p_2\|_{L^2}^2+\|M_1-M_2\|_{L^2}^2
=\|(p_1,M_1)-(p_2,M_2)\|_{\dot H^1\times L^2}^2 ∥Φ ( p 1 , M 1 ) − Φ ( p 2 , M 2 ) ∥ L 2 2 ≲ ∥∇ p 1 − ∇ p 2 ∥ L 2 2 + ∥ M 1 − M 2 ∥ L 2 2 = ∥ ( p 1 , M 1 ) − ( p 2 , M 2 ) ∥ H ˙ 1 × L 2 2
while for Φ − 1 \Phi^{-1} Φ − 1 we have that
∥ Φ − 1 ( u 1 , θ 1 , q 1 ) − Φ − 1 ( u 2 , θ 2 , q 2 ) ∥ H ˙ 1 × L 2 2 = ∥ Δ − 1 [ ∇ h ⊥ ⋅ u h , 1 + ∂ 3 ( θ 1 − min 0 q 1 ) − ∇ h ⊥ ⋅ u h , 2 − ∂ 3 ( θ 2 − min 0 q 2 ) ] ∥ H ˙ 1 2 + ∥ ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ∥ L 2 2 = ∥ ∇ h ⊥ ⋅ u h , 1 + ∂ 3 ( θ 1 − min 0 q 1 ) − ∇ h ⊥ ⋅ u h , 2 − ∂ 3 ( θ 2 − min 0 q 2 ) ∥ H ˙ − 1 2 + ∥ ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ∥ L 2 2 \begin{aligned}
\left\|\Phi^{-1}(u_1,\theta_1,q_1)-\Phi^{-1}(u_2,\theta_2,q_2)\right\|_{\dot{H}^{1}\times L^2}^2 \\
&= \left\|\Delta^{-1}\left[\nabla_h^\perp\cdot u_{h,1}+\partial_3(\theta_1-\min_0 q_1)-\nabla_h^\perp\cdot u_{h,2}-\partial_3(\theta_2-\min_0 q_2)\right]\right\|_{\dot{H}^{1}}^2+\left\|(\theta_1+q_1)-(\theta_2+q_2)\right\|_{L^2}^2 \\
&= \left\|\nabla_h^\perp\cdot u_{h,1}+\partial_3(\theta_1-\min_0 q_1)-\nabla_h^\perp\cdot u_{h,2}-\partial_3(\theta_2-\min_0 q_2)\right\|_{\dot{H}^{-1}}^2+\left\|(\theta_1+q_1)-(\theta_2+q_2)\right\|_{L^2}^2
\end{aligned} Φ − 1 ( u 1 , θ 1 , q 1 ) − Φ − 1 ( u 2 , θ 2 , q 2 ) H ˙ 1 × L 2 2 = Δ − 1 [ ∇ h ⊥ ⋅ u h , 1 + ∂ 3 ( θ 1 − 0 min q 1 ) − ∇ h ⊥ ⋅ u h , 2 − ∂ 3 ( θ 2 − 0 min q 2 ) ] H ˙ 1 2 + ∥ ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ∥ L 2 2 = ∇ h ⊥ ⋅ u h , 1 + ∂ 3 ( θ 1 − 0 min q 1 ) − ∇ h ⊥ ⋅ u h , 2 − ∂ 3 ( θ 2 − 0 min q 2 ) H ˙ − 1 2 + ∥ ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ∥ L 2 2
≲ ∥ u h , 1 − u h , 2 ∥ L 2 2 + ∥ ( θ 1 − min 0 q 1 ) − ( θ 2 − min 0 q 2 ) ∥ L 2 2 + ∥ ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ∥ L 2 2 \lesssim\left\|u_{h,1}-u_{h,2}\right\|_{L^2}^2+\left\|(\theta_1-\min_0 q_1)-(\theta_2-\min_0 q_2)\right\|_{L^2}^2+\left\|(\theta_1+q_1)-(\theta_2+q_2)\right\|_{L^2}^2 ≲ ∥ u h , 1 − u h , 2 ∥ L 2 2 + ( θ 1 − 0 min q 1 ) − ( θ 2 − 0 min q 2 ) L 2 2 + ∥ ( θ 1 + q 1 ) − ( θ 2 + q 2 ) ∥ L 2 2
≲ ∥ ( u 1 , θ 1 , q 1 ) − ( u 2 , θ 2 , q 2 ) ∥ L 2 2 . \lesssim\left\|(u_1,\theta_1,q_1)-(u_2,\theta_2,q_2)\right\|_{L^2}^2. ≲ ∥ ( u 1 , θ 1 , q 1 ) − ( u 2 , θ 2 , q 2 ) ∥ L 2 2 .
Remark 6.26 (Alternate form of Φ − 1 \Phi^{-1} Φ − 1 ). We typically write Φ − 1 \Phi^{-1} Φ − 1 in terms of the nonlinear P V PV P V -and-M M M inversion, and not in terms of the inversion of a Laplacian as is done in the proof of Proposition 6.24 above. This is because the Laplacian inversion does not use P V PV P V , but instead uses a potential vorticity based on the buoyancy (by contrast with P V PV P V which is based on the equivalent potential temperature θ \theta θ ) given by
P V b = ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − min 0 q ) . PV_b=\nabla_h^\perp\cdot u_h+\partial_3(\theta-\min_0 q). P V b = ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − 0 min q ) .
Typically we do not want to use P V b PV_b P V b since it is not slow. That is why P V PV P V is usually preferable. However, for the purpose of the proof of Proposition 6.24 , P V b PV_b P V b is preferable to P V PV P V since it lets us prove more easily that Φ − 1 \Phi^{-1} Φ − 1 is Lipschitz.
So far we have been able to endow the balanced set B \mathcal{B} B with the structure of a Lipschitz-regular manifold. Unfortunately, that is as regular as the balanced set gets due to the presence of min 0 q \min_0 q min 0 q in the buoyancy term, which is Lipschitz but not differentiable.
Nonetheless we can formally differentiate min 0 q \min_0 q min 0 q and thus endow the balanced set B \mathcal{B} B with a formal differentiable structure. In order to do so, we introduce notation for its formal tangent space, for the formal derivative of the chart Φ \Phi Φ , and for the inner product we will use (which comes from the conserved energy).
Definition 6.27. Fix an indicator function H H H .
We define
T H B : = { ( u , θ , q ) ∈ L σ 2 : ∂ 3 u h − ∇ h ⊥ ( θ − q H ) = 0 , u 3 = 0 , ∫ u h = 0 , and ∫ ( θ − q H ) = 0 } . T_H\mathcal{B}:=\left\{(u,\theta,q)\in L^2_\sigma:\ \partial_3u_h-\nabla_h^\perp(\theta-qH)=0,\ u_3=0,\ \int u_h=0,\ \text{and }\int(\theta-qH)=0\right\}. T H B := { ( u , θ , q ) ∈ L σ 2 : ∂ 3 u h − ∇ h ⊥ ( θ − q H ) = 0 , u 3 = 0 , ∫ u h = 0 , and ∫ ( θ − q H ) = 0 } .
We define D H Φ : H → T H B D_H\Phi:\mathbb{H}\to T_H\mathcal{B} D H Φ : H → T H B via
D H Φ ( p , M ) : = ( ∂ 3 p + ∇ h ⊥ 2 ( M − ∂ 3 p ) M − ∂ 3 p − H 2 ( M − ∂ 3 p ) ) . D_H\Phi(p,M):=\begin{pmatrix}
\partial_3p+\dfrac{\nabla_h^\perp}{2}\left(M-\partial_3p\right)\\
M-\partial_3p-\dfrac{H}{2}\left(M-\partial_3p\right)
\end{pmatrix}. D H Φ ( p , M ) := ∂ 3 p + 2 ∇ h ⊥ ( M − ∂ 3 p ) M − ∂ 3 p − 2 H ( M − ∂ 3 p ) .
For any ( u , θ , q ) ∈ L σ 2 (u,\theta,q)\in L^2_\sigma ( u , θ , q ) ∈ L σ 2 we define
∥ ( u , θ , q ) ∥ L H 2 2 : = ∫ ∣ u ∣ 2 + θ 2 + q 2 H + ( θ + q ) 2 \left\|(u,\theta,q)\right\|_{L_H^2}^2:=\int|u|^2+\theta^2+q^2H+(\theta+q)^2 ∥ ( u , θ , q ) ∥ L H 2 2 := ∫ ∣ u ∣ 2 + θ 2 + q 2 H + ( θ + q ) 2
and let ⟨ ⋅ , ⋅ ⟩ L H 2 \langle\cdot,\cdot\rangle_{L_H^2} ⟨ ⋅ , ⋅ ⟩ L H 2 denote the corresponding innner product.
As expected, the formal derivative of Φ \Phi Φ completely characterizes the formal tangent space. (This would be trivially true if the balanced set B \mathcal{B} B was endowed with an honest-to-goodness differentiable structure by its chart Φ \Phi Φ .)
Lemma 6.28. T H B = im D H Φ T_H\mathcal{B}=\operatorname{im}D_H\Phi T H B = im D H Φ .
Proof. Proceeding similarly to the proof of Proposition 3.1 we introduce the “good unknown” (see Remark 3.2)
v H : = − u h ⊥ + ( θ − q H ) e 3 . (71) v_H:=-u_h^\perp+(\theta-qH)e_3. \tag*{(71)} v H := − u h ⊥ + ( θ − q H ) e 3 . (71)
For X = ( u , θ , q ) \mathcal{X}=(u,\theta,q) X = ( u , θ , q ) we have on one hand that, as per Lemma A.5 ,
X ∈ T H B ⟺ ∇ × v H , ∫ v H = 0 , and u 3 = 0. \mathcal{X}\in T_H\mathcal{B}\Longleftrightarrow\nabla\times v_H,\ \int v_H=0,\ \text{and }u_3=0. X ∈ T H B ⟺ ∇ × v H , ∫ v H = 0 , and u 3 = 0.
On the other hand, proceeding as in the proof of Proposition 6.24 and using Lemma A.6 tells us that
X = D H Φ ( p , M ) ⟺ { u H ⊥ + ( θ − q H ) e 3 = ∇ p u 3 = 0 θ + q = 0 ⟺ { v H = ∇ p u 3 = 0 θ + q = M \mathcal{X}=D_H\Phi(p,M)\Longleftrightarrow
\begin{cases}
u_H^\perp+(\theta-qH)e_3=\nabla p\\
u_3=0\\
\theta+q=0
\end{cases}
\Longleftrightarrow
\begin{cases}
v_H=\nabla p\\
u_3=0\\
\theta+q=M
\end{cases} X = D H Φ ( p , M ) ⟺ ⎩ ⎨ ⎧ u H ⊥ + ( θ − q H ) e 3 = ∇ p u 3 = 0 θ + q = 0 ⟺ ⎩ ⎨ ⎧ v H = ∇ p u 3 = 0 θ + q = M
The Helmholtz decomposition (Corollary A.2) allows us to conclude:
X ∈ T H B ⟺ ∇ × v H = 0 , ∫ v H = 0 , and u 3 = 0 ⟺ ∇ × v H = 0 , ∫ v H = 0 , u 3 = 0 , and θ + q = : M ⟺ v H = ∇ p for some p ∈ H ˚ 1 , u 3 = 0 , and θ + q = M ⟺ X = D H Φ ( p , M ) for some ( p , M ) ∈ H ⟺ X ∈ im D H Φ . \begin{aligned}
\mathcal{X}\in T_H\mathcal{B}
&\Longleftrightarrow\nabla\times v_H=0,\ \int v_H=0,\ \text{and }u_3=0\\
&\Longleftrightarrow\nabla\times v_H=0,\ \int v_H=0,\ u_3=0,\ \text{and }\theta+q=:M\\
&\Longleftrightarrow v_H=\nabla p\text{ for some }p\in\mathring{H}^{1},\ u_3=0,\ \text{and }\theta+q=M\\
&\Longleftrightarrow\mathcal{X}=D_H\Phi(p,M)\text{ for some }(p,M)\in\mathbb{H}\\
&\Longleftrightarrow\mathcal{X}\in\operatorname{im}D_H\Phi.
\end{aligned} X ∈ T H B ⟺ ∇ × v H = 0 , ∫ v H = 0 , and u 3 = 0 ⟺ ∇ × v H = 0 , ∫ v H = 0 , u 3 = 0 , and θ + q =: M ⟺ v H = ∇ p for some p ∈ H ˚ 1 , u 3 = 0 , and θ + q = M ⟺ X = D H Φ ( p , M ) for some ( p , M ) ∈ H ⟺ X ∈ im D H Φ.
We now remark that, as suggested by the notation we chose to employ, D H Φ D_H\Phi D H Φ may be viewed formally as the derivative of Φ \Phi Φ . Since the balanced set B \mathcal{B} B is the image of the chart Φ \Phi Φ , it follows that the image of its formal derivative D H Φ D_H\Phi D H Φ may be interpreted formally as the tangent space to that same balanced set. Since Lemma 6.28 above tells us that the image of D H Φ D_H\Phi D H Φ is precisely T H B T_H\mathcal{B} T H B , we see that once again the notation was chosen to be suggestive: T H B T_H\mathcal{B} T H B may be viewed formally as the tangent space to the balanced set B \mathcal{B} B . This is made precise in Remark 6.29 below and is depicted in Figure 8 .
Figure 8. The formal tangent space to the balanced set introduced in Definition 6.27 is the image of the formal derivative of the global chart.
Remark 6.29. Formally, we may compute that, for H : = 1 ( q < 0 ) H := \mathbb{1}(q < 0) H := 1 ( q < 0 ) and for any derivative ∂ \partial ∂ ,
∂ ( min 0 q ) = ( ∂ q ) 1 ( q < 0 ) = ( ∂ q ) H . \partial(\min_{0} q) = (\partial q)\mathbb{1}(q < 0) = (\partial q)H. ∂ ( 0 min q ) = ( ∂ q ) 1 ( q < 0 ) = ( ∂ q ) H .
This shows that D H Φ D_H\Phi D H Φ is, formally, the derivative of Φ \Phi Φ . Since Φ \Phi Φ is a global chart for the Hilbert manifold B \mathcal{B} B this means that we may formally view T H B T_H\mathcal{B} T H B , which is the image of D H Φ D_H\Phi D H Φ , as the tangent space of B \mathcal{B} B .
An alternative justification, more direct albeit still formal, is as follows. Let X = ( u , θ , q ) : ( − 1 , 1 ) → B \mathcal{X}=(u,\theta,q):(-1,1)\to\mathcal{B} X = ( u , θ , q ) : ( − 1 , 1 ) → B be a curve of balanced states and let Z \mathcal{Z} Z denote its derivative at zero, i.e.
Z = ( v , φ , r ) : = d d s X ( s ) ∣ s = 0 . \mathcal{Z}=(v,\varphi,r):=\left.\frac{d}{ds}\mathcal{X}(s)\right|_{s=0}. Z = ( v , φ , r ) := d s d X ( s ) s = 0 .
Since X ∈ B \mathcal{X}\in\mathcal{B} X ∈ B it follows that
j ( X ) = 0 , w ( X ) = 0 , and a ( X ) = 0. j(\mathcal{X})=0,\quad w(\mathcal{X})=0,\quad\text{and}\quad a(\mathcal{X})=0. j ( X ) = 0 , w ( X ) = 0 , and a ( X ) = 0.
Both w w w and a a a are linear, so we deduce immediately that
w ( Z ) = 0 and a ( Z ) = 0. w(\mathcal{Z})=0\quad\text{and}\quad a(\mathcal{Z})=0. w ( Z ) = 0 and a ( Z ) = 0.
What about j j j ? Using the formal identity ∂ ( min 0 q ) = ( ∂ q ) H \partial(\min_{0}q)=(\partial q)H ∂ ( min 0 q ) = ( ∂ q ) H discussed above we compute that
0 = d d s j ( X ) ∣ s = 0 = d d s [ ∂ 3 u h − ∇ h ⊥ ( θ − min 0 q ) ] ∣ s = 0 = ∂ 3 v h − ∇ h ⊥ ( φ − r H ) = : j H ( Z ) . 0=\left.\frac{d}{ds}j(\mathcal{X})\right|_{s=0}
=\left.\frac{d}{ds}\left[\partial_{3}u_h-\nabla_h^{\perp}(\theta-\min_{0}q)\right]\right|_{s=0}
=\partial_{3}v_h-\nabla_h^{\perp}(\varphi-rH)=:j_H(\mathcal{Z}). 0 = d s d j ( X ) s = 0 = d s d [ ∂ 3 u h − ∇ h ⊥ ( θ − 0 min q ) ] s = 0 = ∂ 3 v h − ∇ h ⊥ ( φ − r H ) =: j H ( Z ) .
This means that Z \mathcal{Z} Z belongs to T H B T_H\mathcal{B} T H B . In other words: if X \mathcal{X} X is a curve of balanced states then its formal tangent vector Z \mathcal{Z} Z lies in T H B T_H\mathcal{B} T H B . This warrants treating formally T H B T_H\mathcal{B} T H B as the tangent space of the balanced set B \mathcal{B} B .
Note that since the chart Φ \Phi Φ is invertible and since D H Φ D_H\Phi D H Φ is its formal derivative, we may expect D H Φ D_H\Phi D H Φ to inherit this invertibility. This is indeed the case. We record this result here since it will be of use later.
Lemma 6.30. D H Φ D_H\Phi D H Φ is invertible.
Proof. This follows as in the proof that Φ \Phi Φ is invertible (Proposition 6.24 ) since ( 1 + H ) ( 1 − H / 2 ) = 1 (1+H)(1-H/2)=1 ( 1 + H ) ( 1 − H /2 ) = 1 and so
( u , θ , q ) = D H Φ ( p , M ) ⟺ { u = ∇ h ⊥ p , θ = ∂ 3 p + H 2 ( M − ∂ 3 p ) , q = M − ∂ 3 p − H 2 ( M − ∂ 3 p ) ⟺ { − u h ⊥ + ( θ − q H ) e 3 = ∇ p , u 3 = 0 , θ + q = M (u,\theta,q)=D_H\Phi(p,M)\Longleftrightarrow
\begin{cases}
u=\nabla_h^{\perp}p,\\
\theta=\partial_3p+\dfrac{H}{2}(M-\partial_3p),\\
q=M-\partial_3p-\dfrac{H}{2}(M-\partial_3p)
\end{cases}
\Longleftrightarrow
\begin{cases}
-u_h^{\perp}+(\theta-qH)e_3=\nabla p,\\
u_3=0,\\
\theta+q=M
\end{cases} ( u , θ , q ) = D H Φ ( p , M ) ⟺ ⎩ ⎨ ⎧ u = ∇ h ⊥ p , θ = ∂ 3 p + 2 H ( M − ∂ 3 p ) , q = M − ∂ 3 p − 2 H ( M − ∂ 3 p ) ⟺ ⎩ ⎨ ⎧ − u h ⊥ + ( θ − q H ) e 3 = ∇ p , u 3 = 0 , θ + q = M ⟺ { p = Δ − 1 ∇ ⋅ [ − u h ⊥ + ( θ − q H ) e 3 ] , u 3 = 0 , M = θ + q ⟺ { p = Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − q H ) ] , u 3 = 0 , M = θ + q . \Longleftrightarrow
\begin{cases}
p=\Delta^{-1}\nabla\cdot\left[-u_h^{\perp}+(\theta-qH)e_3\right],\\
u_3=0,\\
M=\theta+q
\end{cases}
\Longleftrightarrow
\begin{cases}
p=\Delta^{-1}\left[\nabla_h^{\perp}\cdot u_h+\partial_3(\theta-qH)\right],\\
u_3=0,\\
M=\theta+q.
\end{cases} ⟺ ⎩ ⎨ ⎧ p = Δ − 1 ∇ ⋅ [ − u h ⊥ + ( θ − q H ) e 3 ] , u 3 = 0 , M = θ + q ⟺ ⎩ ⎨ ⎧ p = Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − q H ) ] , u 3 = 0 , M = θ + q . and so the inverse is given by
( D H Φ ) − 1 ( u , θ , q ) = ( Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − q H ) ] θ + q ) . (D_H\Phi)^{-1}(u,\theta,q)=
\begin{pmatrix}
\Delta^{-1}\left[\nabla_h^{\perp}\cdot u_h+\partial_3(\theta-qH)\right]\\
\theta+q
\end{pmatrix}. ( D H Φ ) − 1 ( u , θ , q ) = ( Δ − 1 [ ∇ h ⊥ ⋅ u h + ∂ 3 ( θ − q H ) ] θ + q ) .
Note that in particular this allows us to view D H Φ D_H\Phi D H Φ as recording the coordinate vectors arising from the chart $\Phi. This is made precise in Remark 6.31 below and depicted in Figure 9 .
Figure 9. The “columns” of the formal derivative D H Φ D_H\Phi D H Φ , denoted by e p e_p e p and e M e_M e M , may be viewed as coordinate vectors on the formal tangent space.
A minor miracle now occurs: while p p p and M M M are coordinates along the balanced set B \mathcal{B} B , their dual coordinates turn out to be, essentially, P V PV P V and M M M ! We see this by computing the L H 2 L_H^2 L H 2 -adjoint of D H Φ D_H\Phi D H Φ .
Lemma 6.32. Note that the dual of H \mathbb{H} H is H ∗ = H − 1 × L 2 \mathbb{H}^*=H^{-1}\times L^2 H ∗ = H − 1 × L 2 . Equipping T H B T_H\mathcal{B} T H B with the inner product L H 2 L_H^2 L H 2 , the adjoint ( D H Φ ) ∗ : T H B → H ∗ (D_H\Phi)^*:T_H\mathcal{B}\to\mathbb{H}^* ( D H Φ ) ∗ : T H B → H ∗ is given by
( D H Φ ) ∗ ( u , θ , q ) = ( − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 1 2 ∂ 3 [ H ( θ + q ) ] ( 1 + H 2 ) ( θ + q ) ) = ( − P V + 1 2 ∂ 3 ( H M ) ( 1 + H 2 ) M ) . (D_H\Phi)^*(u,\theta,q)=
\begin{pmatrix}
-\nabla_h^\perp\cdot u_h-\partial_3\theta+\frac{1}{2}\partial_3[H(\theta+q)]\\
\left(1+\frac{H}{2}\right)(\theta+q)
\end{pmatrix}
=
\begin{pmatrix}
-PV+\frac{1}{2}\partial_3(HM)\\
\left(1+\frac{H}{2}\right)M
\end{pmatrix}. ( D H Φ ) ∗ ( u , θ , q ) = ( − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 2 1 ∂ 3 [ H ( θ + q )] ( 1 + 2 H ) ( θ + q ) ) = ( − P V + 2 1 ∂ 3 ( H M ) ( 1 + 2 H ) M ) .
Proof. For any ( p , M ) ∈ H (p,M)\in\mathbb{H} ( p , M ) ∈ H and any ( u , θ , q ) ∈ T H B (u,\theta,q)\in T_H\mathcal{B} ( u , θ , q ) ∈ T H B we compute by integrating by parts that
⟨ D H Φ ( p , M ) , ( u , θ , q ) ⟩ L H 2 = ∫ u ⋅ ∇ h ⊥ p + θ [ ∂ 3 p + H 2 ( M − ∂ 3 p ) ] + q H [ M − ∂ 3 p − H 2 ( M − ∂ 3 p ) ] + ( θ + q ) M = ∫ u ⋅ ∇ h ⊥ p + θ ∂ 3 p − H 2 ( θ + q ) ∂ 3 p + ( 1 + H 2 ) ( θ + q ) M = ∫ − ( ∇ h ⊥ ⋅ u h ) p − ( ∂ 3 θ ) p + 1 2 ∂ 3 [ H ( θ + q ) ] p + ( 1 + H 2 ) ( θ + q ) M = ⟨ − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 1 2 ∂ 3 [ H ( θ + q ) ] , p ⟩ H − 1 × H ˙ 1 + ⟨ ( 1 + H 2 ) ( θ + q ) , M ⟩ L 2 = ⟨ ( − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 1 2 ∂ 3 [ H ( θ + q ) ] ( 1 + H 2 ) M ) , ( p M ) ⟩ H ∗ × H . \begin{aligned}
\langle D_H\Phi(p,M),(u,\theta,q)\rangle_{L_H^2}
&=\int u\cdot\nabla_h^\perp p+\theta\left[\partial_3p+\frac{H}{2}(M-\partial_3p)\right]\\
&\qquad+qH\left[M-\partial_3p-\frac{H}{2}(M-\partial_3p)\right]+(\theta+q)M\\
&=\int u\cdot\nabla_h^\perp p+\theta\partial_3p-\frac{H}{2}(\theta+q)\partial_3p+\left(1+\frac{H}{2}\right)(\theta+q)M\\
&=\int-\left(\nabla_h^\perp\cdot u_h\right)p-(\partial_3\theta)p+\frac{1}{2}\partial_3[H(\theta+q)]p+\left(1+\frac{H}{2}\right)(\theta+q)M\\
&=\left\langle-\nabla_h^\perp\cdot u_h-\partial_3\theta+\frac{1}{2}\partial_3[H(\theta+q)],p\right\rangle_{H^{-1}\times\dot{H}^1}\\
&\qquad+\left\langle\left(1+\frac{H}{2}\right)(\theta+q),M\right\rangle_{L^2}\\
&=\left\langle
\begin{pmatrix}
-\nabla_h^\perp\cdot u_h-\partial_3\theta+\frac{1}{2}\partial_3[H(\theta+q)]\\
\left(1+\frac{H}{2}\right)M
\end{pmatrix},
\begin{pmatrix}p\\M\end{pmatrix}
\right\rangle_{\mathbb{H}^*\times\mathbb{H}}.
\end{aligned} ⟨ D H Φ ( p , M ) , ( u , θ , q ) ⟩ L H 2 = ∫ u ⋅ ∇ h ⊥ p + θ [ ∂ 3 p + 2 H ( M − ∂ 3 p ) ] + q H [ M − ∂ 3 p − 2 H ( M − ∂ 3 p ) ] + ( θ + q ) M = ∫ u ⋅ ∇ h ⊥ p + θ ∂ 3 p − 2 H ( θ + q ) ∂ 3 p + ( 1 + 2 H ) ( θ + q ) M = ∫ − ( ∇ h ⊥ ⋅ u h ) p − ( ∂ 3 θ ) p + 2 1 ∂ 3 [ H ( θ + q )] p + ( 1 + 2 H ) ( θ + q ) M = ⟨ − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 2 1 ∂ 3 [ H ( θ + q )] , p ⟩ H − 1 × H ˙ 1 + ⟨ ( 1 + 2 H ) ( θ + q ) , M ⟩ L 2 = ⟨ ( − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 2 1 ∂ 3 [ H ( θ + q )] ( 1 + 2 H ) M ) , ( p M ) ⟩ H ∗ × H .
one-forms. Indeed, if we define, for any ( u , θ , q ) ∈ T H B (u,\theta,q) \in T_H\mathcal{B} ( u , θ , q ) ∈ T H B ,
e p ♭ ( u , θ , q ) : = − P V + 1 2 ∂ 3 ( H M ) e_p^{\flat}(u,\theta,q) := -PV + \frac{1}{2}\partial_3(HM) e p ♭ ( u , θ , q ) := − P V + 2 1 ∂ 3 ( H M )
and
e M ♭ ( u , θ , q ) : = ( 1 + H 2 ) ( θ + q ) e_M^{\flat}(u,\theta,q) := \left(1+\frac{H}{2}\right)(\theta+q) e M ♭ ( u , θ , q ) := ( 1 + 2 H ) ( θ + q )
then we see that, for any ( u , θ , q ) ∈ T H B (u,\theta,q) \in T_H\mathcal{B} ( u , θ , q ) ∈ T H B and any ( π , μ ) ∈ H (\pi,\mu) \in\mathbb{H} ( π , μ ) ∈ H ,
⟨ ( u , θ , q ) , e p ( π ) ⟩ L H 2 = ⟨ e p ♭ ( u , θ , q ) , π ⟩ H ˚ 1 \left\langle(u,\theta,q),e_p(\pi)\right\rangle_{L_H^2}=\left\langle e_p^{\flat}(u,\theta,q),\pi\right\rangle_{\mathring{H}^1} ⟨ ( u , θ , q ) , e p ( π ) ⟩ L H 2 = ⟨ e p ♭ ( u , θ , q ) , π ⟩ H ˚ 1
and
⟨ ( u , θ , q ) , e M ( μ ) ⟩ L H 2 = ⟨ e M ♭ ( u , θ , q ) , M ⟩ L 2 . \left\langle(u,\theta,q),e_M(\mu)\right\rangle_{L_H^2}=\left\langle e_M^{\flat}(u,\theta,q),M\right\rangle_{L^2}. ⟨ ( u , θ , q ) , e M ( μ ) ⟩ L H 2 = ⟨ e M ♭ ( u , θ , q ) , M ⟩ L 2 .
This is depicted in Figure 10 .
Figure 10. A depiction of the dual coordinates e M ♭ = ( 1 + H 2 ) M e_M^{\flat} = \left(1 + \frac{H}{2}\right) M e M ♭ = ( 1 + 2 H ) M and e p ♭ = − P V + 1 2 ∂ 3 ( H M ) e_p^{\flat} = -PV + \frac{1}{2}\partial_3(HM) e p ♭ = − P V + 2 1 ∂ 3 ( H M ) on the formal tangent space T H B T_H\mathcal{B} T H B .
We can see that these coordinates p p p and M M M are actually orthogonal. This is recorded below, in a result that will also come in handy when interpreting the Newton descent iteration geometrically.
Lemma 6.34. For any ( p , M ) ∈ H (p,M) \in\mathbb{H} ( p , M ) ∈ H ,
( D H Φ ) ∗ ( D H Φ ) ( p , M ) = ( − ∇ ⋅ ( A H ∇ p ) ( 1 + H 2 ) M ) (D_H\Phi)^*(D_H\Phi)(p,M)=\begin{pmatrix}-\nabla\cdot(A_H\nabla p)\\ \left(1+\frac{H}{2}\right)M\end{pmatrix} ( D H Φ ) ∗ ( D H Φ ) ( p , M ) = ( − ∇ ⋅ ( A H ∇ p ) ( 1 + 2 H ) M ) for A H : = I − H 2 e 3 ⊗ e 3 A_H := I-\frac{H}{2}e_3\otimes e_3 A H := I − 2 H e 3 ⊗ e 3 (as usual).
Remark 6.35. Note that ( D H Φ ) ∗ ( D H Φ ) = ( − Δ ) ⊕ 1 (D_H\Phi)^*(D_H\Phi)=(-\Delta)\oplus1 ( D H Φ ) ∗ ( D H Φ ) = ( − Δ ) ⊕ 1 when H = 0 H=0 H = 0 , which is the canonical isometry from H = H ˚ 1 × L 2 \mathbb{H}=\mathring{H}^1\times L^2 H = H ˚ 1 × L 2 to its dual H ∗ = H − 1 × L 2 \mathbb{H}^*=H^{-1}\times L^2 H ∗ = H − 1 × L 2 . This is because when H = 0 H=0 H = 0 the dry case may be viewed as a special case of the moist Boussinesq system which corresponds to neglecting q q q (i.e. setting q = 0 q=0 q = 0 ).
Proof of Lemma 6.34 . Let ( p , M ) ∈ H (p,M)\in\mathbb{H} ( p , M ) ∈ H and let us write ( u , θ , q ) : = D H Φ ( p , M ) (u,\theta,q):=D_H\Phi(p,M) ( u , θ , q ) := D H Φ ( p , M ) . Then M ( u , θ , q ) : = θ + q = M M(u,\theta,q):=\theta+q=M M ( u , θ , q ) := θ + q = M while
P V ( u , θ , q ) : = ∇ h ⊥ ⋅ u h + ∂ 3 θ = ∇ ⋅ ( A H ∇ p ) + 1 2 ∂ 3 ( H M ) . PV(u,\theta,q):=\nabla_h^\perp\cdot u_h+\partial_3\theta=\nabla\cdot(A_H\nabla p)+\frac{1}{2}\partial_3(HM). P V ( u , θ , q ) := ∇ h ⊥ ⋅ u h + ∂ 3 θ = ∇ ⋅ ( A H ∇ p ) + 2 1 ∂ 3 ( H M ) .
Therefore
( D H Φ ) ∗ ( D H Φ ) ( p , M ) = ( − P V + 1 2 ∂ 3 ( H M ) ( 1 + H 2 ) M ) = ( − ∇ ⋅ ( A H ∇ p ) − 1 2 ∂ 3 ( H M ) + 1 2 ∂ 3 ( H M ) ( 1 + H 2 ) M ) = ( − ∇ ⋅ ( A H ∇ p ) 1 + H / 2 ) \begin{aligned}
(D_H\Phi)^*(D_H\Phi)(p,M)=
\begin{pmatrix}
-PV+\frac{1}{2}\partial_3(HM)\\
\left(1+\frac{H}{2}\right)M
\end{pmatrix} \\
&=
\begin{pmatrix}
-\nabla\cdot(A_H\nabla p)-\frac{1}{2}\partial_3(HM)+\frac{1}{2}\partial_3(HM)\\
\left(1+\frac{H}{2}\right)M
\end{pmatrix} \\
&=
\begin{pmatrix}
-\nabla\cdot(A_H\nabla p)\\
1+H/2
\end{pmatrix}
\end{aligned} ( D H Φ ) ∗ ( D H Φ ) ( p , M ) = ( − P V + 2 1 ∂ 3 ( H M ) ( 1 + 2 H ) M ) = ( − ∇ ⋅ ( A H ∇ p ) − 2 1 ∂ 3 ( H M ) + 2 1 ∂ 3 ( H M ) ( 1 + 2 H ) M ) = ( − ∇ ⋅ ( A H ∇ p ) 1 + H /2 )
as claimed.
Remark 6.36. Written in block form,
( D H Φ ) ∗ ( D H Φ ) = ( − ∇ ⋅ ( A H ∇ ) 0 0 1 + H / 2 ) (D_H\Phi)^*(D_H\Phi)=\begin{pmatrix}-\nabla\cdot(A_H\nabla)&0\\0&1+H/2\end{pmatrix} ( D H Φ ) ∗ ( D H Φ ) = ( − ∇ ⋅ ( A H ∇ ) 0 0 1 + H /2 )
which is block-diagonal. This means that the coordinates p p p and M M M of the Hilbert manifold B \mathcal{B} B , which arise from the global chart Φ \Phi Φ , may formally be viewed as L H 2 L_H^2 L H 2 –orthogonal (but not orthonormal). In other words, for any Z ∈ T H B Z\in T_H\mathcal{B} Z ∈ T H B , which may be written uniquely as
Z = e p ( π ) + e M ( μ ) Z=e_p(\pi)+e_M(\mu) Z = e p ( π ) + e M ( μ )
for ( π , μ ) ∈ H (\pi,\mu)\in\mathbb{H} ( π , μ ) ∈ H and e p e_p e p and e M e_M e M as in Remark 6.31, we have that
⟨ e p ( π ) , e M ( μ ) ⟩ L H 2 = ∫ 0 + ( 1 − H 2 ) ( ∂ 3 π ) ⋅ H 2 μ − ( 1 − H 2 ) ( ∂ 3 π ) ⋅ ( 1 − H 2 ) μ ⋅ H + 0 = ∫ H 4 ( ∂ 3 π ) μ − H 4 ( ∂ 3 π ) μ = 0. \begin{aligned}
\langle e_p(\pi),e_M(\mu)\rangle_{L_H^2}=\int0+\left(1-\frac{H}{2}\right)(\partial_3\pi)\cdot\frac{H}{2}\mu-\left(1-\frac{H}{2}\right)(\partial_3\pi)\cdot\left(1-\frac{H}{2}\right)\mu\cdot H+0 \\
&=\int\frac{H}{4}(\partial_3\pi)\mu-\frac{H}{4}(\partial_3\pi)\mu=0.
\end{aligned} ⟨ e p ( π ) , e M ( μ ) ⟩ L H 2 = ∫ 0 + ( 1 − 2 H ) ( ∂ 3 π ) ⋅ 2 H μ − ( 1 − 2 H ) ( ∂ 3 π ) ⋅ ( 1 − 2 H ) μ ⋅ H + 0 = ∫ 4 H ( ∂ 3 π ) μ − 4 H ( ∂ 3 π ) μ = 0.
This allows us to refine our comment of Remark 6.31: the decomposition T H B = im e p + im e M T_H\mathcal{B}=\operatorname{im}e_p+\operatorname{im}e_M T H B = im e p + im e M is actually L H 2 L_H^2 L H 2 –orthogonal. Moreover, for the dual one-forms e p b e_p^b e p b and e M b e_M^b e M b as defined in Remark 6.33 we have that
μ = θ + q and π solves ∇ ⋅ ( A H ∇ π ) = ∇ h ⊥ ⋅ u h + ∂ 3 θ . (72) \mu=\theta+q\text{ and }\pi\text{ solves }\nabla\cdot(A_H\nabla\pi)=\nabla_h^\perp\cdot u_h+\partial_3\theta. \tag*{(72)} μ = θ + q and π solves ∇ ⋅ ( A H ∇ π ) = ∇ h ⊥ ⋅ u h + ∂ 3 θ . (72)
Indeed:
( 1 + H 2 ) ( θ + q ) = e M b ( Z ) = e M ( b ( e p ( π ) ) + e M b ( e M ( μ ) ) ) = 0 + ( 1 + H 2 ) μ . \left(1+\frac{H}{2}\right)(\theta+q)=e_M^b(Z)=e_M\bigl(b(e_p(\pi))+e_M^b(e_M(\mu))\bigr)=0+\left(1+\frac{H}{2}\right)\mu. ( 1 + 2 H ) ( θ + q ) = e M b ( Z ) = e M ( b ( e p ( π )) + e M b ( e M ( μ )) ) = 0 + ( 1 + 2 H ) μ .
Similarly: on one hand
e p b ( Z ) = − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 1 2 ∂ 3 [ H ( θ + q ) ] = − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 1 2 ∂ 3 ( H M ) (73) e_p^b(Z)=-\nabla_h^\perp\cdot u_h-\partial_3\theta+\frac{1}{2}\partial_3[H(\theta+q)]=-\nabla_h^\perp\cdot u_h-\partial_3\theta+\frac{1}{2}\partial_3(HM) \tag*{(73)} e p b ( Z ) = − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 2 1 ∂ 3 [ H ( θ + q )] = − ∇ h ⊥ ⋅ u h − ∂ 3 θ + 2 1 ∂ 3 ( H M ) (73)
while on the other hand
e p b ( e p ( π ) ) + e p b ( e M ( μ ) ) = − ∇ ⋅ ( A H ∇ π ) + 1 2 ∂ 3 ( H μ ) = − ∇ ⋅ ( A H ∇ π ) + 1 2 ∂ 3 ( H M ) . (74) e_p^b(e_p(\pi))+e_p^b(e_M(\mu))=-\nabla\cdot(A_H\nabla\pi)+\frac{1}{2}\partial_3(H\mu)=-\nabla\cdot(A_H\nabla\pi)+\frac{1}{2}\partial_3(HM). \tag*{(74)} e p b ( e p ( π )) + e p b ( e M ( μ )) = − ∇ ⋅ ( A H ∇ π ) + 2 1 ∂ 3 ( H μ ) = − ∇ ⋅ ( A H ∇ π ) + 2 1 ∂ 3 ( H M ) . (74)
Equating (73) and (6.28) yields the equation for π \pi π recorded in (72) , as desired.
Remark 6.37. (Comparison of the L H 2 L_H^2 L H 2 –orthogonal decomposition with well-known orthogonal decompositions). In Remark 6.33 above we have explained how any formal tangent vector may be written in a unique way in terms of the “coordinate vectors” e p e_p e p and e M e_M e M . More precisely, we know that for any formal tangent vector Z = ( u , θ , q ) ∈ T H B Z=(u,\theta,q)\in T_H\mathcal{B} Z = ( u , θ , q ) ∈ T H B there exists a unique pair ( π , μ ) ∈ H (\pi,\mu)\in\mathbb{H} ( π , μ ) ∈ H such that
Z = e p ( π ) + e M ( μ ) where { e M b ( e M ( μ ) ) = e M b ( Z ) and e p b ( e p ( π ) ) = e p b ( Z ) (75) Z=e_p(\pi)+e_M(\mu)\quad\text{where}\quad
\begin{cases}
e_M^b(e_M(\mu))=e_M^b(Z)\text{ and}\\
e_p^b(e_p(\pi))=e_p^b(Z)
\end{cases} \tag*{(75)} Z = e p ( π ) + e M ( μ ) where { e M b ( e M ( μ )) = e M b ( Z ) and e p b ( e p ( π )) = e p b ( Z ) (75)
⟺ { μ = θ + q and π solves ∇ ⋅ ( A H ∇ π ) = P V . \Longleftrightarrow
\begin{cases}
\mu=\theta+q\text{ and}\\
\pi\text{ solves }\nabla\cdot(A_H\nabla\pi)=PV.
\end{cases} ⟺ { μ = θ + q and π solves ∇ ⋅ ( A H ∇ π ) = P V .
We will now help the reader make better sense of this by showing how their favourite orthogonal decomposition, namely that of Fourier series, may be written in this way.
Fourier series tell us that every f ∈ L 2 ( T 3 ) f\in L^2(\mathbb{T}^3) f ∈ L 2 ( T 3 ) has uniquely determined coordinates ( a k ) k ∈ ℓ 2 ( Z 3 ) (a_k)_k\in\ell^2(\mathbb{Z}^3) ( a k ) k ∈ ℓ 2 ( Z 3 ) such that
f ( x ) = ∑ k a k e 2 π i k ⋅ x . (76) f(x)=\sum_k a_k e^{2\pi i k\cdot x}. \tag*{(76)} f ( x ) = k ∑ a k e 2 π ik ⋅ x . (76)
Given a function f f f , the process producing its Fourier series can be split into two steps.
We compute the amplitudes a k a_k a k .
We reconstruct f f f from its amplitudes.
Let us rewrite each of these two steps in a manner which will make it easier to directly compare with the L H 2 L_H^2 L H 2 –orthogonal decomposition discussed above.
a k ( f ) = ∫ T 3 f ( x ) e − 2 π i k ⋅ x d x . a_k(f)=\int_{\mathbb{T}^3} f(x)e^{-2\pi i k\cdot x}\,dx. a k ( f ) = ∫ T 3 f ( x ) e − 2 π ik ⋅ x d x .
If we view each a k a_k a k as corresponding to a coordinate direction in L 2 ( T 3 ) L^2(\mathbb{T}^3) L 2 ( T 3 ) we may then encode this coordinate direction as a map e k : C → L 2 ( T 3 ) e_k:\mathbb{C}\to L^2(\mathbb{T}^3) e k : C → L 2 ( T 3 ) given by
[ e k ( a ) ] ( x ) = a e 2 π i k ⋅ x . [e_k(a)](x)=ae^{2\pi i k\cdot x}. [ e k ( a )] ( x ) = a e 2 π ik ⋅ x .
Combining these formulations of the two steps used to produce Fourier series we may write (76) in the form
f = ∑ k e k ( a k ) where a k = ⟨ f , e 2 π i k ⋅ x ⟩ L 2 . (77) f=\sum_k e_k(a_k)\ \text{where}\ a_k=\langle f,e^{2\pi i k\cdot x}\rangle_{L^2}. \tag*{(77)} f = k ∑ e k ( a k ) where a k = ⟨ f , e 2 π ik ⋅ x ⟩ L 2 . (77)
In particular, in the language we used to describe the formal tangent space T H B T_H\mathcal{B} T H B , each “coordinate vector” e k e_k e k has a “dual coordinate vector” e k b e_k^b e k b , which is really just its adjoint and is a map e k b : L 2 ( T 3 ) → C e_k^b:L^2(\mathbb{T}^3)\to\mathbb{C} e k b : L 2 ( T 3 ) → C . Using the L 2 L^2 L 2 inner product on L 2 ( T 3 ) L^2(\mathbb{T}^3) L 2 ( T 3 ) and the standard complex inner product on C \mathbb{C} C defined by ⟨ x , y ⟩ C = x y ‾ \langle x,y\rangle_{\mathbb{C}}=x\overline{y} ⟨ x , y ⟩ C = x y for every x , y ∈ C x,y\in\mathbb{C} x , y ∈ C we may compute these dual coordinates vectors: for any k ∈ Z 3 k\in\mathbb{Z}^3 k ∈ Z 3 , any f ∈ L 2 ( T 3 ) f\in L^2(\mathbb{T}^3) f ∈ L 2 ( T 3 ) , and any a ∈ C a\in\mathbb{C} a ∈ C ,
⟨ e k b ( f ) , a ⟩ C = ⟨ f , e k ( a ) ⟩ L 2 = ∫ T 3 f ( x ) a e 2 π i k ⋅ x ‾ d x = a ‾ ∫ T 3 f ( x ) e − 2 π i k ⋅ x d x = a ‾ a k ( f ) = ⟨ a k ( f ) , a ⟩ C . \langle e_k^b(f),a\rangle_{\mathbb{C}}=\langle f,e_k(a)\rangle_{L^2}=\int_{\mathbb{T}^3}f(x)\overline{ae^{2\pi i k\cdot x}}\,dx=\overline{a}\int_{\mathbb{T}^3}f(x)e^{-2\pi i k\cdot x}\,dx=\overline{a}\,a_k(f)=\langle a_k(f),a\rangle_{\mathbb{C}}. ⟨ e k b ( f ) , a ⟩ C = ⟨ f , e k ( a ) ⟩ L 2 = ∫ T 3 f ( x ) a e 2 π ik ⋅ x d x = a ∫ T 3 f ( x ) e − 2 π ik ⋅ x d x = a a k ( f ) = ⟨ a k ( f ) , a ⟩ C .
In other words: e k b = a k e_k^b=a_k e k b = a k .
We may thus write (77) , already an alternate formulation of (76) , in a manner even more reminiscent of our treatment of coordinate vectors and dual coordinate vectors in T H B T_H\mathcal{B} T H B , namely
f = ∑ k e k ( a k ) where a k = e k b ( f ) . (78) f=\sum_k e_k(a_k)\ \text{where}\ a_k=e_k^b(f). \tag*{(78)} f = k ∑ e k ( a k ) where a k = e k b ( f ) . (78)
Finally, to turn this expression into an exact parallel of our treatment of T H B T_H\mathcal{B} T H B we must use the orthonormality of the Fourier basis. That is, we use the fact that
e k b ( e k ( a ) ) = a k ( e k ( a ) ) = ∫ T 3 a e 2 π i k ⋅ x e − 2 π i k ⋅ x d x = a . e_k^b(e_k(a))=a_k(e_k(a))=\int_{\mathbb{T}^3}ae^{2\pi i k\cdot x}e^{-2\pi i k\cdot x}\,dx=a. e k b ( e k ( a )) = a k ( e k ( a )) = ∫ T 3 a e 2 π ik ⋅ x e − 2 π ik ⋅ x d x = a .
We may therefore rewrite (78) in its final form as
f = ∑ k e k ( a k ) where e k b ( e k ( a k ) ) = e k b ( f ) . (79) f=\sum_k e_k(a_k)\ \text{where}\ e_k^b(e_k(a_k))=e_k^b(f). \tag*{(79)} f = k ∑ e k ( a k ) where e k b ( e k ( a k )) = e k b ( f ) . (79)
This is exactly the same form as (75) !
This section is concerned with the geometry of the balanced set B \mathcal{B} B on its own, and not with how it is embedded in the broader state space L σ 2 L^2_\sigma L σ 2 . Nonetheless, at this stage it is worth commenting on how the L H 2 L^2_H L H 2 inner product endowed on the formal tangent space T H B T_H\mathcal{B} T H B of B \mathcal{B} B interacts with the wave set. The result below is depicted in Figure 11 .
Figure 11. Depiction of Lemma 6.38 : the wave set is L H 2 L_{H}^{2} L H 2 -orthogonal to the formal tangent space to the balanced set.
Lemma 6.38. The wave set W \mathcal{W} W is L H 2 L^2_H L H 2 –orthogonal to the formal tangent space T H B T_H\mathcal{B} T H B in the sense that, for any Z ∈ T H B Z\in T_H\mathcal{B} Z ∈ T H B and any Y ∈ W Y\in\mathcal{W} Y ∈ W , ⟨ Z , Y ⟩ L H 2 = 0 \langle Z,Y\rangle_{L^2_H}=0 ⟨ Z , Y ⟩ L H 2 = 0 .
Proof. Lemma 6.28 tells us that Z = D H Φ ( π , μ ) Z=D_H\Phi(\pi,\mu) Z = D H Φ ( π , μ ) for some ( π , μ ) ∈ H (\pi,\mu)\in\mathbb{H} ( π , μ ) ∈ H while Proposition 3.3 tells us that Y = Ψ ( σ , w ) Y=\Psi(\sigma,w) Y = Ψ ( σ , w ) for some σ ∈ H σ 1 \sigma\in H^1_\sigma σ ∈ H σ 1 and w ∈ L 2 w\in L^2 w ∈ L 2 . Then
⟨ Z , Y ⟩ L H 2 = ⟨ D H Φ ( π , M ) , Ψ ( σ , w ) ⟩ L H 2 = ∫ ∇ h ⊥ π ⋅ σ h ⊥ + ( 1 − H 2 ) ( ∂ 3 π ) ⋅ σ 3 + ( 1 − H 2 ) ( ∂ 3 π ) ⋅ σ 3 ⋅ H + 0 = ∫ ∇ h π ⋅ σ h + ( ∂ 3 π ) σ 3 = ∫ ∇ π ⋅ σ = 0 \begin{aligned}
\langle Z,Y\rangle_{L^2_H}
&=\langle D_H\Phi(\pi,M),\Psi(\sigma,w)\rangle_{L^2_H}\\
&=\int\nabla_h^\perp\pi\cdot\sigma_h^\perp+\left(1-\frac{H}{2}\right)(\partial_3\pi)\cdot\sigma_3+\left(1-\frac{H}{2}\right)(\partial_3\pi)\cdot\sigma_3\cdot H+0\\
&=\int\nabla_h\pi\cdot\sigma_h+(\partial_3\pi)\sigma_3=\int\nabla\pi\cdot\sigma=0
\end{aligned} ⟨ Z , Y ⟩ L H 2 = ⟨ D H Φ ( π , M ) , Ψ ( σ , w ) ⟩ L H 2 = ∫ ∇ h ⊥ π ⋅ σ h ⊥ + ( 1 − 2 H ) ( ∂ 3 π ) ⋅ σ 3 + ( 1 − 2 H ) ( ∂ 3 π ) ⋅ σ 3 ⋅ H + 0 = ∫ ∇ h π ⋅ σ h + ( ∂ 3 π ) σ 3 = ∫ ∇ π ⋅ σ = 0
since σ \sigma σ is divergence-free.
We are now ready to use the geometric language introduced in this section in order to provide a geometric interpretation of the Newton descent iteration presented in § 6.1 .
First, we note that we can rewrite the defining equation of the Newton descent search direction, namely (53) , as follows.
In other words, and this will play a key role in the geometric interpretation of Newton descent: π \pi π is determined by the potential vorticity discrepancy, i.e. by how far the potential vorticity at the current guess is from the true potential vorticity.
We are now ready to describe the geometric interpretation of Newton descent.
Figure 12. The Newton descent method follows curves of constant M M M until the appropriate value of P V PV P V is reached: at a point X \mathcal{X} X on the balanced set, the flow corresponding to Newton descent follows the direction of the formal vector Z \mathcal{Z} Z whose p p p -coordinate is given by the Newton descent search direction π \pi π .
To see that the Newton iteration is a discretisation of this flow we rewrite the Newton iteration in a different way.
The Newton iteration proceed as follows.
P V PV P V and M M M are given.
Initialize the iteration by choosing an arbitrary p 0 ∈ H ˚ 1 p_0 \in\mathring{H}^{1} p 0 ∈ H ˚ 1 .
Iterate, for k ≥ 0 k \ge0 k ≥ 0 ,
Let H k : = 1 ( M < ∂ 3 p k ) H_k := \mathbb{1}(M < \partial_3 p_k) H k := 1 ( M < ∂ 3 p k ) and P V k : = Δ p k + 1 2 ∂ 3 min 0 ( M − ∂ 3 p k ) PV_k := \Delta p_k + \frac{1}{2}\partial_3\min_0(M-\partial_3p_k) P V k := Δ p k + 2 1 ∂ 3 min 0 ( M − ∂ 3 p k ) .
π \pi π solves ∇ ⋅ ( A H k ∇ π ) = P V − P V k \nabla\cdot(A_{H_k}\nabla\pi) = PV-PV_k ∇ ⋅ ( A H k ∇ π ) = P V − P V k .
Let p k + 1 : = p k + π p_{k+1} := p_k+\pi p k + 1 := p k + π .
Note that step 3 is written this way because, once the Newton iteration reaches the so-called “quadratic regime” (which is guaranteed to occur once iterates are sufficiently close to the true solution), then exact line search always produces a unit step size. This is why we can update p k + 1 = p k + t π p_{k+1}=p_k+t\pi p k + 1 = p k + t π with t = 1 t=1 t = 1 .
We are now ready to relate the Newton iteration to the flow above. We may interpret (really, rewrite) the Newton iteration geometrically as follows.
P V PV P V and M M M are given.
Initialize the iteration by choosing an arbitrary p 0 ∈ H ˚ 1 p_0 \in\mathring{H}^{1} p 0 ∈ H ˚ 1 , i.e. choosing an arbitrary X 0 = Φ ( p 0 , M ) ∈ B \mathcal{X}_0=\Phi(p_0,M)\in\mathcal{B} X 0 = Φ ( p 0 , M ) ∈ B with the correct M M M –coordinate.
Iterate, for k ≥ 0 k \ge0 k ≥ 0 .
Let ( p k , M ) : = Φ − 1 ( X k ) (p_k,M):=\Phi^{-1}(\mathcal{X}_k) ( p k , M ) := Φ − 1 ( X k ) and P V k : = P V ( X k ) PV_k:=PV(\mathcal{X}_k) P V k := P V ( X k ) .
π \pi π solves ( D H Φ ) ∗ ( D H Φ ) ( π , 0 ) = ( P V − P V k , 0 ) (D_H\Phi)^*(D_H\Phi)(\pi,0)=(PV-PV_k,0) ( D H Φ ) ∗ ( D H Φ ) ( π , 0 ) = ( P V − P V k , 0 ) .
Let X k + 1 : = Φ ( ( p k , M ) + ( π , 0 ) ) \mathcal{X}_{k+1}:=\Phi((p_k,M)+(\pi,0)) X k + 1 := Φ (( p k , M ) + ( π , 0 )) .
Note that step 2 is the same as above, merely written differently (actually: every step coincides!) Indeed Lemma 6.34 tells us that
( D H Φ ) ∗ ( D H Φ ) ( π , 0 ) = ( − ∇ ⋅ ( A H ∇ π ) , 0 ) . (D_H\Phi)^*(D_H\Phi)(\pi,0)=(-\nabla\cdot(A_H\nabla\pi),0). ( D H Φ ) ∗ ( D H Φ ) ( π , 0 ) = ( − ∇ ⋅ ( A H ∇ π ) , 0 ) .
In other words the Newton iteration may be written in one line as
X k + 1 = Φ ( Φ − 1 ( X k ) + ( π , 0 ) ) \mathcal{X}_{k+1}=\Phi\bigl(\Phi^{-1}(\mathcal{X}_k)+(\pi,0)\bigr) X k + 1 = Φ ( Φ − 1 ( X k ) + ( π , 0 ) )
where π \pi π solves ( D H Φ ) ∗ ( D H Φ ) ( π , 0 ) = ( P V − P V k , 0 ) (D_H\Phi)^*(D_H\Phi)(\pi,0)=(PV-PV_k,0) ( D H Φ ) ∗ ( D H Φ ) ( π , 0 ) = ( P V − P V k , 0 ) .
When X k + 1 \mathcal{X}_{k+1} X k + 1 is close to X k \mathcal{X}_k X k , which equivalently means that π \pi π is small, we may thus formally approximate
X k + 1 ∼ X k + ( D H Φ ) ( π , 0 ) . \mathcal{X}_{k+1}\sim\mathcal{X}_k+(D_H\Phi)(\pi,0). X k + 1 ∼ X k + ( D H Φ ) ( π , 0 ) .
We compute that
( D H Φ ) ( π , 0 ) = ( D H Φ ) [ ( D H Φ ) ∗ ( D H Φ ) ] − 1 ( P V − P V k , 0 ) = ( D H Φ ) − ∗ ( P V − P V ( X k ) , 0 ) \begin{aligned}
(D_H\Phi)(\pi,0)&=(D_H\Phi)[(D_H\Phi)^*(D_H\Phi)]^{-1}(PV-PV_k,0)\\
&=(D_H\Phi)^{-*}(PV-PV(\mathcal{X}_k),0)
\end{aligned} ( D H Φ ) ( π , 0 ) = ( D H Φ ) [( D H Φ ) ∗ ( D H Φ ) ] − 1 ( P V − P V k , 0 ) = ( D H Φ ) −∗ ( P V − P V ( X k ) , 0 )
In other words:
X k + 1 ∼ X k + ( D H Φ ) − ∗ ( P V − P V ( X k ) , 0 ) \mathcal{X}_{k+1}\sim\mathcal{X}_k+(D_H\Phi)^{-*}(PV-PV(\mathcal{X}_k),0) X k + 1 ∼ X k + ( D H Φ ) −∗ ( P V − P V ( X k ) , 0 )
which verifies that indeed the Newton iteration is a formal discretisation of the flow in (80) above.
Remark 6.41 (A simple flow). We discuss here a very simple analog of the flow presented in Remark 6.40 . Consider the following flow on the two-dimensional plane:
x ˙ ( s ) = x ∗ − x ( s ) \dot{x}(s)=x_*-x(s) x ˙ ( s ) = x ∗ − x ( s )
y ˙ ( s ) = 0 \dot{y}(s)=0 y ˙ ( s ) = 0
with initial conditions
where x ∗ x_* x ∗ and y ∗ y_* y ∗ are given (they play the role of the given target values of P V PV P V and M M M ) and where x 0 x_0 x 0 (just like p 0 p_0 p 0 ) may be freely chosen. This means that the x x x –coordinate plays the role of the P V PV P V –coordinate whereas the y y y –coordinate plays the role of the M M M –coordinate.
Then the new unknown
x ~ ( s ) : = x ∗ − x ( s ) \widetilde{x}(s):=x_*-x(s) x ( s ) := x ∗ − x ( s )
satisfies
x ~ ˙ ( s ) = − x ˙ ( s ) = − x ~ ( s ) \dot{\widetilde{x}}(s)=-\dot{x}(s)=-\widetilde{x}(s) x ˙ ( s ) = − x ˙ ( s ) = − x ( s )
and so we deduce that the solution is given by
x ~ ( s ) = x ~ ( 0 ) e − s = ( x ∗ − x 0 ) e − s y ( s ) = y ∗ . \begin{aligned}
\widetilde{x}(s) &= \widetilde{x}(0)e^{-s} = (x_* - x_0)e^{-s}\\
y(s) &= y_*.
\end{aligned} x ( s ) y ( s ) = x ( 0 ) e − s = ( x ∗ − x 0 ) e − s = y ∗ .
In terms of the original unknowns x x x and y y y this yields
x ( s ) = x ∗ − x ~ ( s ) = ( 1 − e − s ) x ∗ + e − s x 0 y ( s ) = y ∗ . \begin{aligned}
x(s) &= x_* - \widetilde{x}(s) = (1-e^{-s})x_* + e^{-s}x_0\\
y(s) &= y_*.
\end{aligned} x ( s ) y ( s ) = x ∗ − x ( s ) = ( 1 − e − s ) x ∗ + e − s x 0 = y ∗ .
Therefore the solution ( x , y ) ( s ) (x,y)(s) ( x , y ) ( s ) approaches the point with coordinates ( x ∗ , y ∗ ) (x_*,y_*) ( x ∗ , y ∗ ) exponentially fast, along a trajectory of constant y y y .
Summary and comparison with Helmholtz and the dry case
We are approaching the conclusion of this paper. Before we conclude, we recall in this section that the decomposition we introduced for the moist Boussinesq system under consideration has many different facets. We summarize these many facets here. In order to make it as easy as possible for the reader to follow this discussion, we first recall the many facets of the Helmholtz decomposition and of the vortical-wave decomposition of the dry Boussinesq system.
This is not merely a recapitulation of the discussion of these two classical systems done in the introduction. Throughout the paper we have introduced a variety of concepts and terminology for the sake of the moist decomposition, but we see that many of these have counterparts in these classical systems. It is these counterparts that we summarize first, in § 7.1 and § 7.2 below, before summarizing the various facets of the moist decomposition in § 7.3 .
We could more prosaically call this section: “the twelve aspects of a decomposition found in both Helmholtz, dry, and moist Boussinesq”.
Helmholtz
We first summarize the various facets of the Helmholtz decomposition.
For any three-dimensional vector field u u u the Helmholtz decomposition is
u = ∇ p + σ u = \nabla p + \sigma u = ∇ p + σ
where σ \sigma σ is a divergence-free vector field.
In the context of the Helmholtz decomposition, if we wish to use terminology from atmospheric science and talk about balanced states , a state u u u will be balanced precisely when it is a solenoidal vector field. In other words the balance in question is
where the subscript B B B denotes the balanced component of u u u .
While introducing this terminology in the context of the Helmholtz decomposition may appear overzealous, it does serve to make the parallels with the dry Boussinesq system, and ultimately with the moist Boussinesq system, more evident.
Since the second component of the Helmholtz decomposition is the balanced component, by analogy with the terminology of this paper we refer to the first component as the wave component 2 . By analogy with Proposition 3.3 we may identify the slow measurement in this setting as the measurement which annihilates the wave set. Since the wave set is the set of potential vector fields this means that the slow measurements are precisely the curl and the spatial average : ∇ p \nabla p ∇ p is the wave component of u u u and
∇ × ∇ p = 0 and ∫ ∇ p = 0. \nabla\times\nabla p = 0 \quad\text{and} \quad\int\nabla p = 0. ∇ × ∇ p = 0 and ∫ ∇ p = 0.
In this case there is no subtlety to identifying the “good unknown” since it is the vector field u u u itself. (See Remark 3.2 for a discussion of what is meant by the term “good unknown” and what role they play.)
The elliptic PDE used to reconstruct the balanced component can be viewed as a div-curl system. Indeed σ \sigma σ solves
∇ × σ = ∇ × u subject to ∇ ⋅ σ = 0 and ∫ σ = ∫ u . \nabla\times\sigma= \nabla\times u \quad\text{subject to} \quad\nabla\cdot\sigma= 0 \quad\text{and} \quad\int\sigma= \int u. ∇ × σ = ∇ × u subject to ∇ ⋅ σ = 0 and ∫ σ = ∫ u .
Note that the forcing is precisely the slow measurements: here the curl and the spatial average are the analogs of the potential vorticity in the dry case since they are slow measurements which characterise the balanced component.
Alternatively we can view this div-curl system as a curl-squared system for the vector potential of σ \sigma σ . Writing σ = c + ∇ × ψ \sigma= c + \nabla\times\psi σ = c + ∇ × ψ we see that c = ∫ u c = \int u c = ∫ u while ψ \psi ψ solves ∇ × ( ∇ × ψ ) = ∇ × u \nabla\times(\nabla\times\psi) = \nabla\times u ∇ × ( ∇ × ψ ) = ∇ × u . (This curl-squared system may be written as a Laplacian equation since we may without loss of generality impose that the vector potential ψ \psi ψ be divergence-free, and in that case the curl-squared operator acting on ψ \psi ψ reduces to a Laplacian.)
The elliptic PDE used to reconstruct the wave component is Laplace’s equation:
Δ p = ∇ ⋅ u . \Delta p = \nabla\cdot u. Δ p = ∇ ⋅ u .
Note that it is not the fast measurement, the divergence, which plays the role of a forcing term for the wave component.
Using the decomposition we can obtain a couple of global changes of coordinates. We can fully describe the vector field u u u in three ways.
(a) Describe it using u u u itself.
(b) Describe it using the measurements ∇ ⋅ u \nabla\cdot u ∇ ⋅ u , ∇ × u \nabla\times u ∇ × u , and ∫ u \int u ∫ u .
(c) Describe it using the coordinates p p p and σ \sigma σ .
Recall that the Helmholtz decomposition may be viewed as a slow-fast decomposition for the low Mach number limit of the compressible Euler equations. In that case the leading-order operator of the dynamics is linear, and it is recorded in Remark A.3 . We bring this up now in order to later compare the leading order operators in the case of dry and moist Boussinesq. For now it suffices to note that, as is discussed in more detail in Remark A.3 , the Helmholtz decomposition is a decomposition into the kernel and the image for that operator: the balanced set is its kernel and the wave set is its image.
As discussed in Remark A.3 , the wave component of the Helmholtz decomposition can be seen to obey a wave equation. Once again, this is not necessarily worthy of note by itself, but we record it here to compare it with the dry and moist Boussinesq decompositions in § 7.2 and § 7.3 below.
When viewed as a slow-fast decomposition, the Helmholtz decomposition inherits the additional feature of being orthogonal with respect to the conserved energy, which happens to simply be the L 2 L^2 L 2 norm.
We note that the Helmholtz decomposition and its induced changes of coordinates discussed above yield Parseval–type identities which allow us to rewrite the energy in three different ways: in terms of the state u u u , in terms of the coordinates, and in terms of the measurements. We obtain that
∫ ∣ u ∣ 2 = ∫ ∣ ∇ p ∣ 2 + ∣ σ ∣ 2 = ∫ ∣ ∇ Δ − 1 ( ∇ ⋅ u ) ∣ 2 + ∣ ∇ × − 1 ( ∇ × u , ∫ f u ) ∣ 2 . \int|u|^2 = \int|\nabla p|^2 + |\sigma|^2 = \int\left|\nabla\Delta^{-1}(\nabla\cdot u)\right|^2 + \left|\nabla\times^{-1}\left(\nabla\times u,\int f u\right)\right|^2. ∫ ∣ u ∣ 2 = ∫ ∣∇ p ∣ 2 + ∣ σ ∣ 2 = ∫ ∇ Δ − 1 ( ∇ ⋅ u ) 2 + ∇ × − 1 ( ∇ × u , ∫ f u ) 2 .
Here we write σ = ( ∇ × ) − 1 ( ω , a ) \sigma=(\nabla\times)^{-1}(\omega,a) σ = ( ∇ × ) − 1 ( ω , a ) to mean that σ \sigma σ solves
∇ × σ = ω subject to ∇ ⋅ σ = 0 and ∫ f σ = a . \nabla\times\sigma=\omega\ \text{subject to}\ \nabla\cdot\sigma=0\ \text{and}\ \int f\sigma=a. ∇ × σ = ω subject to ∇ ⋅ σ = 0 and ∫ f σ = a .
Finally we observe that the extraction of the balanced component of a vector field u u u may be viewed as a projection. Indeed, if we define the balanced set, in the context of the Helmholtz decomposition, to be B H e l . : = ( L 2 ) σ ⊥ \mathcal{B}_{Hel.}:=(L^2)_\sigma^\perp B H e l . := ( L 2 ) σ ⊥ , the space of divergence-free three-dimensional vector fields, then we see that
u B = arg min v ∈ B H e l . ∥ u − v ∥ L 2 . u_B=\arg\min_{v\in\mathcal{B}_{Hel.}}\|u-v\|_{L^2}. u B = arg v ∈ B H e l . min ∥ u − v ∥ L 2 .
Dry Boussinesq
We now turn our attention to the various facets of the vortical-wave decomposition used for the dry Boussinesq system. Note that the numbering below is the same as the numbering in the discussion of the Helmholtz decomposition in § 7.1 above. That is not an accident: it serves to highlight the parallels between the two decompositions. Then again we caution the reader that the roles played by the potential and solenoidal pieces are now flipped: ∇ p \nabla p ∇ p appeared in the fast waves in § 7.1 , it now appears in the slow balanced components; σ \sigma σ appeared in the slow balanced components in § 7.1 , it now appears in the fast waves.
For the dry Boussinesq system discussed in this paper, the decomposition of a state consisting of a vector velocity field u u u and a scalar potential temperature field θ \theta θ takes the form
( u θ ) = ( ∇ h ⊥ p ∂ 3 p ) + ( σ h ⊥ + w e 3 σ 3 ) . \begin{pmatrix} u \\ \theta\end{pmatrix}
=
\begin{pmatrix} \nabla_h^\perp p \\ \partial_3 p \end{pmatrix}
+
\begin{pmatrix} \sigma_h^\perp+ w e_3 \\ \sigma_3 \end{pmatrix}. ( u θ ) = ( ∇ h ⊥ p ∂ 3 p ) + ( σ h ⊥ + w e 3 σ 3 ) .
where σ \sigma σ is divergence-free.
In the dry case, the balanced component satisfies the geostrophic and hydrostatic balances:
u B = ∇ h ⊥ p and u_B=\nabla_h^\perp p\quad\text{and} u B = ∇ h ⊥ p and
θ B = ∂ 3 p . \theta_B=\partial_3p. θ B = ∂ 3 p .
The measurement characterising the wave set is, in the dry case, the potential vorticity. Indeed in this case the wave set is precisely the set of states whose potential vorticity vanishes:
P V ( σ h ⊥ + w e 3 , σ 3 ) = ∇ h ⊥ ⋅ ( σ h ⊥ + w e 3 ) h + ∂ 3 σ 3 = ∇ h ⋅ σ h + ∂ 3 σ 3 = 0 \begin{aligned}
PV(\sigma_h^\perp+w e_3,\sigma_3)=\nabla_h^\perp\cdot(\sigma_h^\perp+w e_3)_h+\partial_3\sigma_3 \\
&=\nabla_h\cdot\sigma_h+\partial_3\sigma_3=0
\end{aligned} P V ( σ h ⊥ + w e 3 , σ 3 ) = ∇ h ⊥ ⋅ ( σ h ⊥ + w e 3 ) h + ∂ 3 σ 3 = ∇ h ⋅ σ h + ∂ 3 σ 3 = 0
since σ \sigma σ is divergence-free.
In the dry case the good unknown is the following three-dimensional vector field combining both the horizontal velocity and the potential temperature:
− u h ⊥ + θ e 3 . -u_h^\perp+\theta e_3. − u h ⊥ + θ e 3 .
The elliptic equation used to recover the pressure p p p , and hence the balanced component, is the Laplace/Poisson equation
The wave component is recovered by setting w = u 3 w=u_3 w = u 3 and finding σ \sigma σ to be the solution of the following div-curl system:
∇ × σ = ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 = j + ( ∂ 3 w ) e 3 \nabla\times\sigma=\partial_3u_h-\nabla_h^\perp\theta+(\partial_3w)e_3=j+(\partial_3w)e_3 ∇ × σ = ∂ 3 u h − ∇ h ⊥ θ + ( ∂ 3 w ) e 3 = j + ( ∂ 3 w ) e 3
subject to ∇ ⋅ σ = 0 \nabla\cdot\sigma=0 ∇ ⋅ σ = 0 and
∫ σ = − ∫ u h ⊥ + ∫ θ e 3 = a . \int\sigma=-\int u_h^\perp+\int\theta e_3=a. ∫ σ = − ∫ u h ⊥ + ∫ θ e 3 = a .
We may describe states of the dry Boussinesq system in three equivalent ways.
(a) Describe them using u u u and θ \theta θ .
(b) Describe them using the measurements P V PV P V , j j j , w w w , and a a a .
(c) Describe them using the coordinates p p p , σ \sigma σ , and w w w .
The leading-order operator of the dynamics of the dry Boussinesq system considered in this paper is given by
L ( u θ ) = ( P L ( u h ⊥ − θ e 3 ) u 3 ) . \mathcal{L}\begin{pmatrix}u\\\theta\end{pmatrix}
=
\begin{pmatrix}\mathbb{P}_{L}\left(u_h^\perp-\theta e_3\right)\\u_3\end{pmatrix}. L ( u θ ) = ( P L ( u h ⊥ − θ e 3 ) u 3 ) .
The decomposition is adapted to this operator since the balanced set is its kernel and the wave set is its image.
We may compute immediately that the wave component of a state undergoes linear oscillations under the leading-order operator since, when restricted to the wave set, the square of the operator behaves like multiplication by − 1 -1 − 1 :
L 2 ( σ h ⊥ + w e 3 σ 3 ) = − ( σ h ⊥ + w e 3 σ 3 ) . \mathcal{L}^{2}\begin{pmatrix}\sigma_h^\perp+w e_3\\\sigma_3\end{pmatrix}
=
-\begin{pmatrix}\sigma_h^\perp+w e_3\\\sigma_3\end{pmatrix}. L 2 ( σ h ⊥ + w e 3 σ 3 ) = − ( σ h ⊥ + w e 3 σ 3 ) .
The conserved energy in the dry case is simply the L 2 L^2 L 2 norm given by
E = ∫ ∣ u ∣ 2 + θ 2 . E=\int|u|^2+\theta^2. E = ∫ ∣ u ∣ 2 + θ 2 .
The Parseval-type identity in the dry case lets us rewrite the energy in terms of the coordinates or the measurements as follows:
∫ ∣ u ∣ 2 + θ 2 = ∫ ∣ ∇ p ∣ 2 + ∣ σ ∣ 2 + w 2 = ∫ ∣ ∇ Δ − 1 P V ∣ 2 + ∣ ∇ × − 1 ( j + ( ∂ 3 w ) e 3 , a ) ∣ 2 . \int|u|^2+\theta^2
=
\int|\nabla p|^2+|\sigma|^2+w^2
=
\int|\nabla\Delta^{-1}PV|^2+
|\nabla\times^{-1}(j+(\partial_3w)e_3,a)|^2. ∫ ∣ u ∣ 2 + θ 2 = ∫ ∣∇ p ∣ 2 + ∣ σ ∣ 2 + w 2 = ∫ ∣∇ Δ − 1 P V ∣ 2 + ∣∇ × − 1 ( j + ( ∂ 3 w ) e 3 , a ) ∣ 2 .
For the dry Boussinesq case the balanced set takes the form
B d r y : = { ( u , θ ) ∈ ( L σ 2 ) 3 × L 2 : u = ∇ h ⊥ p , u 3 = 0 , and θ = ∂ 3 p for some p ∈ H ˙ 1 } . \mathcal{B}_{dry}
:=
\left\{
(u,\theta)\in(L^2_\sigma)^3\times L^2:
u=\nabla_h^\perp p,\ u_3=0,\ \text{and }\theta=\partial_3p
\text{ for some }p\in\dot{H}^{1}
\right\}. B d r y := { ( u , θ ) ∈ ( L σ 2 ) 3 × L 2 : u = ∇ h ⊥ p , u 3 = 0 , and θ = ∂ 3 p for some p ∈ H ˙ 1 } .
Therefore the extraction of a balanced component once again has a projection perspective:
( u B , θ B ) = arg min Y ∈ B d r y ∥ ( u , θ ) − Y ∥ L 2 . (u_B,\theta_B)=\underset{\mathcal{Y}\in\mathcal{B}_{dry}}{\arg\min}\|(u,\theta)-\mathcal{Y}\|_{L^2}. ( u B , θ B ) = Y ∈ B d r y arg min ∥ ( u , θ ) − Y ∥ L 2 .
Moist Boussinesq
Finally we perform the same exercise as in § 7.1 and § 7.2 : we list the twelve aspects of the moist decomposition presented in this paper, in the same order as the above sections to make the comparison easy to follow.
Recall that here, in a manner consistent with the rest of this paper (barring the introduction) we write θ \theta θ to denote the equivalent potential temperature (denoted θ e \theta_e θ e in the introduction).
The decomposition of a state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) is given by
( u θ q ) = ( ∇ h ⊥ p ∂ 3 p + 1 2 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 1 2 min 0 ( M − ∂ 3 p ) ) + ( σ h ⊥ + w e 3 σ 3 − σ 3 ) . \begin{pmatrix} u \\ \theta\\ q \end{pmatrix}
=
\begin{pmatrix}
\nabla_h^\perp p \\
\partial_3p+\frac{1}{2}\min_0(M-\partial_3p) \\
M-\partial_3p-\frac{1}{2}\min_0(M-\partial_3p)
\end{pmatrix}
+
\begin{pmatrix}
\sigma_h^\perp+w e_3 \\
\sigma_3 \\
-\sigma_3
\end{pmatrix}. u θ q = ∇ h ⊥ p ∂ 3 p + 2 1 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 2 1 min 0 ( M − ∂ 3 p ) + σ h ⊥ + w e 3 σ 3 − σ 3 .
where p p p has vanishing average and σ \sigma σ is divergence-free.
In the moist case the balanced component satisfies the geostrophic and hydrostatic balances, meaning that
u B = ∇ h ⊥ p u_B=\nabla_h^\perp p u B = ∇ h ⊥ p
and
θ B − min 0 q B = ∂ 3 p . \theta_B-\min_0 q_B=\partial_3p. θ B − 0 min q B = ∂ 3 p .
To fully characterise the balanced component we must also include the moist variable M M M and the following “auxiliary” balance:
θ B + q B = M . \theta_B+q_B=M. θ B + q B = M .
The slow measurements characterising the wave set are now P V PV P V and M M M , and indeed
P V ( σ h ⊥ + w e 3 , σ 3 , − σ 3 ) = ∇ ⋅ σ = 0 PV(\sigma_h^\perp+w e_3,\sigma_3,-\sigma_3)=\nabla\cdot\sigma=0 P V ( σ h ⊥ + w e 3 , σ 3 , − σ 3 ) = ∇ ⋅ σ = 0
and
M ( σ h ⊥ + w e 3 , σ 3 , − σ 3 ) = σ 3 − σ 3 = 0. M(\sigma_h^\perp+w e_3,\sigma_3,-\sigma_3)=\sigma_3-\sigma_3=0. M ( σ h ⊥ + w e 3 , σ 3 , − σ 3 ) = σ 3 − σ 3 = 0.
There are two good unknowns in the moist case, owing to the fact that the leading-order operator is not skew-symmetric (this is discussed in more detail in Remark 3.8). One good unknown is
− u h ⊥ + ( θ − min 0 q ) e 3 , -u_h^\perp+(\theta-\min_0 q)e_3, − u h ⊥ + ( θ − 0 min q ) e 3 ,
which is related to j j j and the null set of the leading-order operator, while its “adjoint” good unknown is
− u h ⊥ + θ e 3 , -u_h^\perp+\theta e_3, − u h ⊥ + θ e 3 ,
which is related to P V PV P V and the image of the leading-order operator.
In order to reconstruct the balanced component we set M = θ + q M=\theta+q M = θ + q and then let p p p be the solution of nonlinear P V PV P V -and-M M M inversion
Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) = P V . \Delta p+\frac{1}{2}\partial_3\min_0(M-\partial_3p)=PV. Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) = P V .
To reconstruct the wave component we let w = u 3 w=u_3 w = u 3 and let σ \sigma σ solve
∇ × ( A B − 1 σ ) = ∂ 3 u h − ∇ h ⊥ ( θ − θ H B ) + ( ∂ 3 w ) e 3 = : j B + ( ∂ 3 w ) e 3 \nabla\times(A_B^{-1}\sigma)=\partial_3u_h-\nabla_h^\perp(\theta-\theta_{H_B})+(\partial_3w)e_3=:j_B+(\partial_3w)e_3 ∇ × ( A B − 1 σ ) = ∂ 3 u h − ∇ h ⊥ ( θ − θ H B ) + ( ∂ 3 w ) e 3 =: j B + ( ∂ 3 w ) e 3
subject to ∇ ⋅ σ = 0 \nabla\cdot\sigma=0 ∇ ⋅ σ = 0 and
∫ A B − 1 σ = − ∫ u h ⊥ + ∫ ( θ − q H B ) e 3 , \int A_B^{-1}\sigma=-\int u_h^\perp+\int(\theta-qH_B)e_3, ∫ A B − 1 σ = − ∫ u h ⊥ + ∫ ( θ − q H B ) e 3 ,
where A B − 1 = I + H B e 3 ⊗ e 3 A_B^{-1}=I+H_B e_3\otimes e_3 A B − 1 = I + H B e 3 ⊗ e 3 for H B : = 1 ( M < ∂ 3 p ) H_B:=\mathbf{1}(M<\partial_3p) H B := 1 ( M < ∂ 3 p ) .
We can describe moist Boussinesq states in three ways.
(a) Describe them using u u u , θ \theta θ , and q q q .
(b) Describe them using the measurements P V PV P V , M M M , j j j , w w w , and a a a .
(c) Describe them using the coordinates p p p , M M M , σ \sigma σ , and w w w .
The leading-order operator of the dynamics in the moist case is
N ( u θ q ) = ( P L [ u h ⊥ − ( θ − min 0 q ) e 3 ] u 3 − u 3 ) . \mathcal{N}
\begin{pmatrix}
u\\
\theta\\
q
\end{pmatrix}
=
\begin{pmatrix}
\mathbb{P}_L\left[u_h^\perp-(\theta-\min_0q)e_3\right]\\
u_3\\
-u_3
\end{pmatrix}. N u θ q = P L [ u h ⊥ − ( θ − min 0 q ) e 3 ] u 3 − u 3 .
As detailed in § 4 the operator of interest acting on the wave set, denoted by L W , H B \mathcal{L}_{W,H_B} L W , H B , is the linearisation of N \mathcal{N} N about a balanced state with phases characterised by H B H_B H B . We then observe that the restriction of this operator to the wave set produces oscillatory dynamics. By contrast with the case of the Helmholtz decomposition and of the dry Boussinesq system, the frequencies of these oscillations are phase-dependent :
L W , H B 2 ( σ h 0 0 ) = − ( σ h 0 0 ) \mathcal{L}_{W,H_B}^{2}
\begin{pmatrix}
\sigma_h\\
0\\
0
\end{pmatrix}
=
-
\begin{pmatrix}
\sigma_h\\
0\\
0
\end{pmatrix} L W , H B 2 σ h 0 0 = − σ h 0 0
and
L W , H B 2 ( w e 3 σ 3 − σ 3 ) = − ( 1 + H B ) ( w e 3 σ 3 − σ 3 ) . \mathcal{L}_{W,H_B}^{2}
\begin{pmatrix}
w e_3\\
\sigma_3\\
-\sigma_3
\end{pmatrix}
=
-(1+H_B)
\begin{pmatrix}
w e_3\\
\sigma_3\\
-\sigma_3
\end{pmatrix}. L W , H B 2 w e 3 σ 3 − σ 3 = − ( 1 + H B ) w e 3 σ 3 − σ 3 .
In the moist case the conserved energy is no longer quadratic. It is now nonlinear and given by
E = ∫ ( ∣ u ∣ 2 + θ 2 + min 0 2 q + ( θ + q ) 2 ) . E = \int\left(|u|^2 + \theta^2 + \min_0^2 q + (\theta+ q)^2\right). E = ∫ ( ∣ u ∣ 2 + θ 2 + 0 min 2 q + ( θ + q ) 2 ) .
In the moist case there is no simple Parseval-type identity which holds: the entirety of § 5 is devoted to finding an alternative! Ultimately we identify in § 5.3 a statistical divergence given by
D ( X 1 ; X 2 ) = 1 2 ∫ ( A H 1 ∇ p ⋅ ∇ p + ( 1 + H 1 2 ) ( M 1 − M 2 ) 2 + A H 1 − 1 σ ⋅ σ + ( w 1 − w 2 ) 2 ) . D(\mathcal{X}_1;\mathcal{X}_2) = \frac{1}{2}\int\left(A_{H_1}\nabla p \cdot\nabla p + \left(1+\frac{H_1}{2}\right)(M_1-M_2)^2 + A_{H_1}^{-1}\sigma\cdot\sigma+ (w_1-w_2)^2\right). D ( X 1 ; X 2 ) = 2 1 ∫ ( A H 1 ∇ p ⋅ ∇ p + ( 1 + 2 H 1 ) ( M 1 − M 2 ) 2 + A H 1 − 1 σ ⋅ σ + ( w 1 − w 2 ) 2 ) .
where, for H 1 : = 1 ( q B , 1 < 0 ) H_1 := \mathbf{1}_{(q_{B,1}<0)} H 1 := 1 ( q B , 1 < 0 ) ,
p p p solves ∇ ⋅ ( A H 1 ∇ p ) = P V 1 − P V 2 − 1 2 ∂ 3 [ H 1 ( M 1 − M 2 ) ] \nabla\cdot(A_{H_1}\nabla p)=PV_1-PV_2-\frac{1}{2}\partial_3[H_1(M_1-M_2)] ∇ ⋅ ( A H 1 ∇ p ) = P V 1 − P V 2 − 2 1 ∂ 3 [ H 1 ( M 1 − M 2 )] and
σ \sigma σ solves ∇ × ( A H 1 − 1 σ ) = j 1 − j 2 + ∂ 3 ( w 1 − w 2 ) e 3 \nabla\times(A_{H_1}^{-1}\sigma)=j_1-j_2+\partial_3(w_1-w_2)e_3 ∇ × ( A H 1 − 1 σ ) = j 1 − j 2 + ∂ 3 ( w 1 − w 2 ) e 3 subject to ∇ ⋅ σ = 0 \nabla\cdot\sigma=0 ∇ ⋅ σ = 0 and ∫ A H 1 − 1 σ = a 1 − a 2 \int A_{H_1}^{-1}\sigma=a_1-a_2 ∫ A H 1 − 1 σ = a 1 − a 2 .
The projection perspective in the moist case takes the following form:
X B = arg min Y ∈ B D ( X ; Y ) . \mathcal{X}_B=\underset{\mathcal{Y}\in\mathcal{B}}{\operatorname{arg\,min}}\,D(\mathcal{X};\mathcal{Y}). X B = Y ∈ B arg min D ( X ; Y ) .
Above we have summarized in detail how the many aspects of the Helmholtz, dry, and moist Boussinesq decompositions are analogous to one another, highlighting their differences. Below we provide a table summarizing § 7.1 –§ 7.3 and which may serve either as a reference or as a more visual way to compare and contrast the various aspects of each decomposition.
Conclusion
In this section we discuss the main contributions of this paper and the main takeaways for the reader.
The core contribution of this paper is the following. We provided a blueprint for decomposing the solutions of ∂ t + N \partial_t+\mathcal{N} ∂ t + N , specifically when N \mathcal{N} N is nonlinear. Recall that otherwise, if N \mathcal{N} N is linear, then we can proceed via spectral methods and identify eigenvalues and eigenvectors of that operator.
This blueprint requires a good understanding of the null set and the image of N \mathcal{N} N , which we refer to as the balanced set and the wave set in the context of this paper.
We actually provide two descriptions of each of these sets: (1) an explicit parametrisation and (2) a characterisation as the joint zero set of some measurements. Note that the measurements annihilating the wave set are known as slow measurements while those annihilating the balanced set are known as fast measurements.
A crucial condition to check in order to follow this general blueprint is that the slow measurements be characteristic of balanced states. In other words: there must be only one balanced set with any given P V PV P V and M M M . Geometrically this is asking that the wave set (i.e. the joint zero set of P V PV P V and M M M ) be transversal to the balanced set.
This allows us to define a global decomposition of state space:
L σ 2 = B + W . \mathcal{L}^2_\sigma= \mathcal{B}+\mathcal{W}. L σ 2 = B + W .
In particular this decomposition gives broader meaning to nonlinear P V PV P V -and-M M M inversion. Previously, this inversion was understood in the context of balanced states: given P V PV P V and M M M , the inversion was used to produce the corresponding balanced state. Now this inversion has relevance for any state: it is used to produce its balanced component .
We then discuss three aspects of this decomposition.
The first aspect we discuss is its interpretation as a slow-fast decomposition. We obtain the three-part decomposition
X = X B ⏟ balanced + X ‾ W ⏟ wave + ( X W − X ‾ W ) ⏟ fast (81) \mathcal{X}=\underbrace{\mathcal{X}_B}_{\text{balanced}}+\underbrace{\overline{\mathcal{X}}_W}_{\text{wave}}+\underbrace{(\mathcal{X}_W-\overline{\mathcal{X}}_W)}_{\text{fast}} \tag*{(81)} X = balanced X B + wave X W + fast ( X W − X W ) (81)
where X ‾ W \overline{\mathcal{X}}_W X W denotes the time average of the wave component.
Contrast this with the dry case: in that case the wave component undergoes linear oscillations, their time averages thus vanish, and so they do not contribute to the slow component at lowest order (they do contribute, a little, when the small parameter ϵ \epsilon ϵ is finite). Here, in the moist case, the wave component undergoes nonlinear oscillations and so its time average contributes to the slow component.
Table 65.
The second aspect of the decomposition we discuss is its connection to the conserved energy. (Recall: when the operator N \mathcal{N} N is linear, the eigendecomposition obtained often has favourable orthogonality properties, e.g. with respect to the inner product generated by the conserved energy as is the case in the case for the dry Boussinesq system.) We obtain a statistical divergence which quantifies the discrepancy between states in a manner consistent with both the conserved energy and the decomposition.
This statistical divergence is inpired by the Parseval-type identities which hold for the Helmholtz decomposition and the dry Boussinesq decomposition. For example in the dry case we ca write the energy either in terms of the state ( u , θ ) (u,\theta) ( u , θ ) or in terms of the measurements ( P V , j , w , a ) (PV,j,w,a) ( P V , j , w , a ) . We proceed similarly in the moist case: we use the nonlinear change of variables from the state ( u , θ , q ) (u,\theta,q) ( u , θ , q ) to the measurements ( P V , M , j , w , a ) (PV,M,j,w,a) ( P V , M , j , w , a ) to define the statistical divergence in terms of these measurements.
The third aspect of the decomposition we discuss is how to compute it in practice. We introduce iterative methods that may be used to compute the decomposition. (These methods produce the balanced component X B X_B X B and the wave component may then be obtained as a residual X W : = X − X B X_W := X-X_B X W := X − X B .) One of these methods, the Newton descent method, has the added benefit of shedding light on past work on this moist Boussinesq system.
This Newton descent iteration corresponds to a previously used numerical method. Our analysis here thus provides the first rigorous proof that this numerical method converges.
This Newton descent iteration may also be interpreted geometrically: it is precisely the discretisation of the flow along the balanced set where the moist variable M M M is kept constant and where we follow the potential vorticity coordinate direction until the desired potential vorticity is attained.
Table 1 (PDF p. 66) below serves as an easy reference for the main results of this paper. This table also lists a few important auxiliary results, remarks, and definitions which either highlight the key ideas underpinning the main results or provide essential context for those results.
Table 1 (PDF p. 66). The main results of this paper and some important auxiliary results.
We now pause to highlight the main difficulty in constructing a decomposition with the favourable properties discussed above. Naively, the difficulty may appear to be the nonlinearities or the lack of skew-symmetry. Both of these challenges come from the phase boundaries, and so the real difficulty is more specifically this: the phase boundaries are dynamic , meaning that they evolve over time.
To explain this we consider what would happen if this difficulty were removed. That is, suppose that ( u , θ , q ) (u,\theta,q) ( u , θ , q ) is a solution of the moist Boussinesq system (2.7a)–(2.7c) whose phase boundaries do not evolve over time. In other words, if H : = 1 ( q < 0 ) H := \mathbf{1}(q<0) H := 1 ( q < 0 ) , suppose that ∂ t H = 0 \partial_t H=0 ∂ t H = 0 .
Then we may treat H H H as a fixed indicator function and observe that N ( u , θ , q ) = L H ( u , θ , q ) \mathcal{N}(u,\theta,q)=\mathcal{L}_H(u,\theta,q) N ( u , θ , q ) = L H ( u , θ , q ) . Crucially: L H \mathcal{L}_H L H is L H 2 L^2_H L H 2 -skew-symmetric (see Lemma A.11 ). This means that, when the phase boundaries are static , the leading-order operator becomes both linear and skew-symmetric. Both symptomatic difficulties (nonlinearities and lack of skew-symmetry) are removed when the root difficulty (dynamic phase boundaries) is removed.
Looking forward
In this section we record some thoughts on future research direction stemming from the topics discussed here.
Can we prove that the PQG system (see [92 ] ), or a variant thereof is the rigorous fast-wave averaging limit of the moist Boussinesq system, in the limit of fast rotation and strong stratification (i.e. as both the Froude and Rossby numbers approach zero, at the same rate)?
In the case of balanced, or well-prepared initial data we expect to be obtain to obtain the PQG system as a rigorous limit of the moist Boussinesq system discussed here. However, in the case of unbalanced, or ill-prepared initial data, there is mounting evidence that the appropriate limiting system is not the PQG system. The crux of the matter are the nonlinear oscillations of the fast waves: for ill-prepared data these waves are present and in the system and, in the moist case, they feed back into the slow components through their average. We thus expect that, at the very least, a modification needs to be made to the PQG system if we hope to obtain it as a rigorous limit of the moist Boussinesq system for generic initial data.
Since this is a challenging problem, and here are two more restricted questions to look at first and whose resolution will shed light on the general case.
Treat the ODE case first. The ODE problem is recorded in Lemma A.23 , which also shows that ODE solutions are PDE solutions. Crucially: this ODE system possesses one of the key difficulties found in the PDE case: its wave solutions, i.e. its fast components, are nonlinear oscillations (which are piecewise linear oscillations, meaning that depending on the sign of q q q they have fixed frequencies, but their frequencies change as the sign of q q q changes) with non-zero averages.
Treat the case of steady interfaces. In other words, look at solutions to the full moist Boussinesq system whose interface (i.e. the zero level set of q q q ) is fixed in time and does not evolve dynamically, even though q q q itself evolves dynamically. Such solutions are readily obtained by considering velocity fields whose support is localized away from the interface. In this case, the Heaviside H = 1 ( q < 0 ) H = \mathbb{1}(q < 0) H = 1 ( q < 0 ) which appears in the moist Boussinesq system can be considered to be fixed, and so the leading-order operator is linear! This removes a key difficulty from the general, nonlinear, moist Boussinesq system due to dynamically-evolving phase boundaries. It is important to note what we expect to glean from studying this “fixed H H H ” case. Indeed, if we consider this example from a physical perspective it is severely contrived and hence there is little hope of extracting meaningful physical intuition from this test case. Nonetheless, studying this “fixed H H H ” case may prove helpful in isolating the mathematical difficulties one at a time, and this is where we may benefit from studying this simpler case first before handling the general case.
Note that carrying out the fast-wave averaging process will require a precise identification of the solvability condition for the next-order equations, something which we have not carried out, even formally, in this paper. Recall that this next-order solvability condition is used to determine the behaviour of the “background profile” at the slow timescale which arises from the solution of the leading-order equation.
In § 6.2 we discussed how the iterative method used to solve nonlinear P V PV P V -and-M M M inversion numerically could be interpreted geometrically. This interpretation viewed the iterative method as the discretisation of a flow along the balanced set. It may be possible to re-interpret this same iterative method in a different way as follows. First, the detailed geometric understanding of the balanced set, in terms of a chart, coordinate vectors, and dual coordinate vectors, would need to be extended to the entire state space, thus describing both the balanced set and the wave set. We may then be able to re-interpret this same iterative method geometrically as a flow in state space which starts at some (possibly unbalanced) initial state and flows towards the balanced set, only reaching it when the iterative method terminates. The appeal of such an interpretation would be that it
matches the following physical intuition: a balanced state is obtained by starting from any state and then progressively removing wave components (i.e. imbalances) until we reach the balanced set.
The two metrics/distance functions defined in § 5.1 and § 5.2 may potentially be used for other applications, such as adjoint models or data-driven models. The “fixed H H H ” energy from § 5.1 appears in situations where linearisation is used, and hence it could find application to adjoint models, sensitivity analysis, and variational data assimilation, which rely on tangent linear approximations to the dynamics [65 , 85 , 37 , 36 , 70 , 5 , 1 , 30 ] . The “fixed H H H ” energy also appears here in the numerical/iterative methods, in § 6.2 where geometrical notions relying on tangent spaces (a.k.a. linearisations) are discussed. The Snell-like distance in § 5.2 here could be useful for applications such as data-driven methods that identify nonlinear features in dynamical systems and utilize a metric, distance function, and/or kernel function [19 , 44 , 9 , 32 , 42 ] .
Can the ocean (or other geophysical or astrophysical fluids) have a nonlinear eigenspace decomposition, similar to the atmospheric decomposition presented here? The nonlinearity in the buoyancy here was a nonlinear switch due to clouds, but the blueprint of the techniques here may also be applicable for other types of nonlinearity. For instance, in oceanography, a nonlinear equation of state describes the nonlinear dependence of oceanic buoyancy on temperature and salinity, and plays an important role in a variety of processes [KM10, HN13, NHR15, RMBN15, Kor21]. Also, PV inversion is used in oceanography as well [Mül95, OSB+01], and it would be interesting to investigate whether the nonlinear eigenspace decomposition here may be applicable for incorporating the effects of a nonlinear equation of state for seawater.
Acknowledgments
The authors are grateful for the support of NSF DMS-1907667. A.R.-T. gratefully acknowledges support from the Simons Foundation, the Isaac Newton Institute for Mathematical Sciences, and the Department of Applied Mathematics and Theoretical Physics of the University of Cambridge where part of this work was carried out.
Appendix
Helmholtz decomposition
In this section we record a precise statement of the Helmholtz decomposition. This is a well-known result which plays an important role in this paper and so we record it here, if only to record it in a manner with the notation consistent with the notation in the rest of this paper.
Theorem A.1 (Helmholtz). For every u ∈ L 2 u \in L^2 u ∈ L 2 there exists a unique p ∈ H ˙ 1 p \in\dot{H}^{1} p ∈ H ˙ 1 and a unique σ ∈ L σ 2 \sigma\in L^2_{\sigma} σ ∈ L σ 2 such that u = σ + ∇ p u = \sigma+ \nabla p u = σ + ∇ p .
Proof. We define p p p to be the solution of ∇ p = ∇ ⋅ u \nabla p = \nabla\cdot u ∇ p = ∇ ⋅ u , ψ \psi ψ to be the solution of ∇ × ( ∇ × ψ ) = ∇ × u \nabla\times(\nabla\times\psi) = \nabla\times u ∇ × ( ∇ × ψ ) = ∇ × u , and let c = ∫ u c = \int u c = ∫ u . Then u = ∇ p + σ u = \nabla p + \sigma u = ∇ p + σ for σ = ∇ × ψ + c \sigma= \nabla\times\psi+ c σ = ∇ × ψ + c (as can be directly verified using the Fourier transform since here we are working on the three-dimensional torus). □
In particular we will use the following corollary of the Helmholtz decomposition.
Corollary A.2. For every u ∈ L 2 u \in L^2 u ∈ L 2 , if ∇ × u = 0 \nabla\times u = 0 ∇ × u = 0 and ∫ u = 0 \int u = 0 ∫ u = 0 then there exists a unique p ∈ H ˙ 1 p \in\dot{H}^{1} p ∈ H ˙ 1 such that u = ∇ p u = \nabla p u = ∇ p .
Proof. It follows from the Helmholtz decomposition that u = ∇ p + σ u = \nabla p + \sigma u = ∇ p + σ for some uniquely determined p p p and σ \sigma σ . Since σ \sigma σ is uniquely determined by the curl and the average of u u u , which both vanish, we see that indeed σ = 0 \sigma= 0 σ = 0 . □
To conclude this section we discuss, as was alluded to in the introduction, how the Helmholtz decomposition may be viewed as a slow-fast decomposition for the compressible Euler equations.
is conserved since
∫ γ p ˉ 1 + 1 / γ u ⋅ ∂ t u + q ∂ t q = ∫ p ˉ 1 + 1 / γ u ⋅ ( − 1 p ˉ 1 / γ ∇ q ) + q ⋅ ( − γ p ˉ ∇ ⋅ u ) = − γ p ˉ ∫ [ u ⋅ ∇ q + q ( ∇ ⋅ u ) ] = 0. \int\gamma\bar{p}^{1+1/\gamma}u\cdot\partial_tu+q\partial_tq
=\int\bar{p}^{1+1/\gamma}u\cdot\left(-\frac{1}{\bar{p}^{1/\gamma}}\nabla q\right)+q\cdot(-\gamma\bar{p}\nabla\cdot u)
=-\gamma\bar{p}\int\left[u\cdot\nabla q+q(\nabla\cdot u)\right]=0. ∫ γ p ˉ 1 + 1/ γ u ⋅ ∂ t u + q ∂ t q = ∫ p ˉ 1 + 1/ γ u ⋅ ( − p ˉ 1/ γ 1 ∇ q ) + q ⋅ ( − γ p ˉ ∇ ⋅ u ) = − γ p ˉ ∫ [ u ⋅ ∇ q + q ( ∇ ⋅ u ) ] = 0.
We now set γ = p ˉ = 1 \gamma=\bar{p}=1 γ = p ˉ = 1 for simplicity, reducing the leading order system to
∂ t q + ∇ ⋅ u = 0 (88) \partial_tq+\nabla\cdot u=0 \tag*{(88)} ∂ t q + ∇ ⋅ u = 0 (88)
∂ t u + ∇ q = 0 (89) \partial_tu+\nabla q=0 \tag*{(89)} ∂ t u + ∇ q = 0 (89)
which we may rewrite as
∂ t ( q u ) = L ( q u ) where L = ( 0 div grad 0 ) . \partial_t\begin{pmatrix}q\\u\end{pmatrix}
=\mathcal{L}\begin{pmatrix}q\\u\end{pmatrix}
\quad\text{where}\quad
\mathcal{L}=\begin{pmatrix}0&\operatorname{div}\\\operatorname{grad}&0\end{pmatrix}. ∂ t ( q u ) = L ( q u ) where L = ( 0 grad div 0 ) .
When γ = p ˉ = 1 \gamma=\bar{p}=1 γ = p ˉ = 1 the conserved energy reduces to the standard L 2 L^2 L 2 inner product. Here we note that here the leading order operator L \mathcal{L} L is skew-symmetric with respect to L 2 L^2 L 2 since the L 2 L^2 L 2 adjoint of div \operatorname{div} div is − grad -\operatorname{grad} − grad .
We then decompose ( L 2 ) 3 = ( L 2 ) σ 3 + H ˙ 1 (L^2)^3=(L^2)^3_\sigma+\dot{H}^1 ( L 2 ) 3 = ( L 2 ) σ 3 + H ˙ 1 , i.e. writing u = σ + ∇ ϕ u=\sigma+\nabla\phi u = σ + ∇ ϕ where σ \sigma σ is divergence-free, such that
( q u ) = ( q ∇ ϕ ) + ( 0 σ ) . \begin{pmatrix}q\\u\end{pmatrix}
=
\begin{pmatrix}q\\\nabla\phi\end{pmatrix}
+
\begin{pmatrix}0\\\sigma\end{pmatrix}. ( q u ) = ( q ∇ ϕ ) + ( 0 σ ) .
The first term in this decomposition is the fast acoustic term. The second term in this decomposition is the slow term obeying incompressible Euler. In particular: the kernel of the leading order operator L \mathcal{L} L is precisely the space { 0 } × ( L 2 ) σ 3 \{0\}\times(L^2)^3_\sigma { 0 } × ( L 2 ) σ 3 to which the second component belongs, and so this is an honest-to-goodness slow-fast decomposition. (We could also readily verify that the image of L \mathcal{L} L is the space in which the first component lives.)
In other words: the Helmholtz decomposition is precisely the slow-fast decomposition for this compressible Euler system!
Moreover considering ∂ t \partial_t ∂ t ((A.7a)) − ∇ ⋅ -\nabla\cdot − ∇ ⋅ ((A.7b)) produces
( ∂ t 2 − Δ ) q = 0. (\partial_t^2-\Delta)q=0. ( ∂ t 2 − Δ ) q = 0.
and so, for S S S denoting the semigroup associated with the wave equation above,
q ( t ) = S ( t ε ) q ˉ ( t ) . q(t)=S\left(\frac{t}{\varepsilon}\right)\bar{q}(t). q ( t ) = S ( ε t ) q ˉ ( t ) .
where q ˉ ( t ) \bar{q}(t) q ˉ ( t ) would then be determined by a solvability condition arising from the next-order dynamics.
Useful computations, identities, and facts
In this section we record a variety of computations, identities, and facts which are of use throughout the paper.
First we record a useful splitting of the curl operator into its horizontal and vertical components.
Lemma A.4 (Horizontal-vertical decomposition of the curl). For any sufficiently regular vector field v v v ,
∇ × v = ∂ 3 v h ⊥ − ∇ h ⊥ v 3 + ( ∇ h ⊥ ⋅ v h ) e 3 . \nabla\times v=\partial_3v_h^\perp-\nabla_h^\perp v_3+(\nabla_h^\perp\cdot v_h)e_3. ∇ × v = ∂ 3 v h ⊥ − ∇ h ⊥ v 3 + ( ∇ h ⊥ ⋅ v h ) e 3 .
Proof. Since ∇ × v = ( ∂ 2 v 3 − ∂ 3 v 2 , ∂ 3 v 1 − ∂ 1 v 3 , ∂ 1 v 2 − ∂ 2 v 1 ) \nabla\times v=(\partial_2v_3-\partial_3v_2,\partial_3v_1-\partial_1v_3,\partial_1v_2-\partial_2v_1) ∇ × v = ( ∂ 2 v 3 − ∂ 3 v 2 , ∂ 3 v 1 − ∂ 1 v 3 , ∂ 1 v 2 − ∂ 2 v 1 ) the result follows from appropriately grouping terms.
We now record a couple of identities concerning the divergence and curl of a vector field of the form of the “good unknowns” (see Remark 3.2). This is a very useful identity since it helps us tie the good unknown to the potential vorticity P V PV P V and to the thermal wind imbalance j j j (see also Remark 3.2).
Lemma A.5. Consider a sufficiently regular vector field v v v and scalar field ϕ \phi ϕ . Then
∇ ⋅ ( − v h ⊥ + ϕ e 3 ) = ∇ h ⊥ ⋅ v h + ∂ 3 ϕ and ∇ × ( − v h ⊥ + ϕ e 3 ) = ∂ 3 v h − ∇ h ⊥ ϕ − ( ∇ h ⋅ v h ) e 3 . \nabla\cdot(-v_h^\perp+\phi e_3)=\nabla_h^\perp\cdot v_h+\partial_3\phi
\quad\text{and}\quad
\nabla\times(-v_h^\perp+\phi e_3)=\partial_3v_h-\nabla_h^\perp\phi-(\nabla_h\cdot v_h)e_3. ∇ ⋅ ( − v h ⊥ + ϕ e 3 ) = ∇ h ⊥ ⋅ v h + ∂ 3 ϕ and ∇ × ( − v h ⊥ + ϕ e 3 ) = ∂ 3 v h − ∇ h ⊥ ϕ − ( ∇ h ⋅ v h ) e 3 . In particular if v v v is divergence-free then
∇ × ( − v h ⊥ + ϕ e 3 ) = ∂ 3 v h − ∇ h ⊥ ϕ + ( ∂ 3 v 3 ) e 3 . \nabla\times(-v_h^\perp+\phi e_3)=\partial_3v_h-\nabla_h^\perp\phi+(\partial_3v_3)e_3. ∇ × ( − v h ⊥ + ϕ e 3 ) = ∂ 3 v h − ∇ h ⊥ ϕ + ( ∂ 3 v 3 ) e 3 .
Proof. This follows immediately from Lemma A.4 .
We now record a useful algebraic inversion which tells us that we may equivalently characterise the thermodynamic variables θ \theta θ and q q q by instead using the buoyancy b = θ − min 0 q b = \theta- \min_0 q b = θ − min 0 q and the moist variable M = θ + q M = \theta+ q M = θ + q .
Lemma A.6. Given any two scalars b b b and M M M the nonlinear algebraic system
{ θ + q = M and θ − min 0 q = b \begin{cases}
\theta+ q = M \quad\text{and}\\
\theta- \min_0 q = b
\end{cases} { θ + q = M and θ − min 0 q = b is invertible, with solution given uniquely by
{ θ = b + 1 2 min 0 ( M − b ) and q = M − b − 1 2 min 0 ( M − b ) . \begin{cases}
\theta= b + \frac{1}{2}\min_0 (M-b) \quad\text{and}\\
q = M-b-\frac{1}{2}\min_0 (M-b).
\end{cases} { θ = b + 2 1 min 0 ( M − b ) and q = M − b − 2 1 min 0 ( M − b ) .
Proof. We can do this easily by introducing H = 1 ( q < 0 ) H = \mathbf{1}(q < 0) H = 1 ( q < 0 ) . Treating H H H as fixed, the system then becomes linear (and invertible) and we obtain
{ θ = b + 1 2 H ( M − b ) and q = ( 1 − H / 2 ) ( M − b ) . \begin{cases}
\theta= b + \frac{1}{2}H(M-b) \quad\text{and}\\
q = (1-H/2)(M-b).
\end{cases} { θ = b + 2 1 H ( M − b ) and q = ( 1 − H /2 ) ( M − b ) .
In particular we note that sign q = sign ( M − b ) \operatorname{sign} q = \operatorname{sign}(M-b) sign q = sign ( M − b ) and so H ( M − b ) = min 0 ( M − b ) H(M-b) = \min_0 (M-b) H ( M − b ) = min 0 ( M − b ) , from which the claim follows.
Below we relate another identity related to “good unknowns”. This time we obtain an identity for the vector field − u h ⊥ + ( θ − q H ) e 3 -u_h^\perp+ (\theta-qH)e_3 − u h ⊥ + ( θ − q H ) e 3 which appears in the “fixed H H H ” version of the measurement-to-coordinates map (see Lemma 5.8 and Remark 5.9 ).
Lemma A.7. If, for some indicator function H H H ,
u = ∇ h ⊥ p + σ h ⊥ + w e 3 , u = \nabla_h^\perp p + \sigma_h^\perp+ we_3, u = ∇ h ⊥ p + σ h ⊥ + w e 3 , θ = ∂ 3 p + H 2 ( M − ∂ 3 p ) + σ 3 , and \theta= \partial_3p + \frac{H}{2}(M-\partial_3p) + \sigma_3,\quad\text{and} θ = ∂ 3 p + 2 H ( M − ∂ 3 p ) + σ 3 , and q = M − ∂ 3 p − H 2 ( M − ∂ 3 p ) − σ 3 q = M-\partial_3p-\frac{H}{2}(M-\partial_3p)-\sigma_3 q = M − ∂ 3 p − 2 H ( M − ∂ 3 p ) − σ 3 for some divergence-free σ \sigma σ satisfying ∇ h ⊥ ⋅ σ h = ∂ 3 w \nabla_h^\perp\cdot\sigma_h = \partial_3w ∇ h ⊥ ⋅ σ h = ∂ 3 w then
− u h ⊥ + ( θ − q H ) e 3 = ∇ p + A H − 1 σ -u_h^\perp+ (\theta-qH)e_3 = \nabla p + A_H^{-1}\sigma − u h ⊥ + ( θ − q H ) e 3 = ∇ p + A H − 1 σ where A H = I − H 2 e 3 ⊗ e 3 A_H = I-\frac{H}{2}e_3\otimes e_3 A H = I − 2 H e 3 ⊗ e 3 and so A H − 1 = I + H e 3 ⊗ e 3 A_H^{-1}=I+He_3\otimes e_3 A H − 1 = I + H e 3 ⊗ e 3 (since ( 1 − H / 2 ) ( 1 + H ) ≡ 1 (1-H/2)(1+H)\equiv1 ( 1 − H /2 ) ( 1 + H ) ≡ 1 ).
Proof. Since ( v ⊥ ) ⊥ = − v h (v^\perp)^\perp=-v_h ( v ⊥ ) ⊥ = − v h we see immediately that − u h ⊥ = ∇ h p + σ h -u_h^\perp=\nabla_hp+\sigma_h − u h ⊥ = ∇ h p + σ h . Then we compute that
θ − q H = ( 1 − H 2 ) ∂ 3 p + H 2 M + σ 3 − 1 2 H ( M − ∂ 3 p ) + H σ 3 = ∂ 3 p + ( 1 + H ) σ 3 . \theta-qH=\left(1-\frac{H}{2}\right)\partial_3p+\frac{H}{2}M+\sigma_3-\frac{1}{2}H(M-\partial_3p)+H\sigma_3=\partial_3p+(1+H)\sigma_3. θ − q H = ( 1 − 2 H ) ∂ 3 p + 2 H M + σ 3 − 2 1 H ( M − ∂ 3 p ) + H σ 3 = ∂ 3 p + ( 1 + H ) σ 3 .
The claim follows from combining these two observations.
We now record a well-posedness result for a div-curl system involving an elliptic matrix. This is used to verify that the “fixed H H H ” version of the measurement–to–coordinate map is invertible.
Proposition A.8 (Solvability of a div-curl system). For any uniformly elliptic B ∈ L ∞ B \in L^\infty B ∈ L ∞ , f ∈ ( H ˙ σ 1 ) ∗ f \in( \dot H^1_\sigma)^* f ∈ ( H ˙ σ 1 ) ∗ , and g ∈ R 3 g \in\mathbb{R}^3 g ∈ R 3 there is a unique σ ∈ ( L σ 2 ) 3 \sigma\in(L^2_\sigma)^3 σ ∈ ( L σ 2 ) 3 satisfying
∇ × ( B σ ) = f , ∇ ⋅ σ = 0 , and ∫ σ = g . (90) \nabla\times(B\sigma)=f,\ \nabla\cdot\sigma=0,\ \text{and}\ \int\sigma=g. \tag*{(90)} ∇ × ( B σ ) = f , ∇ ⋅ σ = 0 , and ∫ σ = g . (90)
Proof. First we show that a solution exists. We let ψ ∈ H ˚ σ 1 \psi\in\mathring{H}^{1}_{\sigma} ψ ∈ H ˚ σ 1 be the unique solution of
∇ × ( ∇ × ψ ) = f \nabla\times(\nabla\times\psi) = f ∇ × ( ∇ × ψ ) = f
and then let π ∈ H ˚ 1 \pi\in\mathring{H}^{1} π ∈ H ˚ 1 be the unique solution of
∇ ⋅ ( B − 1 ∇ π ) = − ∇ ⋅ [ B − 1 ( ∇ × ψ + g ) ] . \nabla\cdot(B^{-1}\nabla\pi) = -\nabla\cdot\left[B^{-1}(\nabla\times\psi+ g)\right]. ∇ ⋅ ( B − 1 ∇ π ) = − ∇ ⋅ [ B − 1 ( ∇ × ψ + g ) ] .
Then
σ : = B − 1 ( ∇ π + ∇ × ψ + g ) \sigma:= B^{-1}(\nabla\pi+ \nabla\times\psi+ g) σ := B − 1 ( ∇ π + ∇ × ψ + g )
is a solution. Indeed
∫ B σ = g and ∇ × ( B σ ) = ∇ × ( ∇ × ψ ) = f \int B\sigma= g \text{ and } \nabla\times(B\sigma) = \nabla\times(\nabla\times\psi) = f ∫ B σ = g and ∇ × ( B σ ) = ∇ × ( ∇ × ψ ) = f
while
∇ ⋅ σ = ∇ ⋅ ( B − 1 ∇ π ) + ∇ ⋅ [ B − 1 ( ∇ × ψ + g ) ] = 0. \nabla\cdot\sigma= \nabla\cdot(B^{-1}\nabla\pi) + \nabla\cdot\left[B^{-1}(\nabla\times\psi+ g)\right] = 0. ∇ ⋅ σ = ∇ ⋅ ( B − 1 ∇ π ) + ∇ ⋅ [ B − 1 ( ∇ × ψ + g ) ] = 0.
The uniqueness of a solution then follows from the uniqueness of the Helmholtz decomposition since we have characterized the solution σ \sigma σ by the Helmholtz decomposition of B σ B\sigma B σ . If σ 1 \sigma_1 σ 1 and σ 2 \sigma_2 σ 2 are solutions of (A.8) then their Helmholtz decompositions
B σ i = ∇ π i + ∇ × ψ i + c i , B\sigma_i = \nabla\pi_i + \nabla\times\psi_i + c_i, B σ i = ∇ π i + ∇ × ψ i + c i ,
for i = 1 , 2 i = 1, 2 i = 1 , 2 , where c i ∈ R 3 c_i \in\mathbb{R}^{3} c i ∈ R 3 , satisfy
c 1 = ∫ B σ 1 = g = ∫ B σ 2 = c 2 c_1 = \int B\sigma_1 = g = \int B\sigma_2 = c_2 c 1 = ∫ B σ 1 = g = ∫ B σ 2 = c 2
and
∇ × ( ∇ × ψ 1 ) = ∇ × ( B σ 1 ) = f = ∇ × ( B σ 2 ) = ∇ × ( ∇ × ψ 2 ) , \nabla\times(\nabla\times\psi_1) = \nabla\times(B\sigma_1) = f = \nabla\times(B\sigma_2) = \nabla\times(\nabla\times\psi_2), ∇ × ( ∇ × ψ 1 ) = ∇ × ( B σ 1 ) = f = ∇ × ( B σ 2 ) = ∇ × ( ∇ × ψ 2 ) ,
which ensures that c 1 = c 2 c_1 = c_2 c 1 = c 2 and ψ 1 = ψ 2 \psi_1 = \psi_2 ψ 1 = ψ 2 . This means that
B ( σ 1 − σ 2 ) = ∇ ( π 1 − π 2 ) B(\sigma_1 - \sigma_2) = \nabla(\pi_1 - \pi_2) B ( σ 1 − σ 2 ) = ∇ ( π 1 − π 2 )
and so
∇ ⋅ [ B − 1 ∇ ( π 1 − π 2 ) ] = ∇ ⋅ ( σ 1 − σ 2 ) = 0 , \nabla\cdot\left[B^{-1}\nabla(\pi_1 - \pi_2)\right] = \nabla\cdot(\sigma_1 - \sigma_2) = 0, ∇ ⋅ [ B − 1 ∇ ( π 1 − π 2 ) ] = ∇ ⋅ ( σ 1 − σ 2 ) = 0 ,
which also ensures that π 1 = π 2 \pi_1 = \pi_2 π 1 = π 2 , and hence that σ 1 = σ 2 \sigma_1 = \sigma_2 σ 1 = σ 2 .
The result below pertains to the leading order dynamics discussed in § 4 . It tells us how the balanced and wave sets relate to the dynamics of solutions of ∂ t + N \partial_t + \mathcal{N} ∂ t + N .
Proposition A.9. Let X = ( u , θ , q ) \mathcal{X} = (u,\theta,q) X = ( u , θ , q ) be a solution of ∂ t X + N X = 0 \partial_t\mathcal{X} + \mathcal{N}\mathcal{X} = 0 ∂ t X + N X = 0 for N \mathcal{N} N as defined in (13) . Then
X = ( ∇ h p ∂ 3 p + 1 2 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 1 2 min 0 ( M − ∂ 3 p ) ) ⏟ X B + ( σ h + w e 3 σ 3 − σ 3 ) ⏟ X W \mathcal{X} =
\underbrace{\begin{pmatrix}
\nabla_h p \\
\partial_3 p + \frac{1}{2}\min_0(M-\partial_3p) \\
M-\partial_3p-\frac{1}{2}\min_0(M-\partial_3p)
\end{pmatrix}}_{\mathcal{X}_{B}}
+
\underbrace{\begin{pmatrix}
\sigma_h+w e_3 \\
\sigma_3 \\
-\sigma_3
\end{pmatrix}}_{\mathcal{X}_{W}} X = X B ∇ h p ∂ 3 p + 2 1 min 0 ( M − ∂ 3 p ) M − ∂ 3 p − 2 1 min 0 ( M − ∂ 3 p ) + X W σ h + w e 3 σ 3 − σ 3 where X B ∈ B \mathcal{X}_{B} \in\mathcal{B} X B ∈ B and X W ∈ W \mathcal{X}_{W} \in\mathcal{W} X W ∈ W for B \mathcal{B} B and W \mathcal{W} W as in Propositions 3.1 and 3.3 , respectively, for M : = θ + q M := \theta+ q M := θ + q and p p p the solution of
Δ p + 1 2 ∂ 3 min 0 ( M − ∂ 3 p ) = ∇ h ⋅ u h + ∂ 3 θ ⏟ P V \Delta p + \frac{1}{2}\partial_3\min_0(M-\partial_3p) = \underbrace{\nabla_h \cdot u_h + \partial_3\theta}_{PV} Δ p + 2 1 ∂ 3 0 min ( M − ∂ 3 p ) = P V ∇ h ⋅ u h + ∂ 3 θ and where X W : = X − X B \mathcal{X}_{W} := \mathcal{X}-\mathcal{X}_{B} X W := X − X B . Moreover
∂ t X B = 0 and ∂ t X W = ( P L ( − σ h − ( 1 + H B ) σ 3 e 3 ) w − w ) = : L W , H B X W ∈ W \partial_t\mathcal{X}_{B} = 0 \text{ and } \partial_t\mathcal{X}_{W} =
\begin{pmatrix}
\mathbb{P}_{L}(-\sigma_h-(1+H_{B})\sigma_3e_3) \\
w \\
-w
\end{pmatrix}
=: \mathcal{L}_{W,H_{B}}\mathcal{X}_{W} \in\mathcal{W} ∂ t X B = 0 and ∂ t X W = P L ( − σ h − ( 1 + H B ) σ 3 e 3 ) w − w =: L W , H B X W ∈ W for H B : = 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) H_{B}:=\mathbf{1}(q_{B}<0)=\mathbf{1}(M<\partial_3p) H B := 1 ( q B < 0 ) = 1 ( M < ∂ 3 p ) . Finally: if X W = ( σ h , 0 , 0 ) \mathcal{X}_{W}=(\sigma_h,0,0) X W = ( σ h , 0 , 0 ) then
L W , H B 2 X W = − X W \mathcal{L}^{2}_{W,H_{B}}\mathcal{X}_{W}=-\mathcal{X}_{W} L W , H B 2 X W = − X W and if X W = ( w e 3 , σ 3 , − σ 3 ) \mathcal{X}_{W}=(we_3,\sigma_3,-\sigma_3) X W = ( w e 3 , σ 3 , − σ 3 ) then
L W , H B 2 X W = − ( 1 + H B ) X W . \mathcal{L}^{2}_{W,H_{B}}\mathcal{X}_{W}=-(1+H_{B})\mathcal{X}_{W}. L W , H B 2 X W = − ( 1 + H B ) X W . In other words: the B \mathcal{B} B -component is constant whereas the dynamics of the W \mathcal{W} W -component induce an X B \mathcal{X}_{B} X B -dependent dynamical system on W \mathcal{W} W .
Proof. The existence and uniqueness of p p p is established in [89 ] . The fact that the B \mathcal{B} B -component is constant follows from the fact that ∂ t P V = ∂ t M = 0 \partial_t P_V=\partial_t M=0 ∂ t P V = ∂ t M = 0 and the fact that ∂ t p \partial_t p ∂ t p solves
∇ ⋅ ( A B ∇ ∂ t p ) = 0 \nabla\cdot(A_{\mathcal{B}}\nabla\partial_t p)=0 ∇ ⋅ ( A B ∇ ∂ t p ) = 0
for A B : = I − 1 2 1 ( M < ∂ 3 p ) e 3 ⊗ e 3 A_{\mathcal{B}}:=I-\frac{1}{2}\mathbf{1}(M<\partial_3p)e_3\otimes e_3 A B := I − 2 1 1 ( M < ∂ 3 p ) e 3 ⊗ e 3 uniformly elliptic, such that indeed ∂ t p \partial_t p ∂ t p vanishes. Then X W : = X − X B \mathcal{X}_{W}:=\mathcal{X}-\mathcal{X}_{\mathcal{B}} X W := X − X B belongs to W \mathcal{W} W and takes the prescribed form by virtue of Proposition 3.3 . Moreover we note that im L W , H B ⊆ W \operatorname{im}\mathcal{L}_{W,H_{\mathcal{B}}}\subseteq\mathcal{W} im L W , H B ⊆ W since
P V ( P L ( − σ h − ( 1 + H B ) σ 3 e 3 ) w − w ) = − ∇ h ⊥ ⋅ σ h + ∂ 3 w = 0. P_V\begin{pmatrix}
P_L(-\sigma_h-(1+H_{\mathcal{B}})\sigma_3e_3)\\
w\\
-w
\end{pmatrix}
=-\nabla_h^\perp\cdot\sigma_h+\partial_3w=0. P V P L ( − σ h − ( 1 + H B ) σ 3 e 3 ) w − w = − ∇ h ⊥ ⋅ σ h + ∂ 3 w = 0.
and
M ( P L ( − σ h − ( 1 + H B ) σ 3 e 3 ) w − w ) = w − w = 0. \mathcal{M}\begin{pmatrix}
P_L(-\sigma_h-(1+H_{\mathcal{B}})\sigma_3e_3)\\
w\\
-w
\end{pmatrix}
=w-w=0. M P L ( − σ h − ( 1 + H B ) σ 3 e 3 ) w − w = w − w = 0.
Finally the identities involving L 2 \mathcal{L}^2 L 2 follow from direct computations.
We now turn our attention to identities involving the conserved energy. First we record the computation verifying that the energy is indeed conserved.
Proposition A.10 (Conserved energy). Let ( u , θ , q ) (u,\theta,q) ( u , θ , q ) be a solution of the moist Boussinesq system (2.7a)–(2.7c). Then the energy
E = 1 2 ∫ ∣ u ∣ 2 + θ 2 + min 0 2 q + ( θ + q ) 2 E=\frac{1}{2}\int\lvert u\rvert^2+\theta^2+\min_0^2q+(\theta+q)^2 E = 2 1 ∫ ∣ u ∣ 2 + θ 2 + 0 min 2 q + ( θ + q ) 2 is conserved, meaning that d E d t = 0 \frac{dE}{dt}=0 d t d E = 0 .
Proof. We immediately note that, for M : = θ + q M:=\theta+q M := θ + q ,
( ∂ t + u ⋅ ∇ ) M = 0. (\partial_t+u\cdot\nabla)M=0. ( ∂ t + u ⋅ ∇ ) M = 0.
Therefore, since u u u is divergence-free,
d d t ∫ 1 2 ( θ + q ) 2 = ∫ ( ∂ t + u ⋅ ∇ ) ( 1 2 M 2 ) = ∫ M ( ∂ t + u ⋅ ∇ ) M = 0. \frac{d}{dt}\int\frac{1}{2}(\theta+q)^2=\int(\partial_t+u\cdot\nabla)\left(\frac{1}{2}M^2\right)=\int M(\partial_t+u\cdot\nabla)M=0. d t d ∫ 2 1 ( θ + q ) 2 = ∫ ( ∂ t + u ⋅ ∇ ) ( 2 1 M 2 ) = ∫ M ( ∂ t + u ⋅ ∇ ) M = 0.
Now observe that d d s min 0 2 s = min 0 s \frac{d}{ds}\min_0^2s=\min_0s d s d min 0 2 s = min 0 s and so
( ∂ t + u ⋅ ∇ ) ( 1 2 min 0 2 q ) = ( min 0 q ) ( ∂ t + u ⋅ ∇ ) q . (\partial_t+u\cdot\nabla)\left(\frac{1}{2}\min_0^2q\right)=(\min_0q)(\partial_t+u\cdot\nabla)q. ( ∂ t + u ⋅ ∇ ) ( 2 1 0 min 2 q ) = ( 0 min q ) ( ∂ t + u ⋅ ∇ ) q .
Therefore
d d t ∫ ∣ u ∣ 2 + 1 2 θ 2 + 1 2 min 0 2 q = ∫ u ⋅ ( ∂ t + u ⋅ ∇ ) u + θ ( ∂ t + u ⋅ ∇ ) θ + ( min 0 q ) ( ∂ t + u ⋅ ∇ ) q = 1 ε ∫ − u ⋅ e 3 × u − u ⋅ ∇ p + u 3 ( θ − min 0 q ) − 1 ε ∫ θ u 3 + 1 ε ∫ ( min 0 q ) u 3 = 1 ε ∫ ( ∇ ⋅ u ) p = 0 , \begin{aligned}
\frac{d}{dt}\int\lvert u\rvert^2+\frac{1}{2}\theta^2+\frac{1}{2}\min_0^2q
&=\int u\cdot(\partial_t+u\cdot\nabla)u+\theta(\partial_t+u\cdot\nabla)\theta+(\min_0q)(\partial_t+u\cdot\nabla)q\\
&=\frac{1}{\varepsilon}\int-u\cdot e_3\times u-u\cdot\nabla p+u_3(\theta-\min_0q)-\frac{1}{\varepsilon}\int\theta u_3+\frac{1}{\varepsilon}\int(\min_0q)u_3\\
&=\frac{1}{\varepsilon}\int(\nabla\cdot u)p=0,
\end{aligned} d t d ∫ ∣ u ∣ 2 + 2 1 θ 2 + 2 1 0 min 2 q = ∫ u ⋅ ( ∂ t + u ⋅ ∇ ) u + θ ( ∂ t + u ⋅ ∇ ) θ + ( 0 min q ) ( ∂ t + u ⋅ ∇ ) q = ε 1 ∫ − u ⋅ e 3 × u − u ⋅ ∇ p + u 3 ( θ − 0 min q ) − ε 1 ∫ θ u 3 + ε 1 ∫ ( 0 min q ) u 3 = ε 1 ∫ ( ∇ ⋅ u ) p = 0 ,
and so d E d t = 0 \frac{dE}{dt}=0 d t d E = 0 as claimed.
We now turn our attention to another variant of the energy, namely the “fixed H H H ” energy used in § 5 . We first show that the (formally) linearised operator is skew-adjoint with respect to that energy, which is also known in the geometric context of § 6.2 as the L H 2 L_H^2 L H 2 inner product.
Lemma A.11. Let H H H be an indicator function. The L H 2 L_H^2 L H 2 -adjoint of L H L_H L H is − L H -L_H − L H .
Proof. For any ( u , θ , q ) (u,\theta,q) ( u , θ , q ) and ( v , ϕ , r ) (v,\phi,r) ( v , ϕ , r ) in L σ 2 L^2_\sigma L σ 2 we compute that
⟨ L H ( u , θ , q ) , ( v , ϕ , r ) ⟩ L H 2 = ∫ P L ( u h ⊥ − θ e 3 + q H e 3 ) ⋅ v + u 3 ϕ − u 3 r H + 0 = ∫ − u ⋅ v h ⊥ − θ v 3 + q v 3 H + u 3 ϕ − u 3 r H = ∫ u ⋅ ( − v h ⊥ + ϕ e 3 − r H e 3 ) + θ ( − v 3 ) + q ⋅ v 3 ⋅ H = ⟨ ( u , θ , q ) , ( P L ( − v h ⊥ + ϕ e 3 − r H e 3 ) , − v 3 , v 3 ) ⟩ L H 2 . \begin{aligned}
\langle\mathcal{L}_H(u,\theta,q),(v,\phi,r)\rangle_{L^2_H}
&=\int\mathbb{P}_L\left(u_h^\perp-\theta e_3+qHe_3\right)\cdot v+u_3\phi-u_3rH+0\\
&=\int-u\cdot v_h^\perp-\theta v_3+qv_3H+u_3\phi-u_3rH\\
&=\int u\cdot\left(-v_h^\perp+\phi e_3-rHe_3\right)+\theta(-v_3)+q\cdot v_3\cdot H\\
&=\left\langle(u,\theta,q),\left(\mathbb{P}_L\left(-v_h^\perp+\phi e_3-rHe_3\right),-v_3,v_3\right)\right\rangle_{L^2_H}.
\end{aligned} ⟨ L H ( u , θ , q ) , ( v , ϕ , r ) ⟩ L H 2 = ∫ P L ( u h ⊥ − θ e 3 + q H e 3 ) ⋅ v + u 3 ϕ − u 3 r H + 0 = ∫ − u ⋅ v h ⊥ − θ v 3 + q v 3 H + u 3 ϕ − u 3 r H = ∫ u ⋅ ( − v h ⊥ + ϕ e 3 − r H e 3 ) + θ ( − v 3 ) + q ⋅ v 3 ⋅ H = ⟨ ( u , θ , q ) , ( P L ( − v h ⊥ + ϕ e 3 − r H e 3 ) , − v 3 , v 3 ) ⟩ L H 2 .
and so indeed
L H ∗ ( v ϕ r ) = ( P L ( − v h ⊥ + ϕ e 3 − r H e 3 ) − v 3 v 3 ) = − L H ( v ϕ r ) . \mathcal{L}_H^*
\begin{pmatrix}
v\\
\phi\\
r
\end{pmatrix}
=
\begin{pmatrix}
\mathbb{P}_L\left(-v_h^\perp+\phi e_3-rHe_3\right)\\
-v_3\\
v_3
\end{pmatrix}
=-\mathcal{L}_H
\begin{pmatrix}
v\\
\phi\\
r
\end{pmatrix}. L H ∗ v ϕ r = P L ( − v h ⊥ + ϕ e 3 − r H e 3 ) − v 3 v 3 = − L H v ϕ r .
We now record another property of the “fixed H H H ” energy, showing that if the indicator is time-dependent, this energy fails to be conserved.
Proposition A.12. Suppose that H H H is a given indicator function and that X = ( u , θ , q ) \mathcal{X}=(u,\theta,q) X = ( u , θ , q ) is a solution of ∂ t X + L H X = 0 \partial_t\mathcal{X}+\mathcal{L}_H\mathcal{X}=0 ∂ t X + L H X = 0 for L H \mathcal{L}_H L H as defined in (28) . For
E H ( X ) : = 1 2 ∫ ∣ u ∣ 2 + θ 2 + q 2 H + ( θ + q ) 2 E_H(\mathcal{X}):=\frac{1}{2}\int|u|^2+\theta^2+q^2H+(\theta+q)^2 E H ( X ) := 2 1 ∫ ∣ u ∣ 2 + θ 2 + q 2 H + ( θ + q ) 2 we have that
d d t E H = 1 2 ∫ q 2 ∂ t H . \frac{d}{dt}E_H=\frac{1}{2}\int q^2\partial_tH. d t d E H = 2 1 ∫ q 2 ∂ t H . In particular E H E_H E H is constant if and only if the H H H -interface is contained in the q q q -interface, i.e. ∂ { H = 1 } ⊆ { q = 0 } \partial\{H=1\}\subseteq\{q=0\} ∂ { H = 1 } ⊆ { q = 0 } .
Proof. This follows from the fact that if X \mathcal{X} X solves ∂ t X + L X \partial_t\mathcal{X}+\mathcal{L}\mathcal{X} ∂ t X + LX then M : = θ + q M:=\theta+q M := θ + q satisfies ∂ t M = 0 \partial_tM=0 ∂ t M = 0 and so we may compute that
∫ u ⋅ ∂ t u + θ ∂ t θ + q ( ∂ t q ) H + M ( ∂ t M ) = ∫ u h ⊥ ⋅ u h ⊥ − u 3 ( θ − q H ) + θ u 3 − q H u 3 = 0 \int u\cdot\partial_tu+\theta\partial_t\theta+q(\partial_tq)H+M(\partial_tM)=\int u_h^\perp\cdot u_h^\perp-u_3(\theta-qH)+\theta u_3-qHu_3=0 ∫ u ⋅ ∂ t u + θ ∂ t θ + q ( ∂ t q ) H + M ( ∂ t M ) = ∫ u h ⊥ ⋅ u h ⊥ − u 3 ( θ − q H ) + θ u 3 − q H u 3 = 0
(in other words: if H H H is constant in time then the operator L H \mathcal{L}_H L H is skew-adjoint with respect to the inner product induced by E H E_H E H ).
We now turn our attention to a standard result verifying that pullback metrics are themselves metrics. This plays a crucial role in § 5 in order to eventually define the statistical divergence we introduce in this paper and which agrees with both the decomposition and the conserved energy.
Proposition A.13 (Pullback metric). Let S S S be a set, let ( M , d ) (M,d) ( M , d ) be a metric space, and let F : S → M F:S\to M F : S → M be injective. The map F ∗ d : S × S → [ 0 , ∞ ) F^*d:S\times S\to[0,\infty) F ∗ d : S × S → [ 0 , ∞ ) defined by
F ∗ d ( x , y ) = d ( F ( x ) , F ( y ) ) F^*d(x,y)=d(F(x),F(y)) F ∗ d ( x , y ) = d ( F ( x ) , F ( y )) for every x , y ∈ S x,y\in S x , y ∈ S is a metric on S S S known as the pullback metric of d d d under F F F .
Proof. The pullback metric immediately inherits symmetry and the triangle inequality from d d d . Its positive-definiteness then follows from the positive-definiteness of d d d and the injectivity of F F F : for any x , y ∈ S x,y\in S x , y ∈ S ,
F ∗ d ( x , y ) = 0 ⟺ d ( F ( x ) , F ( y ) ) = 0 ⟺ F ( x ) = F ( y ) ⟺ x = y . F^*d(x,y)=0\Longleftrightarrow d(F(x),F(y))=0\Longleftrightarrow F(x)=F(y)\Longleftrightarrow x=y. F ∗ d ( x , y ) = 0 ⟺ d ( F ( x ) , F ( y )) = 0 ⟺ F ( x ) = F ( y ) ⟺ x = y .
The last result we record in this appendix is a result used in § 6.2 to verify that the balanced set is an honest-to-goodness Lipschitz–regular Hilbert manifold. The result in question is an elementary estimate.
Lemma A.14. For any v ∈ ( L 2 ) 3 v\in(L^2)^3 v ∈ ( L 2 ) 3 and any ϕ ∈ L 2 \phi\in L^2 ϕ ∈ L 2 ,
∥ ∇ h ⊥ ⋅ v h + ∂ 3 ϕ ∥ H − 1 ≤ ∥ v h ∥ L 2 + ∥ ϕ ∥ L 2 . \left\|\nabla_h^\perp\cdot v_h+\partial_3\phi\right\|_{H^{-1}}\leq\|v_h\|_{L^2}+\|\phi\|_{L^2}. ∇ h ⊥ ⋅ v h + ∂ 3 ϕ H − 1 ≤ ∥ v h ∥ L 2 + ∥ ϕ ∥ L 2 .
Proof. This follows from a direct computation since, for any ψ ∈ H ˙ 1 \psi\in\dot{H}^{1} ψ ∈ H ˙ 1 ,
⟨ ∇ h ⊥ ⋅ v h + ∂ 3 ϕ , ψ ⟩ H − 1 × H ˙ 1 = ∫ ( ∇ h ⊥ ⋅ v h + ∂ 3 ϕ ) ψ = − ∫ v h ⋅ ∇ h ⊥ ψ + ϕ ∂ 3 ψ ≤ ( ∥ v h ∥ L 2 + ∥ ϕ ∥ L 2 ) ∥ ψ ∥ H ˙ 1 , \left\langle\nabla_{h}^{\perp} \cdot v_{h} + \partial_{3}\phi,\psi\right\rangle_{H^{-1}\times\dot{H}^{1}} = \int\left(\nabla_{h}^{\perp} \cdot v_{h}+\partial_{3}\phi\right)\psi
= -\int v_{h}\cdot\nabla_{h}^{\perp}\psi+\phi\partial_{3}\psi\leq\left(\|v_{h}\|_{L^{2}}+\|\phi\|_{L^{2}}\right)\|\psi\|_{\dot{H}^{1}}, ⟨ ∇ h ⊥ ⋅ v h + ∂ 3 ϕ , ψ ⟩ H − 1 × H ˙ 1 = ∫ ( ∇ h ⊥ ⋅ v h + ∂ 3 ϕ ) ψ = − ∫ v h ⋅ ∇ h ⊥ ψ + ϕ ∂ 3 ψ ≤ ( ∥ v h ∥ L 2 + ∥ ϕ ∥ L 2 ) ∥ ψ ∥ H ˙ 1 ,
as desired.
In this section we record various results from convex analysis that are of use to us. This begins with a result on various characterisations of strong convexity and concludes with results pertaining to the quadratic remainder (which are used in § 6.1 to obtain a quadratic upper bound on our variational energy, one of the two key ingredients used to deduce convergence rates for various descent methods).
First we recall a few first-order characterisations of strong convexity.
Lemma A.15 (Equivalent characterisations of strong convexity). Let ( H , ⟨ ⋅ , ⋅ ⟩ ) (H,\langle\cdot,\cdot\rangle) ( H , ⟨ ⋅ , ⋅ ⟩) be a Hilbert space and let f : H → R f:H\to\mathbb{R} f : H → R be Fréchet differentiable. For any μ > 0 \mu>0 μ > 0 the following are equivalent, where each inequality holds for every x , y ∈ H x,y\in H x , y ∈ H and θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] .
f ( θ x + ( 1 − θ ) y ) ≤ θ f ( x ) + ( 1 − θ ) f ( y ) − θ ( 1 − θ ) μ 2 ∥ x − y ∥ 2 (91) f(\theta x+(1-\theta)y)\leq\theta f(x)+(1-\theta)f(y)-\theta(1-\theta)\frac{\mu}{2}\|x-y\|^{2} \tag*{(91)} f ( θ x + ( 1 − θ ) y ) ≤ θ f ( x ) + ( 1 − θ ) f ( y ) − θ ( 1 − θ ) 2 μ ∥ x − y ∥ 2 (91) f ( y ) ≥ f ( x ) + ⟨ D f ( x ) , y − x ⟩ + μ 2 ∥ x − y ∥ 2 (92) f(y)\geq f(x)+\langle Df(x),y-x\rangle+\frac{\mu}{2}\|x-y\|^{2} \tag*{(92)} f ( y ) ≥ f ( x ) + ⟨ D f ( x ) , y − x ⟩ + 2 μ ∥ x − y ∥ 2 (92) ⟨ D f ( x ) − D f ( y ) , x − y ⟩ ≥ μ ∥ x − y ∥ 2 (93) \langle Df(x)-Df(y),x-y\rangle\geq\mu\|x-y\|^{2} \tag*{(93)} ⟨ D f ( x ) − D f ( y ) , x − y ⟩ ≥ μ ∥ x − y ∥ 2 (93) If any of these inequalities hold recall that we say that f f f is μ \mu μ -strongly convex.
Proof. See Definition 2.1.2 and Theorem 2.1.9 in [79 ] .
We now turn our attention towards properties of the quadratic remainder. First we note that the map f ↦ T f f\mapsto T^{f} f ↦ T f is linear.
Lemma A.16 (Linearity of the quadratic remainder map). For any x ∈ R d x\in\mathbb{R}^{d} x ∈ R d the map T x : C 1 ( R d ; R ) → C 0 ( R d ; R ) T_{x}:C^{1}(\mathbb{R}^{d};\mathbb{R})\to C^{0}(\mathbb{R}^{d};\mathbb{R}) T x : C 1 ( R d ; R ) → C 0 ( R d ; R ) defined by f ↦ T x f f\mapsto T_{x}^{f} f ↦ T x f is linear.
Proof. In light of the definition of the quadratic remainder this follows immediately from the derivative being linear.
Second we observe that quadratic remainders can be used to characterise affine functions. This is also the first of two results teasing out the relation between affine functions and quadratic remainders.
Lemma A.17 (Vanishing quadratic remainders characterise affine functions). A continuously differentiable function f : R d → R f:\mathbb{R}^{d}\to\mathbb{R} f : R d → R is affine, meaning that there exist v ∈ R d v\in\mathbb{R}^{d} v ∈ R d and s ∈ R s\in\mathbb{R} s ∈ R for which f ( x ) = v T x + s f(x)=v^{T}x+s f ( x ) = v T x + s for every x ∈ R d x\in\mathbb{R}^{d} x ∈ R d , if and only if its quadratic remainder T x f T_{x}^{f} T x f vanishes everywhere, i.e. T x f ( y ) = 0 T_{x}^{f}(y)=0 T x f ( y ) = 0 for every x , y ∈ R d x,y\in\mathbb{R}^{d} x , y ∈ R d .
Proof. The “only if” direction follows from a direct computation. Conversely suppose that T x f ≡ 0 T_{x}^{f}\equiv0 T x f ≡ 0 . In particular this means that T 0 f ( y ) = 0 T_{0}^{f}(y)=0 T 0 f ( y ) = 0 , which may be rearranged into f ( y ) = f ( 0 ) + ∇ f ( 0 ) T y f(y)=f(0)+\nabla f(0)^{T}y f ( y ) = f ( 0 ) + ∇ f ( 0 ) T y , proving that f f f is affine.
To conclude this section we note how the quadratic remainder behaves under affine transformations – this is the second of two results on the relation between affine functions and quadratic remainders.
Lemma A.18 (Equivariance of the quadratic remainder map under affine transformations). Consider a continuously differentiable function f : R d → R f:\mathbb{R}^{d}\to\mathbb{R} f : R d → R and an affine function Φ : R d → R d \Phi:\mathbb{R}^{d}\to\mathbb{R}^{d} Φ : R d → R d . Then T x f ∘ Φ ( y ) = T Φ ( x ) f ( Φ ( y ) ) T_{x}^{f\circ\Phi}(y)=T_{\Phi(x)}^{f}(\Phi(y)) T x f ∘ Φ ( y ) = T Φ ( x ) f ( Φ ( y )) for every x , y ∈ R d x,y\in\mathbb{R}^{d} x , y ∈ R d .
Proof. We compute that
T x f ∘ Φ ( y ) = f ( Φ ( y ) ) − f ( Φ ( x ) ) − ( D Φ ) T ( x ) ( ∇ f ) ( Φ ( x ) ) ⋅ ( y − x ) . (94) T_{x}^{f\circ\Phi}(y)=f(\Phi(y))-f(\Phi(x))-(D\Phi)^{T}(x)(\nabla f)(\Phi(x))\cdot(y-x). \tag*{(94)} T x f ∘ Φ ( y ) = f ( Φ ( y )) − f ( Φ ( x )) − ( D Φ ) T ( x ) ( ∇ f ) ( Φ ( x )) ⋅ ( y − x ) . (94)
Since we seek to write T f ∘ Φ T^{f \circ\Phi} T f ∘ Φ in terms of T f T^f T f the key step in (A.12) is to write y − x y-x y − x in terms of Φ ( y ) − Φ ( x ) \Phi(y)-\Phi(x) Φ ( y ) − Φ ( x ) . To do so we note that Φ \Phi Φ is affine and so Lemma A.17 tells us that T Φ T^\Phi T Φ vanishes, i.e. Φ ( y ) − Φ ( x ) = D Φ ( x ) ( y − x ) \Phi(y)-\Phi(x)=D\Phi(x)(y-x) Φ ( y ) − Φ ( x ) = D Φ ( x ) ( y − x ) for every x , y ∈ R d x,y \in\mathbb{R}^d x , y ∈ R d . Plugging this into (A.12) yields
T x f ∘ Φ ( y ) = f ( Φ ( y ) ) − f ( Φ ( x ) ) − ( ∇ f ) ( Φ ( x ) ) ⋅ ( D Φ ) ( x ) ( y − x ) T_x^{f\circ\Phi}(y)=f(\Phi(y))-f(\Phi(x))-(\nabla f)(\Phi(x))\cdot(D\Phi)(x)(y-x) T x f ∘ Φ ( y ) = f ( Φ ( y )) − f ( Φ ( x )) − ( ∇ f ) ( Φ ( x )) ⋅ ( D Φ ) ( x ) ( y − x )
= f ( Φ ( y ) ) − f ( Φ ( x ) ) − ( ∇ f ) ( Φ ( x ) ) ⋅ ( Φ ( y ) − Φ ( x ) ) = T Φ ( x ) f ( Φ ( y ) ) , \qquad=f(\Phi(y))-f(\Phi(x))-(\nabla f)(\Phi(x))\cdot(\Phi(y)-\Phi(x))=T_{\Phi(x)}^f(\Phi(y)), = f ( Φ ( y )) − f ( Φ ( x )) − ( ∇ f ) ( Φ ( x )) ⋅ ( Φ ( y ) − Φ ( x )) = T Φ ( x ) f ( Φ ( y )) ,
as claimed.
Exact nonlinear waves
In this section we record results pertaining to the nonlinear oscillations with non-vanishing averages which are present in the moist Boussinesq system. The key results are Proposition A.19 , which shows that a simple ODE leads to nonlinear oscillations with non-vanishing averages, and Corollary A.22 which explains how this solutions of this simple ODE give rise to honest-to-goodness solutions of the moist Boussinesq system with oscillations of the same nature.
First we find an exact solution of this simple ODE.
Proposition A.19. The ODE q ′ ′ + q + min 0 q = 0 q''+q+\min_0 q=0 q ′′ + q + min 0 q = 0 admits periodic weak solutions with nonzero average over their period.
Proof. Define
Q ( t ) : = { 2 sin t if 0 ≤ t ≤ π and sin 2 ( t − t ˉ ) if π ≤ t < T Q(t):=
\begin{cases}
\sqrt{2}\sin t & \text{if } 0\le t\le\pi\text{ and}\\
\sin\sqrt{2}(t-\bar{t}) & \text{if } \pi\le t<T
\end{cases} Q ( t ) := { 2 sin t sin 2 ( t − t ˉ ) if 0 ≤ t ≤ π and if π ≤ t < T for T : = π + π 2 T:=\pi+\frac{\pi}{\sqrt{2}} T := π + 2 π and t ˉ : = π − π 2 \bar{t}:=\pi-\frac{\pi}{\sqrt{2}} t ˉ := π − 2 π , and note that Q Q Q is differentiable. Then define q : [ 0 , ∞ ) → R q:[0,\infty)\to\mathbb{R} q : [ 0 , ∞ ) → R to be the T T T -periodic extention of Q Q Q , noting that q q q is also differentiable. Moreover: q ′ q' q ′ is continuous on { 0 , π } + T N \{0,\pi\}+T\mathbb{N} { 0 , π } + T N , i.e. on the set where it changes sign. Therefore q q q is a weak solution of q ′ ′ + q + min 0 q = 0 q''+q+\min_0q=0 q ′′ + q + min 0 q = 0 since
q ′ ′ = q when q > 0 and q ′ ′ = 2 q when q < 0. q''=q\text{ when }q>0\text{ and }q''=2q\text{ when }q<0. q ′′ = q when q > 0 and q ′′ = 2 q when q < 0. Finally the simple substitution s = 2 ( t − π ) + π s=\sqrt{2}(t-\pi)+\pi s = 2 ( t − π ) + π shows that
∫ 0 T q = ∫ 0 T Q = ∫ 0 π 2 sin t d t + ∫ π 2 π 1 2 sin s d s = ( 2 − 1 2 ) ∫ 0 π sin t d t ≠ 0 , \int_0^Tq=\int_0^TQ=\int_0^\pi\sqrt{2}\sin t\,dt+\int_\pi^{2\pi}\frac{1}{\sqrt{2}}\sin s\,ds=\left(\sqrt{2}-\frac{1}{\sqrt{2}}\right)\int_0^\pi\sin t\,dt\ne0, ∫ 0 T q = ∫ 0 T Q = ∫ 0 π 2 sin t d t + ∫ π 2 π 2 1 sin s d s = ( 2 − 2 1 ) ∫ 0 π sin t d t = 0 , as desired.
We now seek to use Proposition A.19 above to show that the moist Boussinesq system (2.7a)–(2.7c) admits solutions with the same nonlinear oscillations. As an intermediate step we consider an ODE system contained within the moist Boussinesq system (this notion is made precise in Lemma A.21 below).
Corollary A.20. The ODE system
{ d w d t = θ − min 0 q , (A.13a) d θ d t = − w , and (A.13b) d q d t = w (A.13c) \begin{cases}
\frac{dw}{dt}=\theta-\min_0 q, & \text{(A.13a)}\\
\frac{d\theta}{dt}=-w, \text{ and} & \text{(A.13b)}\\
\frac{dq}{dt}=w & \text{(A.13c)}
\end{cases} ⎩ ⎨ ⎧ d t d w = θ − min 0 q , d t d θ = − w , and d t d q = w (A.13a) (A.13b) (A.13c)
admits periodic solutions with nonzero averages over their period. Moreover these solutions may be chosen to satisfy M : = θ + q = 0 M := \theta+ q = 0 M := θ + q = 0 .
Proof. Suppose that M : = θ + q M := \theta+ q M := θ + q vanishes. Then q = − θ q = -\theta q = − θ and so the ODE system reduces to
d w d t = − q − min 0 q and d q d t = w \frac{dw}{dt} = -q - \min_{0} q \quad\text{and}\quad\frac{dq}{dt} = w d t d w = − q − 0 min q and d t d q = w
such that q q q must satisfy
d 2 q d t 2 = d w d t = − q − min 0 q , \frac{d^{2}q}{dt^{2}} = \frac{dw}{dt} = -q - \min_{0} q, d t 2 d 2 q = d t d w = − q − 0 min q ,
or equivalently
q ′ ′ + q + min 0 q = 0. (95) q^{\prime\prime} + q + \min_{0} q = 0. \tag*{(95)} q ′′ + q + 0 min q = 0. (95)
Proposition A.19 then guarantees the existence of a periodic weak solution q ∗ q_{*} q ∗ of (A.14) with nonzero average over its period. Then ( w ∗ , θ ∗ , q ∗ ) (w_{*},\theta_{*},q_{*}) ( w ∗ , θ ∗ , q ∗ ) is a solution of the original ODE system (A.13a)–(A.13c) satisfying θ ∗ + q ∗ = 0 \theta_{*} + q_{*} = 0 θ ∗ + q ∗ = 0 provided that
θ ∗ = − q ∗ and w ∗ = d q d t . \theta_{*} = -q_{*}\quad\text{and}\quad w_{*} = \frac{dq}{dt}. θ ∗ = − q ∗ and w ∗ = d t d q .
So finally, for the period T T T and for t ˉ \bar{t} t ˉ being defined as in Proposition A.19 we have that
w ∗ ∣ ( 0 , T ) = { 2 cos t if 0 ≤ t < π and 2 cos 2 ( t − t ˉ ) if π ≤ t < T w_{*}|_{(0,T)} =
\begin{cases}
\sqrt{2}\cos t & \text{if } 0 \le t < \pi\text{ and}\\
\sqrt{2}\cos\sqrt{2}(t-\bar{t}) & \text{if } \pi\le t < T
\end{cases} w ∗ ∣ ( 0 , T ) = { 2 cos t 2 cos 2 ( t − t ˉ ) if 0 ≤ t < π and if π ≤ t < T
and so
∫ 0 T w ∗ = ∫ 0 π 2 cos t d t + ∫ π 2 π cos s d s = ( 2 − 1 ) ∫ 0 π cos t d t ≠ 0. \int_{0}^{T} w_{*} = \int_{0}^{\pi}\sqrt{2}\cos t\,dt + \int_{\pi}^{2\pi}\cos s\,ds = (\sqrt{2}-1)\int_{0}^{\pi}\cos t\,dt \ne0. ∫ 0 T w ∗ = ∫ 0 π 2 cos t d t + ∫ π 2 π cos s d s = ( 2 − 1 ) ∫ 0 π cos t d t = 0.
So indeed w ∗ w_{*} w ∗ and θ ∗ \theta_{*} θ ∗ also have nonzero averages over their period.
As alluded to earlier we now show that solutions of the ODE system of Corollary A.20 give rise to solutions of the moist Boussinesq system (2.7a)–(2.7c).
Lemma A.21. Let w ˉ , θ ˉ , q ˉ : [ 0 , ∞ ) → R \bar{w},\bar{\theta},\bar{q} : [0,\infty) \to\mathbb{R} w ˉ , θ ˉ , q ˉ : [ 0 , ∞ ) → R be weak solutions of the ODE system (A.13a)–(A.13c). If we define
u ( t , x ) = w ˉ ( t ) e 3 , θ ( t , x ) = θ ˉ ( t ) , and q ( t , x ) = q ˉ ( t ) u(t,x)=\bar{w}(t)e_{3},\quad\theta(t,x)=\bar{\theta}(t),\quad\text{and}\quad q(t,x)=\bar{q}(t) u ( t , x ) = w ˉ ( t ) e 3 , θ ( t , x ) = θ ˉ ( t ) , and q ( t , x ) = q ˉ ( t ) then ( u , θ , q ) (u,\theta,q) ( u , θ , q ) is a weak solution of the moist Boussinesq system (2.7a)–(2.7c) with ϵ = 1 \epsilon= 1 ϵ = 1 .
Proof. When u u u , θ \theta θ , and q q q are constant in space the advective terms vanish. Since the pressure necessarily solves − Δ p = ∇ ⋅ ( u ⋅ ∇ u ) -\Delta p = \nabla\cdot(u\cdot\nabla u) − Δ p = ∇ ⋅ ( u ⋅ ∇ u ) it must therefore also vanish. Finally, since e 3 × u = u h ⊥ e_{3}\times u=u^{\perp}_{h} e 3 × u = u h ⊥ , which also vanishes when u = w ˉ e 3 u=\bar{w}e_{3} u = w ˉ e 3 , the claim follows.
We are finally ready to show that the moist Boussinesq system (2.7a)–(2.7c) admits solutions with nonlinear oscillations whose time averages do not vanish.
Corollary A.22. The moist Boussinesq system (2.7a)–(2.7c) admits periodic weak solutions with nonzero averages over their period. Moreover these solutions may be chosen to belong to the wave set W W W at every instant (for W W W as defined in Proposition 3.3).
Proof. This follows from Corollary A.20 and Lemma A.21 since together they guarantee the existence of weak solutions of the moist Boussinesq system (2.7a)–(2.7c) with M : = θ + q = 0 M := \theta+ q = 0 M := θ + q = 0 . Moreover we know from Lemma A.21 that these solutions take the form u = w ˉ e 3 u = \bar{w}e_{3} u = w ˉ e 3 and θ = θ ˉ \theta= \bar{\theta} θ = θ ˉ . Therefore these solutions have vanishing potential vorticity, and they satisfy M = 0 M = 0 M = 0 . Proposition 3.3 therefore tells us that, indeed, these solutions belong to the wave set W W W .
To conclude this section we record a result analogous to Lemma A.21 above: now the full velocity is incorporated into the ODE system, not just its vertical component. This is of independent interest from Lemma A.21 since it is used in the study of the Snell-type metric in Section 5.2.
Lemma A.23. Let u ˉ \bar{u} u ˉ , θ ˉ \bar{\theta} θ ˉ , q ˉ : [ 0 , ∞ ) → R \bar{q} : [0,\infty) \to\mathbb{R} q ˉ : [ 0 , ∞ ) → R be weak solutions of the ODE system
d u d t = − u h ⊥ + ( θ − min 0 q ) e 3 (96) \frac{du}{dt}=-u_h^{\perp}+(\theta-\min_{0}q)e_3 \tag*{(96)} d t d u = − u h ⊥ + ( θ − 0 min q ) e 3 (96) d θ d t = − u 3 , and (97) \frac{d\theta}{dt}=-u_3,\qquad\text{and} \tag*{(97)} d t d θ = − u 3 , and (97) d q d t = u 3 (98) \frac{dq}{dt}=u_3 \tag*{(98)} d t d q = u 3 (98) If we define u ( t , x ) = u ˉ ( t ) u(t,x)=\bar{u}(t) u ( t , x ) = u ˉ ( t ) , θ ( t , x ) = θ ˉ ( t ) \theta(t,x)=\bar{\theta}(t) θ ( t , x ) = θ ˉ ( t ) , and q ( t , x ) = q ˉ ( t ) q(t,x)=\bar{q}(t) q ( t , x ) = q ˉ ( t ) then ( u , θ , q ) (u,\theta,q) ( u , θ , q ) is a weak solution of the moist Boussinesq system (2.7a)–(2.7c) with ε = 1 \varepsilon=1 ε = 1 .
Proof. This follows exactly as for Lemma A.21 .
References