betaFeedback

WorksMathematics

Curvilinear coordinates and curvature in radiative transport

Abstract

We derive a general expression for the streaming term in radiative transport equa- tions and other transport problems when formulated in curvilinear coordinates, emphasizing coordinate systems adapted to the geometry of the domain and the directional dependence of particle transport. By parametrizing the angular vari- able using a local orthonormal frame, we express directional derivatives in terms of curvature-related quantities that reflect the geometry of underlying spatial man- ifolds. Our formulation highlights how the interaction between coordinate choices and curvature influences the streaming operator, offering geometric interpretations of its components. The resulting framework offers intuitive insight into when and how angular dependence can be simplified and may guide the selection of coordinate systems that balance analytical tractability and computational efficiency.

Introduction

Equations of the form

1c∂Ψ∂t+Ω⋅∇Ψ+G[Ψ]=Q,(1)\frac{1}{c}\frac{\partial\Psi}{\partial t}+\boldsymbol{\Omega}\cdot\nabla\Psi+\mathcal{G}[\Psi]=Q, \tag*{(1)}

arise in a wide variety of physical and engineering contexts involving the directional transport of particles, energy or other agents. Here, Ψ=Ψ(r,Ω,t)\Psi=\Psi(\mathbf{r},\boldsymbol{\Omega},t) denotes a density or distribution function depending on spatial position r∈R3\mathbf{r}\in\mathbb{R}^{3}, direction of movement or momentum Ω∈S2\boldsymbol{\Omega}\in S^{2}, and time tt. The streaming term Ω⋅∇Ψ\boldsymbol{\Omega}\cdot\nabla\Psi captures advective transport along characteristics determined by Ω\boldsymbol{\Omega}, while the operator G[Ψ]\mathcal{G}[\Psi] accounts for local or nonlocal interactions such as absorption, scattering, collisions, or relaxation effects. The source term QQ models external or internal production mechanisms.

Equation (1) and closely related formulations find application in a wide range of domains. Examples range from rarefied gas dynamics where it appears as the Boltzmann equation describing non-equilibrium molecular flows Cercignani [5], to plasma physics, where the Vlasov equation models the evolution of charged particle distributions under electromagnetic fields Nicholson [16]. The phonon Boltzmann equation describes microscale heat transport in solids Majumdar [15], and similar forms govern the propagation of cosmic rays in astrophysics Schlickeiser [17]. Radiative transfer models photon transport in stellar and atmospheric environments Chandrasekhar [6], as well as in optical tomography Arridge [1]. While in the context of biological and social systems, kinetic equations with directional transport can be used to describe traffic flow, swarming, and crowd dynamics Bellomo et al. (2013).

Despite their disparate origins, these systems all share a common mathematical core structured around the advective transport term Ω⋅∇Ψ\boldsymbol{\Omega}\cdot\nabla\Psi. This term fundamentally couples direction and space, giving rise to anisotropy, long-range effects, and complex interactions that distinguish kinetic equations from classical PDEs in purely spatial domains.

As with many problems in physics and engineering, it is common practice to address this problem using coordinates tailored to the geometry of the domain. For instance, spherical coordinates for scenarios where most parameters exhibit rotational symmetry. Notably, equation (1) offers additional degrees of freedom in choosing coordinates for Ω\boldsymbol{\Omega}, beyond those for r\boldsymbol{r}. However, this introduces added complexity: the optimal representation of r\boldsymbol{r} and Ω\boldsymbol{\Omega} may depend on each other.

For Ω\boldsymbol{\Omega}, a vector on the unit sphere, a common parametrization is the standard spherical coordinates (polar and azimuthal angles) or equivalently:

Ω(ω,μ)=μn+1−μ2cos⁡(ω)t+1−μ2sin⁡(ω)b,(2)\boldsymbol{\Omega}(\omega,\mu)=\mu\boldsymbol{n}+\sqrt{1-\mu^2}\cos(\omega)\boldsymbol{t} +\sqrt{1-\mu^2}\sin(\omega)\boldsymbol{b}, \tag*{(2)}

where ω∈[0,2π]\omega\in[0,2\pi], μ∈[−1,1]\mu\in[-1,1], and n,t,b∈S2\boldsymbol{n},\boldsymbol{t},\boldsymbol{b}\in\mathbb{S}^2 form an orthogonal basis. While this formulation may appear convoluted, it simply means that Ω\boldsymbol{\Omega} is parametrized by its projection μ\mu onto a designated vector n\boldsymbol{n} and a polar angle ω\omega measured in a plane orthogonal to n\boldsymbol{n} from a reference vector t\boldsymbol{t} in that plane. A visualization is provided in Figure 1.

Visualization of the parametrization of $\Omega(\mu,\omega) \in S^2$

Figure 1. Visualization of the parametrization of Ω(μ,ω)∈S2\Omega(\mu,\omega) \in S^2: μ=Ω⋅n\mu= \Omega\cdot\mathbf{n} is the projection of Ω\Omega onto n\mathbf{n}, ω\omega measures angle from t\mathbf{t} in t−b\mathbf{t}-\mathbf{b}-plane. n,t,b\mathbf{n},\mathbf{t},\mathbf{b} might depend on r\mathbf{r} making μ\mu amd ω\omega functions of r\mathbf{r} even for fixed Ω\Omega

Given a parametrization, i.e., a set of coordinates for r∈R3\boldsymbol{r}\in\mathbb{R}^3, say r=r(u1,u2,u3)\boldsymbol{r}=\boldsymbol{r}(u^1,u^2,u^3), a natural choice for n,t\boldsymbol{n},\boldsymbol{t}, and b\boldsymbol{b} would be the canonical tangent vectors n=c1∂r∂u1\boldsymbol{n}=c_1\frac{\partial\boldsymbol{r}}{\partial u^1}, t=c2∂r∂u2\boldsymbol{t}=c_2\frac{\partial\boldsymbol{r}}{\partial u^2}, and b=n×t\boldsymbol{b}=\boldsymbol{n}\times\boldsymbol{t}, where the coefficients c1,c2c_1,c_2 ensure normalization. Note that this canonical choice is not necessarily orthogonal; if it is not, additional modifications are required. Orthogonalization via Gram-Schmidt is one possible approach. Other choices in terms of contra and covariant coordinates of the canonical tangent vectors, guaranteeing orthogonality, are described in Freimanis [8]. Moving forward, we will not rely on any specific choice of n,t\boldsymbol{n},\boldsymbol{t}, and b\boldsymbol{b}, but will assume they form an integrable orthonormal vector field.

The parametrization of Ω\boldsymbol{\Omega} presented in (2) simplifies integrals over Ω\boldsymbol{\Omega}, which frequently appear in scattering operators S\mathcal{S}, as:

∫S2(⋅) dΩ=∫−11∫02π(⋅) dω dμ.(3)\int_{\mathbb{S}^2}(\cdot)\,\mathrm{d}\boldsymbol{\Omega} =\int_{-1}^{1}\int_{0}^{2\pi}(\cdot)\,\mathrm{d}\omega\,\mathrm{d}\mu. \tag*{(3)}

A more significant advantage, however, is that n\boldsymbol{n} can often be chosen as the normal vector to level sets of Ψ\Psi, leading to cases where Ψ\Psi is effectively independent of ω\omega, i.e.,

Ψ(r,Ω,E,t)=Ψ(r,(ω,μ),E,t)=Ψ(r,μ,E,t).(4)\Psi(\boldsymbol{r},\boldsymbol{\Omega},E,t)=\Psi(\boldsymbol{r},(\omega,\mu),E,t)=\Psi(\boldsymbol{r},\mu,E,t). \tag*{(4)}

This reduction in the dimensionality of the parameter space can greatly simplify solving equations (1). However, achieving this simplification with constant n\mathbf{n} is generally only possible in trivial cases. In most scenarios, n,t\mathbf{n},\mathbf{t}, and b\mathbf{b} must depend on r\mathbf{r}, making the computation of the term Ω⋅∇Ψ\Omega\cdot\nabla\Psi particularly cumbersome.

In the literature, the exact form of this term for a given choice of coordinates for r\mathbf{r} and Ω\Omega is often derived through tedious geometric arguments. These derivations can be challenging to follow and even harder to adapt to slightly different geometries. Examples of such derivations can be seen in [2, 12]

In the next section, we derive general expressions for Ω⋅∇Ψ\Omega\cdot\nabla\Psi for arbitrary choices of r\mathbf{r} and n,t,b\mathbf{n},\mathbf{t},\mathbf{b} in a purely analytical manner. We denote by ∇r\nabla^{r} the partial derivatives in space and identify

Ω⋅∇Ψ=Ω⋅∇rΨ+∂Ψ∂μΩ⋅∇μ+∂Ψ∂ωΩ⋅∇ω,(5)\Omega\cdot\nabla\Psi= \Omega\cdot\nabla^{r}\Psi+ \frac{\partial\Psi}{\partial\mu}\Omega\cdot\nabla\mu+ \frac{\partial\Psi}{\partial\omega}\Omega\cdot\nabla\omega, \tag*{(5)}

presenting alternative formulations for the terms Ω⋅∇μ\Omega\cdot\nabla\mu and Ω⋅∇ω\Omega\cdot\nabla\omega in section 2. We relate these expressions to the geometric properties of curves and surfaces embedded in R3\mathbb{R}^{3}, which we believe will make the process of deriving Ω⋅∇Ψ\Omega\cdot\nabla\Psi more straightforward, in some cases, and provides additional intuition and guidance in general. Namely we will rewrite the terms Ω⋅∇μ\Omega\cdot\nabla\mu and Ω⋅∇ω\Omega\cdot\nabla\omega as terms looking as follows:

Ω⋅∇μ=(1−μ2)The curvature of some manifold X+μ1−μ2The curvature of a curve orthogonal to X(6)\Omega\cdot\nabla\mu= (1-\mu^{2})\boxed{\text{The curvature of some manifold }X} + \mu\sqrt{1-\mu^{2}}\boxed{\text{The curvature of a curve orthogonal to }X} \tag*{(6)}

and

Ω⋅∇ω=1−μ2 The curvature of some curves in X+μNon-curvature term quantifying spinning of t around n.(7)\boldsymbol{\Omega}\cdot\nabla\omega=\sqrt{1-\mu^2}\,\boxed{\text{The curvature of some curves in }X}+\mu\boxed{\text{Non-curvature term quantifying spinning of }t\text{ around }n}. \tag*{(7)}

In Freimanis [8] equation (5) is calculated for arbitrary coordinates rr, but restricted to two choices of nn, tt, bb relative to the chosen coordinates, and with focus on the coordinate representation alone. We pose severely fewer restrictions on nn, tt, bb and focus on interpretations regarding the embeddings into R3\mathbb{R}^3 rather than purely on the coordinate level. In § 3, we provide examples illustrating this approach in some classical settings and some less common scenarios.

Derivation of Ω⋅∇Ψ\boldsymbol{\Omega}\cdot\nabla\Psi

In this section, we set out to calculate an expression Ω⋅∇Ψ\boldsymbol{\Omega}\cdot\nabla\Psi when Ω\boldsymbol{\Omega} is parametrized as

Ω(r)=μn+1−μ2cos⁡(ω)t+1−μ2sin⁡(ω)b,(8)\boldsymbol{\Omega}(r)=\mu n+\sqrt{1-\mu^2}\cos(\omega)t+\sqrt{1-\mu^2}\sin(\omega)b, \tag*{(8)}

where n,t∈C1(R3,R3)n,t\in C^1(\mathbb{R}^3,\mathbb{R}^3), and b(r):=n×tb(r):=n\times t are orthonormal, differentiable vector fields and thus Ψ(r,Ω,E,t)=Ψ(r,(ω(r),μ(r)),E,t)\Psi(r,\boldsymbol{\Omega},E,t)=\Psi(r,(\omega(r),\mu(r)),E,t). Given coordinates for rr it is not always the simplest choice to guarantee all these conditions. Seemingly obvious choices might have neither normality nor orthogonality. We will stick to these assumptions regardless and discuss possible fixes, if they do not hold in some of the examples in § 3.

We will begin by establishing some notation and simple properties of these vector fields, as well as ω\omega and μ\mu, that will be helpful in later calculations. We will then proceed to calculate Ω⋅∇Ψ\boldsymbol{\Omega}\cdot\nabla\Psi, providing several alternative forms for most terms involved, allowing practitioners to choose the one best suited to their needs. Throughout the bold letters will be used for vectors in R3\mathbb{R}^3. Further, given a vector h1,h2∈R3h_1,h_2\in\mathbb{R}^3, we denote by h1⋅h2h_1\cdot h_2 their standard inner product, by ∥h1∥=h1⋅h1\lVert h_1\rVert=\sqrt{h_1\cdot h_1} the euclidean norm and their normalization as h^1:=h1∥h1∥\hat{h}_1:=\frac{h_1}{\lVert h_1\rVert}. We will denote the directional derivative in direction h1h_1 as ∇h1:=h1⋅∇\nabla_{h_1}:=h_1\cdot\nabla. Lastly, let us define the projection of Ω(r)\boldsymbol{\Omega}(r) onto the orthogonal complement of nn, i.e., onto the plane spanned by tt and bb, as Ω∥(r)\boldsymbol{\Omega}_{\parallel}(r). Specifically,

Ω∥(r):=1−μ2cos⁡(ω)t+1−μ2sin⁡(ω)b=Ω(r)−μn.(9)\boldsymbol{\Omega}_{\parallel}(r):=\sqrt{1-\mu^2}\cos(\omega)t+\sqrt{1-\mu^2}\sin(\omega)b=\boldsymbol{\Omega}(r)-\mu n. \tag*{(9)}

We will analyze equation (5) piece by piece and start with the term Ω⋅∇μ=∇Ωμ\boldsymbol{\Omega}\cdot\nabla\mu=\nabla_{\boldsymbol{\Omega}}\mu in section 2.1.

Between ∇Ωμ\nabla_{\boldsymbol{\Omega}}\mu and ∇Ωω\nabla_{\boldsymbol{\Omega}}\omega the prior is arguably more important, since terms stemming from the second one can often be ignored if nn is chosen orthogonal to level sets of Ψ\Psi, which under mild conditions leads to ∂Ψ/∂ω=0\partial\Psi/\partial\omega=0. (Compare A.1) For cases where ∂Ψ/∂ω≠0\partial\Psi/\partial\omega\ne0 we will derive an expression for ∇Ωω\nabla_{\boldsymbol{\Omega}}\omega in § 2.2. Most terms in the streaming will be related to curvatures of either curves or surfaces, or both, defined by the vectors nn, tt and bb, which lends itself to a more intuitive approach to choosing coordinates.

Derivation of Ω⋅∇μ\boldsymbol{\Omega} \cdot\nabla\mu

In this section we set out to further evaluate and interpret the term ∇Ωμ\nabla_{\boldsymbol{\Omega}}\mu in equation (5).

Since μ=Ω⋅n\mu= \boldsymbol{\Omega} \cdot\mathbf n we first find

∇Ωμ=Ω⋅∇Ωn=Ω∥⋅(∇Ω∥n+μ∇nn)+μn⋅∇Ωn(10)\nabla_{\boldsymbol{\Omega}}\mu= \boldsymbol{\Omega} \cdot\nabla_{\boldsymbol{\Omega}}\mathbf n = \boldsymbol{\Omega}_{\parallel} \cdot\left(\nabla_{\boldsymbol{\Omega}_{\parallel}}\mathbf n + \mu\nabla_{\mathbf n}\mathbf n\right) + \mu\mathbf n \cdot\nabla_{\boldsymbol{\Omega}}\mathbf n \tag*{(10)}

using the linearity of the derivative and the inner product. Note that the last term μn⋅∇Ωn=0\mu\mathbf n \cdot\nabla_{\boldsymbol{\Omega}}\mathbf n = 0, because n\mathbf n is assumed to be normalized. It also is a consequence of a more general property of orthonormal frames detailed in the following remark.

Remark 1. Properties of orthonormal frames

Let h∈S2\mathbf h \in S^2 and let u:R3→S2\mathbf u:\mathbb{R}^3 \to S^2 and v:R3→S2\mathbf v:\mathbb{R}^3 \to S^2 be distinct, differentiable orthonormal vector fields, i.e. u⋅u=v⋅v=1\mathbf u \cdot\mathbf u = \mathbf v \cdot\mathbf v = 1 and u⋅v=0\mathbf u \cdot\mathbf v = 0 everywhere. Since these inner products are constant in space their derivatives disappear and therefore

u⋅∇hu=0(11)\mathbf u \cdot\nabla_{\mathbf h}\mathbf u = 0 \tag*{(11)}

and

u⋅∇hv=−v⋅∇hu.(12)\mathbf u \cdot\nabla_{\mathbf h}\mathbf v = -\mathbf v \cdot\nabla_{\mathbf h}\mathbf u. \tag*{(12)}

Note that this holds for u,v∈{n,t,b}\mathbf u,\mathbf v \in\{\mathbf n,\mathbf t,\mathbf b\}.

Using Remark 1 we are left to simplify

∇Ωμ=Ω∥⋅∇Ω∥n+μ(Ω∥⋅∇nn)(13)\nabla_{\boldsymbol{\Omega}}\mu= \boldsymbol{\Omega}_{\parallel} \cdot\nabla_{\boldsymbol{\Omega}_{\parallel}}\mathbf n + \mu(\boldsymbol{\Omega}_{\parallel} \cdot\nabla_{\mathbf n}\mathbf n) \tag*{(13)}

The second term can be expressed in terms of the normal curvature of an integral curve.

Remark 2. Integral Curves and Normal Curvature

Given a differentiable vector field u:R3→S2\mathbf u:\mathbb{R}^3 \to S^2 let γu(r,τ)\gamma^{\mathbf u}(\mathbf r,\tau) be the integral curve of u\mathbf u passing through r\mathbf r, i.e. let it solve the following differential equation:

dγu(τ,r)dτ=u(γu(τ,r)),τ∈(−ε,ε)\frac{d\gamma^{\mathbf u}(\tau,\mathbf r)}{d\tau} = \mathbf u\left(\gamma^{\mathbf u}(\tau,\mathbf r)\right), \qquad\tau\in(-\varepsilon,\varepsilon)
γu(0,r)=r(14)\gamma^{\mathbf u}(0,\mathbf r) = \mathbf r \tag*{(14)}

Existence of γu\gamma^{\mathbf u} is guaranteed by Picard-Lindelöf, but since we will only use this curve locally we simply note that we can solve this differential equation in a neighborhood of γu(0,r)\gamma^{\mathbf u}(0,\mathbf r). We shall further, sloppily, refer to γu(0,r)\gamma^{\mathbf u}(0,\mathbf r) as γu(r)\gamma^{\mathbf u}(\mathbf r). Since u\mathbf u is a unit vector everywhere γu\gamma^{\mathbf u} is parametrized by arc length and has a normal curvature vector

κu(r):=−d2γu(τ,r)dτ2∣τ=0=−du(γu(τ,r))dτ∣τ=0=−(u(r)⋅∇)u(r)=−∇uu(r).(15)\boldsymbol{\kappa}^{\mathbf u}(\mathbf r) := -\left.\frac{d^2\gamma^{\mathbf u}(\tau,\mathbf r)}{d\tau^2}\right|_{\tau=0} = -\left.\frac{d\mathbf u(\gamma^{\mathbf u}(\tau,\mathbf r))}{d\tau}\right|_{\tau=0} = -(\mathbf u(\mathbf r)\cdot\nabla)\mathbf u(\mathbf r) = -\nabla_{\mathbf u}\mathbf u(\mathbf r). \tag*{(15)}

(The sign convention here is chosen such that were γu\gamma^{\mathbf u} a curve on a sphere with outward normal N\mathbf N, the value of κu⋅N\boldsymbol{\kappa}^{\mathbf u}\cdot\mathbf N would be positive.)

Setting u=nu=n in the previous remark, we find that the second term in equation (13) is

μ(Ω∥⋅∇nn)=μ1−μ2(Ω^∥⋅∇nn)=μ1−μ2(Ω^∥⋅κn)(16)\mu(\boldsymbol{\Omega}_{\parallel}\cdot\nabla_n n)=\mu\sqrt{1-\mu^2}(\hat{\boldsymbol{\Omega}}_{\parallel}\cdot\nabla_n n)=\mu\sqrt{1-\mu^2}\left(\hat{\boldsymbol{\Omega}}_{\parallel}\cdot\boldsymbol{\kappa}^n\right) \tag*{(16)}

expressing the term entirely in terms of μ\mu and the curvature of the integral curve of nn in the direction of Ω∥\boldsymbol{\Omega}_{\parallel}. This term quantifies how much γn\gamma^n curves in direction Ω∥\boldsymbol{\Omega}_{\parallel}. It disappears, if γn\gamma^n is a straight line. If we imagine a set of surfaces orthogonal to γn\gamma^n along the curve, this is simultaneously a measure for how parallel such surfaces are along the curve.

As an example, if n(r)n(\boldsymbol{r}) is chosen to be the normalized gradient of f(r)=∥r∥2f(\boldsymbol{r})=\lVert\boldsymbol{r}\rVert^2, the term Ω∥⋅κn=0\boldsymbol{\Omega}_{\parallel}\cdot\boldsymbol{\kappa}^n=0 everywhere, since the level sets of ff are spheres centered at the origin and all these spheres are, locally parallel along γn\gamma^n. On the other hand, if nn is chosen to be the normalized gradient of f(r)=rTArf(\boldsymbol{r})=\boldsymbol{r}^{\mathsf{T}}A\boldsymbol{r}, where AA is some positive definite matrix with at least two distinct eigenvalues, then Ω∥⋅κn\boldsymbol{\Omega}_{\parallel}\cdot\boldsymbol{\kappa}^n will not vanish. In the latter case, the level sets of ff are ellipses, excluding spheres, which are not parallel along γn\gamma^n. This is illustrated in Figure 2. We will see these two specific examples in § 3.

2D visualization of $\gamma^n$ in spherical coordinates and elliptical coordinates

Figure 2. 2D visualization of γn\gamma^n in spherical coordinates and elliptical coordinates. The integral curves γn\gamma^n following the normal vectors of circles are straight lines, while those of ellipses are curved.

Having established the notion of curvature of integral curves we shall first express the remaining term in equation (5) in terms of integral curves as well. Recall the definition of Ω∥\boldsymbol{\Omega}_{\parallel} and therefor note that its normalization is given by Ω^∥=cos⁡(ω)t+sin⁡(ω)b\hat{\boldsymbol{\Omega}}_{\parallel}=\cos(\omega)t+\sin(\omega)b. Thus using the linearity of the derivative and the inner product again we can now also tackle the first term in equation (13) and see that

Ω∥⋅∇Ω∥n=(1−μ2)(Ω^∥⋅∇Ω^∥n)=(1−μ2)[cos⁡2(ω)(t⋅∇tn)+sin⁡2(ω)(b⋅∇bn)+sin⁡(ω)cos⁡(ω)(t⋅∇bn+b⋅∇tn)](17)\begin{aligned} \boldsymbol{\Omega}_{\parallel}\cdot\nabla_{\boldsymbol{\Omega}_{\parallel}}n&=(1-\mu^2)\left(\hat{\boldsymbol{\Omega}}_{\parallel}\cdot\nabla_{\hat{\boldsymbol{\Omega}}_{\parallel}}n\right)\\ &=(1-\mu^2)\left[\cos^2(\omega)(t\cdot\nabla_t n)+\sin^2(\omega)(b\cdot\nabla_b n)+\sin(\omega)\cos(\omega)(t\cdot\nabla_b n+b\cdot\nabla_t n)\right] \tag*{(17)} \end{aligned}

Using Remark 1 the last set of derivatives can be rewritten as t⋅∇bn+b⋅∇tn=t⋅∇bn−t⋅∇bn=0\mathbf{t} \cdot\nabla_{\mathbf{b}}\mathbf{n} + \mathbf{b} \cdot\nabla_{\mathbf{t}}\mathbf{n} = \mathbf{t} \cdot\nabla_{\mathbf{b}}\mathbf{n} - \mathbf{t} \cdot\nabla_{\mathbf{b}}\mathbf{n} = 0, causing the last summand to disappear. Further using Remark 1 again we can replace t⋅∇tn=−n⋅∇tt\mathbf{t} \cdot\nabla_{\mathbf{t}}\mathbf{n} = -\mathbf{n} \cdot\nabla_{\mathbf{t}}\mathbf{t} and b⋅∇bn=−n⋅∇bb\mathbf{b} \cdot\nabla_{\mathbf{b}}\mathbf{n} = -\mathbf{n} \cdot\nabla_{\mathbf{b}}\mathbf{b} and get

Ω∥⋅∇Ω∥n=−(1−μ2)[cos⁡(ω)2(n⋅∇tt)+sin⁡(ω)2(n⋅∇bb)]=(1−μ2)[cos⁡(ω)2n⋅κt+sin⁡(ω)2n⋅κb].(18)\begin{aligned} \boldsymbol{\Omega}_{\parallel} \cdot\nabla_{\boldsymbol{\Omega}_{\parallel}}\mathbf{n} &= -(1-\mu^2)\left[\cos(\omega)^2(\mathbf{n}\cdot\nabla_{\mathbf{t}}\mathbf{t})+\sin(\omega)^2(\mathbf{n}\cdot\nabla_{\mathbf{b}}\mathbf{b})\right]\\ &= (1-\mu^2)\left[\cos(\omega)^2\mathbf{n}\cdot\boldsymbol{\kappa}^{\mathbf{t}}+\sin(\omega)^2\mathbf{n}\cdot\boldsymbol{\kappa}^{\mathbf{b}}\right]. \tag*{(18)} \end{aligned}

In the last step Remark 2 was used to identify ∇tt=κt\nabla_{\mathbf{t}}\mathbf{t}=\boldsymbol{\kappa}^{\mathbf{t}} and ∇bb=κb\nabla_{\mathbf{b}}\mathbf{b}=\boldsymbol{\kappa}^{\mathbf{b}} as the curvatures of the corresponding integral curves.

While the interpretation in terms of the curvature of integral curves already offers some intuition, we would like to note that n⋅κt\mathbf{n}\cdot\boldsymbol{\kappa}^{\mathbf{t}} and n⋅κb\mathbf{n}\cdot\boldsymbol{\kappa}^{\mathbf{b}} can also be interpreted as the curvature of a manifold orthogonal to n\mathbf{n} in direction t\mathbf{t} and b\mathbf{b} respectively, if such a manifold exists. We will continue to look at Ω∥⋅∇Ω∥n\boldsymbol{\Omega}_{\parallel}\cdot\nabla_{\boldsymbol{\Omega}_{\parallel}}\mathbf{n} a little longer to make its relation to the shape of such manifolds orthogonal to n\mathbf{n} clearer. To that end, we will establish under which conditions such orthogonal manifolds exist and introduce the shape operator of a manifold.

Remark 3. Induced Foliation

Let V:R3→S2\mathbf{V}:\mathbb{R}^{3}\to S^{2} be a differentiable normalized vector field and let U⊂R3U\subset\mathbb{R}^{3} be a connected open set. If V⋅rot⁡(V)=0\mathbf{V}\cdot\operatorname{rot}(\mathbf{V})=0 on UU, then for each r∈U\mathbf{r}\in U there is a manifold such that r∈XrV\mathbf{r}\in X_{\mathbf{r}}^{\mathbf{V}} and so that V(r′)\mathbf{V}(\mathbf{r}') is the normal vector of XX at all r′∈XrV\mathbf{r}'\in X_{\mathbf{r}}^{\mathbf{V}}. Since r′∈XrV\mathbf{r}'\in X_{\mathbf{r}}^{\mathbf{V}} implies XrV=Xr′VX_{\mathbf{r}}^{\mathbf{V}}=X_{\mathbf{r}'}^{\mathbf{V}} this introduces an equivalence relation on UU called a foliation, with equivalence classes, called leaves, being surfaces normal to V\mathbf{V}. Due to this we will drop the r\mathbf{r} from XrVX_{\mathbf{r}}^{\mathbf{V}} going forward and write XVX^{\mathbf{V}} to mean the leaf of the foliation that is orthogonal to V\mathbf{V} at r\mathbf{r}. In R3\mathbb{R}^{3} the condition V⋅rot⁡(V)=0\mathbf{V}\cdot\operatorname{rot}(\mathbf{V})=0 is both necessary and sufficient for the existence of a foliation. This follows from Frobenius theorem [9, 13] applied to vector fields in R3\mathbb{R}^{3}.

A maybe more applicable, yet still sufficient (not necessary condition), for the existence of XVX^{\mathbf{V}} can be expressed as follows: Let r(u1,u2,u3)\mathbf{r}(u^{1},u^{2},u^{3}) be a (local) parametrization of R3\mathbb{R}^{3}, i.e. let (u1,u2,u3)(u^{1},u^{2},u^{3}) be the (local) coordinates. If 0≠V∝∂r∂u1\mathbf{0}\ne\mathbf{V}\propto\frac{\partial\mathbf{r}}{\partial u^{1}} then XV(u2,u3)=r(uˉ1,u2,u3)X^{\mathbf{V}}(u^{2},u^{3})=\mathbf{r}(\bar{u}^{1},u^{2},u^{3}) is a parametrization of XVX^{\mathbf{V}} containing the point rˉ=r(uˉ1,uˉ2,uˉ3)\bar{\mathbf{r}}=\mathbf{r}(\bar{u}^{1},\bar{u}^{2},\bar{u}^{3}), i.e. fixing one of the coordinates of r\mathbf{r}, but altering the others generates a surface containing r\mathbf{r}.

Remark 4. Manifolds and their Shape Operator

Let XN:R2⊃U→R3X^{\mathbf{N}}:\mathbb{R}^{2}\supset U\to\mathbb{R}^{3} be the parametrization of an embedded surface X(u1,u2)\mathbf{X}(u^{1},u^{2}) with parameters (u1,u2)∈U(u^{1},u^{2})\in U and tangent vectors Ti=∂X∂ui\mathbf{T}_{i}=\frac{\partial\mathbf{X}}{\partial u^{i}} (i=1,2)(i=1,2) and normal vector N∈R3\mathbf{N}\in\mathbb{R}^{3}. For simplicity assume N\mathbf{N} is not just defined and differentiable on X\mathbf{X}, but on an open neighborhood UU of of X⊂U⊂R3\mathbf{X}\subset U\subset\mathbb{R}^{3}. The metric tensor gg of X\mathbf{X} can be expressed in coordinates as the matrix gα,β=Tα⋅Tβg_{\alpha,\beta}=\mathbf{T}_{\alpha}\cdot\mathbf{T}_{\beta} where α,β∈{u1,u2}\alpha,\beta\in\{u^{1},u^{2}\}, while the second fundamental form hh of X\mathbf{X} is given by the matrix defined through hα,β=−∂Tα∂β⋅N=−∂2X∂α∂β⋅Nh_{\alpha,\beta}=-\frac{\partial\mathbf{T}_{\alpha}}{\partial\beta}\cdot\mathbf{N}=-\frac{\partial^{2}\mathbf{X}}{\partial\alpha\partial\beta}\cdot\mathbf{N}. From gg and hh the shape operator, or Weingarten map SXS_{\mathbf{X}}, can be calculated as SX=g−1hS_{\mathbf{X}}=g^{-1}h. The shape operator contains information about the curvature of the manifold X\mathbf{X}. Given a unit tangent vector u∈TX\mathbf{u}\in T_{\mathbf{X}} the curvature in direction u\mathbf{u} can be calculated as uTSXu\mathbf{u}^{T}S_{\mathbf{X}}\mathbf{u}. The shape operator shares its Eigenvalues with hh, meaning Eigenvalues of SXS_{\mathbf{X}} are the principal curvatures of X\mathbf{X}, its determinant is the Gauss curvature and its trace is the mean curvature of X\mathbf{X}.

By Weingarten’s theorem, the coordinate representation of SXS_X can also be calculated as

(SX)αβ=Tα⋅∇TβNα,β∈{u1,u2}.(19)\left(S_X\right)_{\alpha}^{\beta}=T_{\alpha}\cdot\nabla_{T_{\beta}}N \qquad\alpha,\beta\in\{u^1,u^2\}. \tag*{(19)}

Weingarten [18]; Kreyszig [11]

With notation from Remarks 3 and 4 and N=nN=n and T1=tT_1=t and T2=bT_2=b we can interpret the derivative terms in equation (10) as components of the shape operator SXnS_{X^n} of the manifold XnX^n and thus

Ω∥⋅∇Ω∥n=(1−μ2)[(cos⁡(ω),sin⁡(ω))⋅SXn⋅(cos⁡(ω)sin⁡(ω))]=:(1−μ2)C(r,ω)(20)\begin{aligned} \Omega_{\parallel}\cdot\nabla_{\Omega_{\parallel}}n &=(1-\mu^2)\left[(\cos(\omega),\sin(\omega))\cdot S_{X^n}\cdot \begin{pmatrix} \cos(\omega)\\ \sin(\omega) \end{pmatrix}\right]\\ &=: (1-\mu^2)C(r,\omega) \tag*{(20)} \end{aligned}

Note that (cos⁡(ω),sin⁡(ω))(\cos(\omega),\sin(\omega)) is merely Ω^∥\widehat{\Omega}_{\parallel} restricted to the tangent space of XnX^n and that therefore the term in brackets is nothing but the curvature of XnX^n in direction Ω^∥\widehat{\Omega}_{\parallel}, which we denoted by C(r,ω)C(r,\omega). A visualization can be seen in Figure 3.

Example manifold $X^n$ with normal vector $n$ and tangent vectors $t$ and $\Omega_{\parallel}$

Figure 3. Example manifold XnX^n with normal vector nn and tangent vectors tt and Ω∥\Omega_{\parallel}. The curvature at rr in direction ω\omega relative to tt, denoted by C(r,ω)C(r,\omega) is the normal curvature of red curve starting at rr.

Overall we found two curvature forms for ∇Ωμ\nabla_{\Omega}\mu in

∇Ωμ=(1−μ2)[cos⁡(ω)2(n⋅κt)+sin⁡(ω)2(n⋅κb)]+μ1−μ2(Ω^∥⋅κn)=(1−μ2)C(r,ω)+μ1−μ2(Ω^∥⋅κn).(21)\begin{aligned} \nabla_{\Omega}\mu &=(1-\mu^2)\left[\cos(\omega)^2\left(n\cdot\kappa^t\right)+\sin(\omega)^2\left(n\cdot\kappa^b\right)\right]+\mu\sqrt{1-\mu^2}\left(\widehat{\Omega}_{\parallel}\cdot\kappa^n\right)\\ &=(1-\mu^2)C(r,\omega)+\mu\sqrt{1-\mu^2}\left(\widehat{\Omega}_{\parallel}\cdot\kappa^n\right). \tag*{(21)} \end{aligned}

We will refer to the first one of these as the curve-curvature form of ∇Ωμ\nabla_{\Omega}\mu and to the second one as the surface-curvature form of ∇Ωμ\nabla_{\Omega}\mu when using them in examples later on.

Derivation of Ω⋅∇ω\Omega\cdot\nabla\omega

We will continue to calculate the remaining term in equation (5), namely ∇Ωω\nabla_{\Omega}\omega. To this end note that ω=arccos⁡(Ω∥⋅t1−μ2)=arccos⁡(Ω^∥⋅t)\omega= \arccos\left(\frac{\Omega_{\parallel}\cdot t}{\sqrt{1-\mu^2}}\right) = \arccos\left(\hat{\Omega}_{\parallel}\cdot t\right) or alternatively ω=arcsin⁡(Ω∥⋅b1−μ2)=arcsin⁡(Ω^∥⋅b)\omega= \arcsin\left(\frac{\Omega_{\parallel}\cdot b}{\sqrt{1-\mu^2}}\right) = \arcsin\left(\hat{\Omega}_{\parallel}\cdot b\right). Between these two notions of ω\omega, the angle is uniquely determined up to multiples of 2π2\pi. We will see that both lead to the same derivative. With arcsin⁡(x)′=−arccos⁡(x)′=11−x2\arcsin(x)'=-\arccos(x)'=\frac{1}{\sqrt{1-x^2}} one sees either

∇Ωω=1cos⁡(ω)(Ω^∥⋅∇Ωb)=t⋅∇Ωb(22)\nabla_{\Omega}\omega=\frac{1}{\cos(\omega)}\left(\hat{\Omega}_{\parallel}\cdot\nabla_{\Omega}b\right)=t\cdot\nabla_{\Omega}b \tag*{(22)}

or

∇Ωω=1sin⁡(ω)(Ω^∥⋅∇Ωt)=−b⋅∇Ωt(23)\nabla_{\Omega}\omega=\frac{1}{\sin(\omega)}\left(\hat{\Omega}_{\parallel}\cdot\nabla_{\Omega}t\right)=-b\cdot\nabla_{\Omega}t \tag*{(23)}

In both of these equations, we used Remark 1, specifically that b∇b=t∇t=0b\nabla b=t\nabla t=0, along with the definition of Ω∥\Omega_{\parallel} to get to the right-hand side. Using Remark 1 again one can also conclude that equations (22) and (23) are indeed the same, where they are both defined. To provide additional interpretation and intuition we will manipulate this further to also derive curvature forms for this equation. First using the linearity of the derivative

∇Ωω=t∇Ωb=−b∇Ωt=1−μ2[cos⁡(ω)(t⋅∇tb)+sin⁡(ω)(t⋅∇bb)]+μ(t⋅∇nb)\begin{aligned} \nabla_{\Omega}\omega=t\nabla_{\Omega}b=-b\nabla_{\Omega}t \\ &=\sqrt{1-\mu^2}\left[\cos(\omega)(t\cdot\nabla_t b)+\sin(\omega)(t\cdot\nabla_b b)\right]+\mu(t\cdot\nabla_n b) \end{aligned}

is obtained. This is then rewritten, using Remark 1 and the definition of κu\kappa^u in Remark 2, as

∇Ωω=1−μ2[−cos⁡(ω)(b⋅∇tt)+sin⁡(ω)(t⋅∇bb)]+μ(t⋅∇nb)=1−μ2[cos⁡(ω)(b⋅κt)−sin⁡(ω)(t⋅κb)]+μ(t⋅∇nb)=∂Ω∥∂ω⋅(κt+κb)+μ(t⋅∇nb)(24)\begin{aligned} \nabla_{\Omega}\omega=\sqrt{1-\mu^2}\left[-\cos(\omega)(b\cdot\nabla_t t)+\sin(\omega)(t\cdot\nabla_b b)\right]+\mu(t\cdot\nabla_n b) \\ &=\sqrt{1-\mu^2}\left[\cos(\omega)(b\cdot\kappa^t)-\sin(\omega)(t\cdot\kappa^b)\right]+\mu(t\cdot\nabla_n b) \\ &=\frac{\partial\Omega_{\parallel}}{\partial\omega}\cdot(\kappa^t+\kappa^b)+\mu(t\cdot\nabla_n b) \tag*{(24)} \end{aligned}

The terms b⋅κtb\cdot\kappa^t and t⋅κbt\cdot\kappa^b are the curvatures of γt\gamma^t and γb\gamma^b within XnX^n. They are zero if γt\gamma^t and γb\gamma^b are geodesics and otherwise measure how rapidly these curves bend away from geodesics. An example where γb\gamma^b is not a geodesic and such κb⋅t≠0\kappa^b\cdot t\ne0 is depicted in figure 4

Notably the final term t⋅∇nb=−b⋅∇ntt\cdot\nabla_n b=-b\cdot\nabla_n t was not rewritten in terms of a curvature. It expresses how tt and bb spiral around the curve γn\gamma^n.

The previous section raises the question of whether a surface-curvature form of ∇Ω\nabla_{\Omega} exists as well. Going back to equation (24) we can indeed identify the terms t⋅∇tbt\cdot\nabla_t b and t⋅∇nbt\cdot\nabla_n b in terms of the shape operator SXbS_{X^b} of the surface XbX^b, if it exists.(Compare conditions of existence in Remark 3). The equation then becomes

Example of non-geodesic integral curve $\gamma^u$ (blue) along a tangent vector field $u$ (magenta), with a geodesic starting at a common point (green)

Figure 4. Example of non-geodesic integral curve γu\gamma^u (blue) along a tangent vector field uu (magenta). For reference a geodesic starting at a common point (green). Since γu\gamma^u is non-geodesic the curvature κu\kappa^u projected onto the tangent space is not zero.

∇Ωω=(1,0)SXb(cos⁡(ω)1−μ2μ)sin⁡(ω)1−μ2(t⋅∇bb)=(1,0)SXb(cos⁡(ω)1−μ2μ)−sin⁡(ω)1−μ2(t⋅κb)\begin{aligned} \nabla_{\Omega}\omega= (1,0)S_{X^b}\begin{pmatrix}\frac{\cos(\omega)\sqrt{1-\mu^2}}{\mu}\end{pmatrix}\sin(\omega)\sqrt{1-\mu^2}(t\cdot\nabla_b b) \\ &= (1,0)S_{X^b}\begin{pmatrix}\frac{\cos(\omega)\sqrt{1-\mu^2}}{\mu}\end{pmatrix}-\sin(\omega)\sqrt{1-\mu^2}(t\cdot\kappa^b) \end{aligned}

Alternatively rewriting t⋅∇nb=−b⋅∇ntt\cdot\nabla_n b=-b\cdot\nabla_n t one can arrive at a formulation in terms of SXtS_{X^t}, again, if the surface XtX^t exists.

∇Ωω=−(1,0)SXt(sin⁡(ω)1−μ2μ)−cos⁡(ω)1−μ2(t⋅∇bb)(25)\nabla_{\Omega}\omega=-(1,0)S_{X^t}\begin{pmatrix}\frac{\sin(\omega)\sqrt{1-\mu^2}}{\mu}\end{pmatrix}-\cos(\omega)\sqrt{1-\mu^2}(t\cdot\nabla_b b) \tag*{(25)}
=−(1,0)SXt(sin⁡(ω)1−μ2μ)+cos⁡(ω)1−μ2(b⋅κt)(26)=-(1,0)S_{X^t}\begin{pmatrix}\frac{\sin(\omega)\sqrt{1-\mu^2}}{\mu}\end{pmatrix}+\cos(\omega)\sqrt{1-\mu^2}(b\cdot\kappa^t) \tag*{(26)}

Unfortunately, terms including the shape operator are not as symmetric as in the first section, so the interpretation becomes less straightforward. For example, while (1,0)SXb(1,0)T(1,0)S_{X^b}(1,0)^{\mathsf{T}} still describes the curvature of XbX^b in the first tangent direction, here in direction tt, the terms proportional to (1,0)SXb(0,1)T(1,0)S_{X^b}(0,1)^{\mathsf{T}} correspond to a twisting of XbX^b in direction of the second tangent vector, when moving along the first tangent vector. An example visualizing such twisting behavior is depicted in figure 5. Depicted is a scenario, where XbX^b is a helicoid an intrinsically flat, but twisting surface, for which diagonal entries of SXbS_{X^b} would be zero, while the off-diagonal entries are not.

Examples

In the previous section we found that (5) can, in one of several ways, be recast as

Ω⋅∇Ψ=Ω⋅∇rΨ+[(1−μ2)C(r,ω)+μ1−μ2(Ω^∥⋅κn)]∂Ψ∂μ+[∂Ω∥∂ω⋅(κt+κb)+μ(t⋅∇nb)]∂Ψ∂ω.(27)\begin{aligned} \Omega\cdot\nabla\Psi&= \Omega\cdot\nabla^r\Psi\\ &\quad+\left[(1-\mu^2)C(r,\omega)+\mu\sqrt{1-\mu^2}\left(\hat{\Omega}_{\parallel}\cdot\kappa^n\right)\right]\frac{\partial\Psi}{\partial\mu} \\ &\quad+\left[\frac{\partial\Omega_{\parallel}}{\partial\omega}\cdot\left(\kappa^t+\kappa^b\right)+\mu(t\cdot\nabla_n b)\right]\frac{\partial\Psi}{\partial\omega}. \tag*{(27)} \end{aligned}
Example of a twisting, but flat manifold (helicoid)

Figure 5. Example of a twisting, but flat manifold (helicoid). Diagonal entries of the shape operator disappear, despite off-diagonal terms being non-zero. The tangent vector tt spins around nn when moving along γ⊥n\gamma^\perp n.

Before looking at specific examples, let us note that, for general curves and manifolds, curvatures can be rather complicated to calculate. However, the curvature approach lends itself nicely to a quick and intuitive assessment of which terms in (28) to expect in the first place. At first glance equation (28) should be considered as

Ω⋅∇Ψ=Ω⋅∇rΨ(28)\boldsymbol{\Omega} \cdot\nabla\Psi= \boldsymbol{\Omega} \cdot\nabla^{r}\Psi \tag*{(28)}
+[(1−μ2)Is Xn flat?+μ1−μ2Is γn curved towards Xn?]∂Ψ∂μ(29)+ \left[(1-\mu^{2})\boxed{\text{Is }X^{n}\text{ flat?}}+\mu\sqrt{1-\mu^{2}}\boxed{\text{Is }\gamma^{n}\text{ curved towards }X^{n}?}\right]\frac{\partial\Psi}{\partial\mu} \tag*{(29)}
+[Are γt and γb curved in Xn?+μAre γt and γb spinning around γn?]∂Ψ∂ω.(30)+\left[\boxed{\text{Are }\gamma^{t}\text{ and }\gamma^{b}\text{ curved in }X^{n}?}+\mu\boxed{\text{Are }\gamma^{t}\text{ and }\gamma^{b}\text{ spinning around }\gamma^{n}?}\right]\frac{\partial\Psi}{\partial\omega}. \tag*{(30)}

If either of the boxed questions in the expression above is answered positively the corresponding term is known to be zero without any further calculation. Which question to ask for each term can be split further or altered based on whether a surface-curvature or a curve-curvature formulation is chosen in the previous section. Either way we believe this to be a quick and intuitive way to decide, if a term needs to be calculated explicitly or not.

We will continue this section with a handful of examples, calculating all non-zero terms, but following this approach to skip the calculation of all terms that will end up being zero anyway. Various choices for the parametrization of Ω\boldsymbol{\Omega} are explored to see how the introduced frameworks can make finding the coordinate form of Ω⋅∇Ψ\boldsymbol{\Omega} \cdot\nabla\Psi, easier and more intuitive. We will start with classical known examples and build up from there to increasingly more complicated scenarios.

Visualization of $\Omega$-basis choice $n$, $t$, $b$ in cylindrical symmetry

Figure 6. Visualization of Ω\Omega-basis choice nn, tt, bb in cylindrical symmetry.

Cylindrical I

Let us begin with a classical example useful in cases with cylindrical symmetry. Assume that rr is expressed as r=(ρcos⁡(ϕ),ρsin⁡(ϕ),z)r = (\rho\cos(\phi), \rho\sin(\phi), z) through the coordinates ρ\rho, ϕ\phi and zz. Common choices in the literature are equivalent to picking n(r)=(0,0,1)n(r) = (0, 0, 1) constant along the zz-axis, t(r)=(cos⁡(ϕ),sin⁡(ϕ),0)t(r) = (\cos(\phi), \sin(\phi), 0) to be the outward normal vector of cylinders of radius ρ\rho and therefore b=(−sin⁡(ϕ),cos⁡(ϕ),0)b = (-\sin(\phi), \cos(\phi), 0) the tangent vector of those cylinders in the x−yx-y-plane. Bell and Glasstone [2] We show this choice in figure 6. Based on the set up we realize that a good choice to express equation (5) here is as

Ω⋅∇Ψ=Ω⋅∇rΨ+[(1−μ2)C(r,ω)+μ1−μ2(Ω^∥⋅κn)]∂Ψ∂μ+[(1,0)SXb(cos⁡(ω)1−μ2μ−sin⁡(ω)1−μ2(t⋅κb))]∂Ψ∂ω.(31)\begin{aligned} \Omega\cdot\nabla\Psi={}& \Omega\cdot\nabla^{r}\Psi\\ &+ \left[(1-\mu^{2})C(r,\omega)+\mu\sqrt{1-\mu^{2}}\left(\hat{\Omega}_{\parallel}\cdot\kappa^{n}\right)\right]\frac{\partial\Psi}{\partial\mu} \\ &+ \left[(1,0)S_{X^{b}}\left(\frac{\cos(\omega)\sqrt{1-\mu^{2}}}{\mu}-\sin(\omega)\sqrt{1-\mu^{2}}\left(t\cdot\kappa^{b}\right)\right)\right]\frac{\partial\Psi}{\partial\omega}. \tag*{(31)} \end{aligned}

which right away leads us towards the informal

Ω⋅∇Ψ=Ω⋅∇rΨ+[(1−μ2)Is Xn flat?+μ1−μ2Is γn curved towards Xn?]∂Ψ∂μ+[(1,0)Is Xb flat?(cos⁡(ω)1−μ2μ−sin⁡(ω)1−μ2Is γb curved towards t?)]∂Ψ∂ω.\begin{aligned} \Omega\cdot\nabla\Psi={}& \Omega\cdot\nabla^{r}\Psi\\ &+ \left[(1-\mu^{2})\boxed{\text{Is }X^{n}\text{ flat?}}+\mu\sqrt{1-\mu^{2}}\boxed{\text{Is }\gamma^{n}\text{ curved towards }X^{n}?}\right]\frac{\partial\Psi}{\partial\mu} \\ &+ \left[(1,0)\boxed{\text{Is }X^{b}\text{ flat?}}\left(\frac{\cos(\omega)\sqrt{1-\mu^{2}}}{\mu}-\sin(\omega)\sqrt{1-\mu^{2}}\boxed{\text{Is }\gamma^{b}\text{ curved towards }t?}\right)\right]\frac{\partial\Psi}{\partial\omega}. \end{aligned}

The answer to the first three questions will be ’Yes’, telling us that the corresponding terms will disappear without further calculation. Here is why. Since nn is constant, it is straightforward to see, that all XnX^{n} are planes and therefor flat. Similarly, γn\gamma^{n} are straight lines in zz-direction. Thus κn=0\kappa^{n}=0 as well, leading to ∇Ωμ=0\nabla_{\Omega}\mu=0 every where. Since bb is independent of both ρ\rho and zz the surface XbX^{b} are just planes as well. The only term remaining is the fourth one, which in this case can also be determined without direct calculations. The curves γb\gamma^{b} generated by bb are great circles around cylinders of radius ρ\rho with normal vector tt. Therefore t⋅κb=1ρt\cdot\kappa^{b}=\frac{1}{\rho} is given by the principal curvature of such cylinders. We therefore concluded ∇Ωω=−sin⁡(ω)1−μ2ρ\nabla_{\Omega}\omega=-\sin(\omega)\frac{\sqrt{1-\mu^{2}}}{\rho} and hence

∇ΩΨ=∇ΩrΨ+0∂Ψ∂μ−1−μ2sin⁡(ω)ρ∂Ψ∂ω.(32)\nabla_{\Omega}\Psi=\nabla_{\Omega}^{r}\Psi+0\frac{\partial\Psi}{\partial\mu}-\frac{\sqrt{1-\mu^{2}}\sin(\omega)}{\rho}\frac{\partial\Psi}{\partial\omega}. \tag*{(32)}

Expressing ∇ΩrΨ\nabla_{\Omega}^{r}\Psi in cylinder coordinates also makes sense in practice, but is not required.

Cylindrical II

While the previous example is a common choice it is not the only way Ω\Omega could be parametrized if the underlying system has cylindrical symmetry. Here is an example, where the roles of nn, tt and bb are switched compared to the previous example. Assume rr again to be given as r=(ρcos⁡(ϕ),ρsin⁡(ϕ),z)r=(\rho\cos(\phi),\rho\sin(\phi),z). Now let nn be the radial vector (cos⁡(ϕ),sin⁡(ϕ),0)(\cos(\phi),\sin(\phi),0), i.e. the normal vector to cylinders of radius ρ\rho. Let t=(−cos⁡(ϕ),sin⁡(ϕ),0)t=(-\cos(\phi),\sin(\phi),0) be the tangent vector to these cylinders orthogonal to the zz-axis and let b=(0,0,1)b=(0,0,1) be the parallel to the zz-axis. Since bb is constant, we can see immediately, that both SXbS_{X^{b}} and κb\kappa^{b} are 0. Therefore ∇Ωω=0\nabla_{\Omega}\omega=0.

The trade-off is that, since nn is now the normal vector on a cylinder everywhere, i.e. XnX^{n} are cylinders, the corresponding shape operators do not disappear anymore. It is in fact given by

SXn=(1ρ000).(33)S_{X^{n}}=\begin{pmatrix}\frac{1}{\rho}&0\\0&0\end{pmatrix}. \tag*{(33)}

Since γn\gamma^{n} are still straight lines and thus κn=0\kappa^{n}=0 we have ∇Ωμ=1−μ2cos⁡2(ω)ρ\nabla_{\Omega}\mu=\frac{\sqrt{1-\mu^{2}}\cos^{2}(\omega)}{\rho} or

∇ΩΨ=∇ΩrΨ+1−μ2cos⁡2(ω)ρ∂Ψ∂μ+0∂Ψ∂ω.(34)\nabla_{\Omega}\Psi=\nabla_{\Omega}^{r}\Psi+\frac{\sqrt{1-\mu^{2}}\cos^{2}(\omega)}{\rho}\frac{\partial\Psi}{\partial\mu}+0\frac{\partial\Psi}{\partial\omega}. \tag*{(34)}

Sphere

We will continue with another classical example: spherical coordinates. Assume r=ρ(cos⁡(ϕ)sin⁡(θ),sin⁡(ϕ)sin⁡(θ),cos⁡(θ))r=\rho(\cos(\phi)\sin(\theta),\sin(\phi)\sin(\theta),\cos(\theta)). The vector n(r)n(r) is commonly chosen to be the normal vector to the sphere of radius ρ\rho at rr, i.e. n=(cos⁡(ϕ)sin⁡(θ),sin⁡(ϕ)sin⁡(θ),cos⁡(θ))n=(\cos(\phi)\sin(\theta),\sin(\phi)\sin(\theta),\cos(\theta)). The tangent vectors tt and bb are picked as t=(cos⁡(ϕ)cos⁡(θ),sin⁡(ϕ)cos⁡(θ),sin⁡(θ))t=(\cos(\phi)\cos(\theta),\sin(\phi)\cos(\theta),\sin(\theta)), which is the tangent vector of the geodesic connecting north and south pole under the angle θ\theta and b=(cos⁡(ϕ),sin⁡(ϕ),0)b=(\cos(\phi),\sin(\phi),0), parallel to curves circling the poles parallel to the equator. The choice of nn, tt and bb is depicted in Figure 7 for clarity.

Visualization of $\Omega$-basis choice $n$, $t$, $b$ in spherical symmetry.

Figure 7. Visualization of Ω\Omega-basis choice nn, tt, bb in spherical symmetry.

We express equation (5) here is as

Ω⋅∇Ψ=Ω⋅∇rΨ+[(1−μ2)C(r,ω)+μ1−μ2(Ω^⋅κn)]∂Ψ∂μ+[1−μ2[cos⁡(ω)(b⋅κt)−sin⁡(ω)(t⋅κb)]+μ(t⋅∇nb)]∂Ψ∂ω,(35)\begin{aligned} \boldsymbol{\Omega} \cdot\nabla\Psi&= \boldsymbol{\Omega} \cdot\nabla^{r}\Psi\\ &\quad+ \left[(1-\mu^{2})C(r,\omega)+\mu\sqrt{1-\mu^{2}}\left(\hat{\boldsymbol{\Omega}}\cdot\boldsymbol{\kappa}^{n}\right)\right]\frac{\partial\Psi}{\partial\mu} \\ &\quad+ \left[\sqrt{1-\mu^{2}}\left[\cos(\omega)\left(\boldsymbol{b}\cdot\boldsymbol{\kappa}^{t}\right)-\sin(\omega)\left(\boldsymbol{t}\cdot\boldsymbol{\kappa}^{b}\right)\right]+\mu\left(\boldsymbol{t}\cdot\nabla_{n}b\right)\right]\frac{\partial\Psi}{\partial\omega}, \tag*{(35)} \end{aligned}

which gives the informal version

Ω⋅∇Ψ=Ω⋅∇rΨ+[(1−μ2)Is Xn flat?+μ1−μ2Is γn curved towards Xn?]∂Ψ∂μ+[1−μ2[cos⁡(ω)Is γt a geodesic of Xn?−sin⁡(ω)Is γb a geodesic of Xn?]+μDoes b not spin around n of Xn?]∂Ψ∂ω.\begin{aligned} \boldsymbol{\Omega}\cdot\nabla\Psi&= \boldsymbol{\Omega}\cdot\nabla^{r}\Psi\\ &\quad+ \left[(1-\mu^{2})\boxed{\text{Is }X^{n}\text{ flat?}}+\mu\sqrt{1-\mu^{2}}\boxed{\text{Is }\gamma^{n}\text{ curved towards }X^{n}?}\right]\frac{\partial\Psi}{\partial\mu} \\ &\quad+ \left[\sqrt{1-\mu^{2}}\left[\cos(\omega)\boxed{\text{Is }\gamma^{t}\text{ a geodesic of }X^{n}?}-\sin(\omega)\boxed{\text{Is }\gamma^{b}\text{ a geodesic of }X^{n}?}\right]+\mu\boxed{\text{Does }b\text{ not spin around }n\text{ of }X^{n}?}\right]\frac{\partial\Psi}{\partial\omega}. \end{aligned}

Since XnX^{n} are spheres here the first answer to the first question is negative2. However, nn is aligned with the radial vector everywhere, and hence γn\gamma^{n} is a straight line meaning the second term does indeed disappear. Next note that γt\gamma^{t} are in fact geodesics on the spheres, while bb is constant along γn\gamma^{n}. From this we conclude that all terms in ∇Ωμ\nabla_{\Omega}\mu except (1−μ2)C(ω)(1-\mu^{2})C(\omega) are zero as well as all terms in ∇Ωω\nabla_{\Omega}\omega except 1−μ2sin⁡(ω)(t⋅κb)\sqrt{1-\mu^{2}}\sin(\omega)(\boldsymbol{t}\cdot\boldsymbol{\kappa}^{b}).

Due to the symmetry of the sphere C(ω)=1ρC(\omega)=\frac{1}{\rho}, however, is independent of ω\omega and is just the curvature of the sphere.

Considering the remaining term t⋅κb\mathbf{t}\cdot\boldsymbol{\kappa}^{b}, note that γb\boldsymbol{\gamma}^{b} at r\mathbf{r} are circles of radius sin⁡(θ)ρ\sin(\theta)\rho with κb\boldsymbol{\kappa}^{b} pointing radially to the zz-axis. Thus ∥κb∥=1sin⁡(θ)ρ\lVert\boldsymbol{\kappa}^{b}\rVert=\frac{1}{\sin(\theta)\rho} and therefore κb⋅t=∥κb∥∥t∥cos⁡(θ)=cot⁡(θ)ρ\boldsymbol{\kappa}^{b}\cdot\mathbf{t}=\lVert\boldsymbol{\kappa}^{b}\rVert\lVert\mathbf{t}\rVert\cos(\theta)=\frac{\cot(\theta)}{\rho}. Piecing everything together

∇ΩΨ=∇ΩrΨ+1−μ2ρ∂Ψ∂μ−1−μ2sin⁡(ω)cot⁡(θ)ρ∂f∂ω(36)\nabla_{\Omega}\Psi=\nabla_{\Omega}^{r}\Psi+\frac{1-\mu^{2}}{\rho}\frac{\partial\Psi}{\partial\mu}-\sqrt{1-\mu^{2}}\sin(\omega)\frac{\cot(\theta)}{\rho}\frac{\partial f}{\partial\omega} \tag*{(36)}

Ellipse

We will continue with an example that, on the one hand, is a natural next step from spherical coordinates, but on the other hand, will showcase a few practical issues that can arise as well as how quickly complexity can increase: Elliptical coordinates. For this example we will express equation (5) as

Ω⋅∇Ψ=Ω⋅∇rΨ+[(1−μ2)[cos⁡(ω)2n⋅κt+sin⁡(ω)2n⋅κb]+μ1−μ2(Ω^∥⋅κn)]∂Ψ∂μ+[1−μ2[cos⁡(ω)(b⋅κt)−sin⁡(ω)(t⋅κb)]+μ(t⋅∇nb)]∂Ψ∂ω.(37)\begin{aligned} \boldsymbol{\Omega}\cdot\nabla\Psi={}&\boldsymbol{\Omega}\cdot\nabla^{r}\Psi\\ &+\left[(1-\mu^{2})\left[\cos(\omega)^{2}\mathbf{n}\cdot\boldsymbol{\kappa}^{t}+\sin(\omega)^{2}\mathbf{n}\cdot\boldsymbol{\kappa}^{b}\right]+\mu\sqrt{1-\mu^{2}}\left(\widehat{\boldsymbol{\Omega}}_{\parallel}\cdot\boldsymbol{\kappa}^{n}\right)\right]\frac{\partial\Psi}{\partial\mu}\\ &+\left[\sqrt{1-\mu^{2}}\left[\cos(\omega)(\mathbf{b}\cdot\boldsymbol{\kappa}^{t})-\sin(\omega)(\mathbf{t}\cdot\boldsymbol{\kappa}^{b})\right]+\mu(\mathbf{t}\cdot\nabla_{\mathbf{n}}\mathbf{b})\right]\frac{\partial\Psi}{\partial\omega}. \tag*{(37)} \end{aligned}

We are about to see that even in this still rather simple case none of these terms will be zero for practical choices of n\mathbf{n}, t\mathbf{t} and b\mathbf{b}.

Assume R3\mathbb{R}^{3} is parametrized by ρ,ϕ,θ\rho,\phi,\theta again, such that r=ρ(acos⁡(ϕ)sin⁡(θ),bsin⁡(ϕ)sin⁡(θ),ccos⁡(θ))\mathbf{r}=\rho(a\cos(\phi)\sin(\theta),b\sin(\phi)\sin(\theta),c\cos(\theta)), where a,b,c>0a,b,c>0. A natural choice, and one that would get rid of ∂Ψ/∂ω\partial\Psi/\partial\omega if ellipses are indeed level sets, would be to pick n\mathbf{n} as the normal vector to these ellipses. Thus

n∝(cos⁡(ϕ)sin⁡(θ)a,sin⁡(ϕ)sin⁡(θ)b,cos⁡(θ)c)T.(38)\mathbf{n}\propto\left(\frac{\cos(\phi)\sin(\theta)}{a},\frac{\sin(\phi)\sin(\theta)}{b},\frac{\cos(\theta)}{c}\right)^{\mathsf{T}}. \tag*{(38)}

Easy choices for t\mathbf{t} and b\mathbf{b} would be

t∝∂X∂θ=(acos⁡(ϕ)cos⁡(θ),bcos⁡(θ)sin⁡(ϕ),csin⁡(θ))T.(39)\mathbf{t}\propto\frac{\partial\mathbf{X}}{\partial\theta}=(a\cos(\phi)\cos(\theta),b\cos(\theta)\sin(\phi),c\sin(\theta))^{\mathsf{T}}. \tag*{(39)}

and

b~∝∂X∂ϕ=(−asin⁡(ϕ)sin⁡(θ),bcos⁡(ϕ)sin⁡(θ),0)T.(40)\widetilde{\mathbf{b}}\propto\frac{\partial\mathbf{X}}{\partial\phi}=(-a\sin(\phi)\sin(\theta),b\cos(\phi)\sin(\theta),0)^{\mathsf{T}}. \tag*{(40)}

We will assume all three of these are normalized, even though we refrained from providing the normalization constants on the right-hand sides to simplify the presentation. This is essentially the same choice we made in the previous section on the sphere. In fact, if a=b=c=1a=b=c=1 all vectors are identical. Note however that on the ellipsoid t\mathbf{t} and b~\tilde{\mathbf b} are not always orthogonal unless a=ba=b:

t⋅b~=[(b2−a2)cos⁡(ϕ)cos⁡(θ)sin⁡(ϕ)sin⁡(θ)][sin⁡(θ)2(a2sin⁡(ϕ)2+b2cos⁡(ϕ)2)(a2cos⁡(ϕ)2cos⁡(θ)2+b2cos⁡(θ)2sin⁡(ϕ)2+c2sin⁡(θ)2)]−1/2(41)\mathbf t \cdot\tilde{\mathbf b} = \left[(b^{2}-a^{2})\cos(\phi)\cos(\theta)\sin(\phi)\sin(\theta)\right]\left[\sin(\theta)^{2}\left(a^{2}\sin(\phi)^{2}+b^{2}\cos(\phi)^{2}\right)\left(a^{2}\cos(\phi)^{2}\cos(\theta)^{2}+b^{2}\cos(\theta)^{2}\sin(\phi)^{2}+c^{2}\sin(\theta)^{2}\right)\right]^{-1/2} \tag*{(41)}

The vector b=n×t\mathbf b=\mathbf n\times\mathbf t can be calculated as

b=b~−(t⋅b~)t1−t⋅b~∝(−(b2+c2)cos⁡(θ)sin⁡(ϕ)sin⁡(θ)bc(a2+c2)cos⁡(ϕ)cos⁡(θ)sin⁡(θ)ac−(a2−b2)cos⁡(ϕ)cos⁡(θ)2sin⁡(ϕ)ab).(42)\mathbf b=\frac{\tilde{\mathbf b}-(\mathbf t\cdot\tilde{\mathbf b})\mathbf t}{1-\mathbf t\cdot\tilde{\mathbf b}}\propto\begin{pmatrix} -\dfrac{(b^{2}+c^{2})\cos(\theta)\sin(\phi)\sin(\theta)}{bc}\\ \dfrac{(a^{2}+c^{2})\cos(\phi)\cos(\theta)\sin(\theta)}{ac}\\ -\dfrac{(a^{2}-b^{2})\cos(\phi)\cos(\theta)^{2}\sin(\phi)}{ab} \end{pmatrix}. \tag*{(42)}

Note further that on an ellipsoid neither γt\gamma^{\mathbf t} nor γb~\gamma^{\tilde{\mathbf b}} are geodesics, forcing us to calculate the terms t⋅κb\mathbf t\cdot\boldsymbol{\kappa}^{\mathbf b} and b⋅κt\mathbf b\cdot\boldsymbol{\kappa}^{\mathbf t} or the corresponding shape operators. Although b\mathbf b can be used directly, many quantities, specifically γu\gamma^{\mathbf u} and by extension κu\boldsymbol{\kappa}^{\mathbf u}, are simpler to calculate for u=b~\mathbf u=\tilde{\mathbf b} than for u=b\mathbf u=\mathbf b. Conveniently the map u↦κu\mathbf u\mapsto\boldsymbol{\kappa}^{\mathbf u} is a linear mapping, as long as r\mathbf r is fixed, allowing us to work with b~\tilde{\mathbf b}, where useful, and then obtain κb\boldsymbol{\kappa}^{\mathbf b} via equation (42).

κb=κb~−(t⋅κt)b~1−t⋅b~.(43)\boldsymbol{\kappa}^{\mathbf b}=\frac{\boldsymbol{\kappa}^{\tilde{\mathbf b}}-(\mathbf t\cdot\boldsymbol{\kappa}^{\mathbf t})\tilde{\mathbf b}}{1-\mathbf t\cdot\tilde{\mathbf b}}. \tag*{(43)}

The shape operator can be transformed in a similar fashion, if the surface curvature form is preferred. The curvature C(ω)C(\omega) does not depend on the choice of basis of the tangent space. While it therefore does not depend on b\mathbf b or b~\tilde{\mathbf b} it still implicitly depends on t\mathbf t through ω\omega. Because of this, it can be beneficial to use more convenient choices, such as b~\tilde{\mathbf b}, even if not all conditions are met, if appropriate adjustments are made. We will demonstrate this in the current section.

In order to calculate κn\boldsymbol{\kappa}^{\mathbf n}, κt\boldsymbol{\kappa}^{\mathbf t} and κb~\boldsymbol{\kappa}^{\tilde{\mathbf b}} we need the curves γn\gamma^{\mathbf n}, γt\gamma^{\mathbf t} and γb~\gamma^{\tilde{\mathbf b}}. Here is where choosing b~\tilde{\mathbf b} over n×t\mathbf n\times\mathbf t becomes beneficial for now as parametrizations of γt\gamma^{\mathbf t} and γb~\gamma^{\tilde{\mathbf b}} are given through r(ρ,ϕ,θ)\mathbf r(\rho,\phi,\theta) by fixing ρ\rho and one of the angles respectively. Thus γt(τ,r(ρ,ϕ,θ))=r(ρ,ϕ,θ+τ)\gamma^{\mathbf t}(\tau,\mathbf r(\rho,\phi,\theta))=\mathbf r(\rho,\phi,\theta+\tau) and γb~(τ,r(ρ,ϕ,θ))=r(ρ,ϕ+τ,θ)\gamma^{\tilde{\mathbf b}}(\tau,\mathbf r(\rho,\phi,\theta))=\mathbf r(\rho,\phi+\tau,\theta) with τ∈(−ϵ,ϵ)\tau\in(-\epsilon,\epsilon). These are however not arc length parametrizations so either reparametrization is necessary or the formulas for κu\boldsymbol{\kappa}^{\mathbf u} introduced in Remark 1 need adjustment. We correct for parametrization by setting

κu=−(γu)′′∥(γu)′′∥∥(γu)′×(γu)′′∥∥(γu)′∥3,(44)\boldsymbol{\kappa}^{\mathbf u}=-\frac{(\gamma^{\mathbf u})''}{\lVert(\gamma^{\mathbf u})''\rVert}\frac{\lVert(\gamma^{\mathbf u})'\times(\gamma^{\mathbf u})''\rVert}{\lVert(\gamma^{\mathbf u})'\rVert^{3}}, \tag*{(44)}

where the first term preserves the direction of the original definition, while the second term accounts for rescaling to arc-length. Using this we can calculate κt\boldsymbol{\kappa}^{\mathbf t} and κb~\boldsymbol{\kappa}^{\tilde{\mathbf b}}:

κt=1kt(asin⁡(θ)cos⁡(ϕ)c2((a2−b2)cos⁡(2ϕ)+a2+b2)bsin⁡(θ)sin⁡(ϕ)c2((a2−b2)cos⁡(2ϕ)+a2+b2)ccos⁡(θ)c2((a2−b2)cos⁡(2ϕ)+a2+b2))(45)\boldsymbol{\kappa}^{\mathbf t}=\frac{1}{k_{\mathbf t}}\begin{pmatrix} a\sin(\theta)\cos(\phi)\sqrt{c^{2}\left((a^{2}-b^{2})\cos(2\phi)+a^{2}+b^{2}\right)}\\ b\sin(\theta)\sin(\phi)\sqrt{c^{2}\left((a^{2}-b^{2})\cos(2\phi)+a^{2}+b^{2}\right)}\\ c\cos(\theta)\sqrt{c^{2}\left((a^{2}-b^{2})\cos(2\phi)+a^{2}+b^{2}\right)} \end{pmatrix} \tag*{(45)}

with

kt=k_t =
2ρ(cos⁡2(θ)(a2cos⁡2(ϕ)+b2sin⁡2(ϕ))+c2sin⁡2(θ))3/2sin⁡2(θ)(a2cos⁡2(ϕ)+b2sin⁡2(ϕ))+c2cos⁡2(θ)\sqrt{2}\rho\left(\cos^2(\theta)\left(a^2\cos^2(\phi)+b^2\sin^2(\phi)\right)+c^2\sin^2(\theta)\right)^{3/2}\sqrt{\sin^2(\theta)\left(a^2\cos^2(\phi)+b^2\sin^2(\phi)\right)+c^2\cos^2(\theta)}

and

κb=1kb(asin⁡(θ)cos⁡(ϕ)a2b2sin⁡4(θ)bsin⁡(θ)sin⁡(ϕ)a2b2sin⁡4(θ)0),(46)\boldsymbol{\kappa}^{b}=\frac{1}{k_b}\begin{pmatrix}a\sin(\theta)\cos(\phi)\sqrt{a^2b^2\sin^4(\theta)}\\b\sin(\theta)\sin(\phi)\sqrt{a^2b^2\sin^4(\theta)}\\0\end{pmatrix}, \tag*{(46)}

where

kb=ρ(sin⁡2(θ)(a2sin⁡2(ϕ)+b2cos⁡2(ϕ)))3/2sin⁡2(θ)(a2cos⁡2(ϕ)+b2sin⁡2(ϕ))k_b=\rho\left(\sin^2(\theta)\left(a^2\sin^2(\phi)+b^2\cos^2(\phi)\right)\right)^{3/2}\sqrt{\sin^2(\theta)\left(a^2\cos^2(\phi)+b^2\sin^2(\phi)\right)}

Now κb\boldsymbol{\kappa}^{b} can be calculated through (43), as can all inner products of κt\boldsymbol{\kappa}^{t} and κb\boldsymbol{\kappa}^{b} with n,tn,t and bb required for the calculation ∇ΩΨ\nabla\Omega\Psi.

Finding a parametrization of γn\boldsymbol{\gamma}^{n} to calculate κn\boldsymbol{\kappa}^{n} is less straightforward. Note in particular that fixing the two angles and varying ρ\rho does not give such a parametrization. A direct calculation of κn=−∇nn\boldsymbol{\kappa}^{n}=-\nabla_n n can be simpler and is what we did to obtain

κn=1kn(acos⁡(ϕ)sin⁡(θ)(b4(c2−a2)cos⁡(θ)2+c4(b2−a2)sin⁡(ϕ)2sin⁡(θ)2)bsin⁡(ϕ)sin⁡(θ)(a4(c2−b2)cos⁡(θ)2+c4(a2−b2)cos⁡(ϕ)2sin⁡(θ)2)csin⁡(θ)2cos⁡(θ)(a4(b2−c2)sin⁡(ϕ)2−b4(c2−a2)cos⁡(ϕ)2)),(47)\boldsymbol{\kappa}^{n}=\frac{1}{k_n}\begin{pmatrix} a\cos(\phi)\sin(\theta)\left(b^4(c^2-a^2)\cos(\theta)^2+c^4(b^2-a^2)\sin(\phi)^2\sin(\theta)^2\right)\\ b\sin(\phi)\sin(\theta)\left(a^4(c^2-b^2)\cos(\theta)^2+c^4(a^2-b^2)\cos(\phi)^2\sin(\theta)^2\right)\\ c\sin(\theta)^2\cos(\theta)\left(a^4(b^2-c^2)\sin(\phi)^2-b^4(c^2-a^2)\cos(\phi)^2\right) \end{pmatrix}, \tag*{(47)}

with

kn=ρ(c2sin⁡(θ)2(a2sin⁡(ϕ)+b2cos⁡(ϕ)2)+a2b2cos⁡(θ)2)2.k_n=\rho\left(c^2\sin(\theta)^2\left(a^2\sin(\phi)+b^2\cos(\phi)^2\right)+a^2b^2\cos(\theta)^2\right)^2.

Calculating the inner product with Ω∥\boldsymbol{\Omega}_{\parallel} is thus simply a, somewhat tedious, exercise in arithmetic. Similarly, i.e. through direct calculation, we find the remaining term

b⋅∇nt=−t⋅∇nb=abc(a2−b2)sin⁡(ϕ)cos⁡(ϕ)cos⁡(θ)[ρ(a2sin⁡(ϕ)2+b2cos⁡2(ϕ))][a2b2cos⁡(θ)2+c2sin⁡(θ)2(a2sin⁡2(ϕ)+b2cos⁡(ϕ)2)](48)b\cdot\nabla_n t=-t\cdot\nabla_n b= \frac{abc\left(a^2-b^2\right)\sin(\phi)\cos(\phi)\cos(\theta)} {\left[\rho\left(a^2\sin(\phi)^2+b^2\cos^2(\phi)\right)\right]\left[a^2b^2\cos(\theta)^2+c^2\sin(\theta)^2\left(a^2\sin^2(\phi)+b^2\cos(\phi)^2\right)\right]} \tag*{(48)}

All that is left is to plug these terms back into the formulas for ∇ΩΨ\nabla_{\Omega}\Psi, which we will refrain from due to the length of the terms in this example. We saw that going from spherical to elliptical coordinates significantly increased the complexity of the involved terms as well as the procedure of calculating them. We also highlighted, that convenient choices might not always fulfill the conditions used in the derivations in the previous sections, but demonstrated how they might still work.

Remark 5. Homothetic Scaling A shared feature of all the preceding examples is the appearance of the scaling factor 1/ρ1/\rho in every term. This is not a coincidence, but rather a direct consequence of the fact that in each case, the rescaled embedding satisfies Xρrn=ρXrn\mathbf{X}^{n}_{\rho r}=\rho\mathbf{X}^{n}_{r}. In other words, the foliation induced by n\mathbf{n} can be generated by scaling a single base manifold by a factor of ρ\rho. Such families of manifolds are said to be generated by homothetic/self-similar rescaling. These commonly arise in coordinate systems and in the context of curvature flows. Kohn [10]; Drugan et al. (2018)

More specifically, for r1,r2∈R3\mathbf{r}_1,\mathbf{r}_2 \in\mathbb{R}^{3} with r2=ρr1\mathbf{r}_2=\rho\mathbf{r}_1 for some ρ>0\rho>0, we observe that the normal, tangent, and binormal vectors satisfy n(r1)=n(r2)\mathbf{n}(\mathbf{r}_1)=\mathbf{n}(\mathbf{r}_2), t(r1)=t(r2)\mathbf{t}(\mathbf{r}_1)=\mathbf{t}(\mathbf{r}_2), and b(r1)=b(r2)\mathbf{b}(\mathbf{r}_1)=\mathbf{b}(\mathbf{r}_2). Consequently, the associated quantities Xu\mathbf{X}^{u} and γu\boldsymbol{\gamma}^{u} are simply rescaled by a factor of ρ\rho when transitioning from r1\mathbf{r}_1 to r2\mathbf{r}_2.

An illustrative 2D example of such homothetic rescaling is shown in Figure 8.

To see the effect on curvatures let X∈R3\mathbf{X}\in\mathbb{R}^{3} be an embedded manifold parametrized by u1,u2∈U2⊂R2u^{1},u^{2}\in U^{2}\subset\mathbb{R}^{2}, if X\mathbf{X} is a surface, and u1∈U1⊂Ru^{1}\in U^{1}\subset\mathbb{R}, if X\mathbf{X} is a curve. Next, let Y=ρX\mathbf{Y}=\rho\mathbf{X}. If Xi=∂X∂ui\mathbf{X}_{i}=\frac{\partial\mathbf{X}}{\partial u^{i}} with i=1(,2)i=1(,2) are picked as tangent vectors to X\mathbf{X}, then If Yi=∂Y∂ui=ρXi\mathbf{Y}_{i}=\frac{\partial\mathbf{Y}}{\partial u^{i}}=\rho\mathbf{X}_{i} are tangent vectors to Y\mathbf{Y}. Since Xi\mathbf{X}_{i} and Yi\mathbf{Y}_{i} are colinear, the normal vectors NX=NY\mathbf{N}_{X}=\mathbf{N}_{Y} are the same for fixed u1(,u2)u^{1}(,u^{2}). Consequently the metric tensors of Y\mathbf{Y} gYα,β=Yα⋅Yβ=ρ2Xα⋅Xβ=ρ2gXα,βg_{Y_{\alpha,\beta}}=\mathbf{Y}_{\alpha}\cdot\mathbf{Y}_{\beta}=\rho^{2}\mathbf{X}_{\alpha}\cdot\mathbf{X}_{\beta}=\rho^{2}g_{X_{\alpha,\beta}} is a scalar multiple of the metric tensor gXg_{X} of X\mathbf{X}. Similarly for the second fundamental form:

hYα,β=NY⋅∂2Y∂α∂β=ρNX⋅∂2X∂α∂β=ρhXα,β.h_{Y_{\alpha,\beta}}=\mathbf{N}_{Y}\cdot\frac{\partial^{2}\mathbf{Y}}{\partial\alpha\partial\beta}=\rho\mathbf{N}_{X}\cdot\frac{\partial^{2}\mathbf{X}}{\partial\alpha\partial\beta}=\rho h_{X_{\alpha,\beta}}.

This now finally leads to

SY=(gY)−1hY=1ρ2(gX)−1ρhX=1ρSX.(49)S_{Y}=(g_{Y})^{-1}h_{Y}=\frac{1}{\rho^{2}}(g_{X})^{-1}\rho h_{X}=\frac{1}{\rho}S_{X}. \tag*{(49)}
2D example of homothetically rescaled manifolds

Figure 8. 2D example of homothetically rescaled manifolds. Base manifold X\mathbf{X} in red, rescaled copies ρX\rho\mathbf{X} in blue for various ρ\rho.

This relation between shape factors of scaled manifolds, intuitively implies that manifolds scaled by a factor ρ\rho are flatter by a factor 1/ρ1/\rho. This applies to manifolds in arbitrary dimensions, but specifically to the curves and surfaces of interest for us here thus explaining the factor 1/ρ1/\rho in all terms from previous examples bar b⋅∇nt\mathbf{b} \cdot\nabla_n t. To explain the scaling of this factor consider γu(r1,τ)\gamma^u(r_1,\tau) and assume this is a parametrization by arc length. Then γu(r2,τ)=ργ(r1,τ/ρ)\gamma^u(r_2,\tau)=\rho\gamma(r_1,\tau/\rho) is also an arc length parametrization. Noting that this implies κu(r1)=1ρκu(r2)\kappa^u(r_1)=\frac{1}{\rho}\kappa^u(r_2) gives an additional explanation of the factor 1/ρ1/\rho in all curvature related terms. However, we can also use this to see that

∇nt(r2)=ddτt(γn(r2,τ))∣τ=0=ddτt(ργn(r1,τ/ρ))∣τ=0=ddτt(γn(r1,τ/ρ))∣τ=0=1ρ∇nt(r1)(50)\begin{aligned} \nabla_n t(r_2)=\left.\frac{d}{d\tau}t\left(\gamma^n(r_2,\tau)\right)\right|_{\tau=0}=\left.\frac{d}{d\tau}t\left(\rho\gamma^n(r_1,\tau/\rho)\right)\right|_{\tau=0} \\ &=\left.\frac{d}{d\tau}t\left(\gamma^n(r_1,\tau/\rho)\right)\right|_{\tau=0}=\frac{1}{\rho}\nabla_n t(r_1) \tag*{(50)} \end{aligned}

thereby explaining the 1/ρ1/\rho factor in the remaining term. Here we used γu(r2,τ)=ργ(r1,τ/ρ)\gamma^u(r_2,\tau)=\rho\gamma(r_1,\tau/\rho) in the first line and t(r)=t(ρr)t(r)=t(\rho r) going from the first to the second line.

While the example of elliptical coordinates gave us a glimpse into how complex things can get one might find some relief in the fact that it suffices to handle the case ρ=1\rho=1 and simply scale everything in the end, when dealing with coordinates based on homothetic foliations.

Translating Graph coordinates

We will close the example section with a class of coordinates that is not based on homothetic foliation, but translated graphs instead. For this let f∈C(R2,R)f\in C(\mathbb{R}^2,\mathbb{R}) be twice differentiable function with graph G0(x,y)=(x,y,f(x,y))G_0(x,y)=(x,y,f(x,y)) and assume that R3\mathbb{R}^3 is parametrized along translated copies of GG:

r(x,y,z):=Gz(x,y):=(x,y,f(x,y)+z).(51)r(x,y,z):=G_z(x,y):=(x,y,f(x,y)+z). \tag*{(51)}

An example is depicted in Figure 9. We will consider the special case of a translating paraboloid first and then provide formulas for the case of general GzG_z. For these two examples we will use the form of Ω⋅∇Ψ\Omega\cdot\nabla\Psi given at the start of the example section in (28). As was the case for ellipses unfortunately none of the terms will end up being zero. However we will see that all terms in the streaming operator can be calculated with knowledge of G0G_0 alone, i.e. independently of zz.

2D example of translating graphs

Figure 9. 2D example of translating graphs. Base graph G0G_0 in red and several translated copies Gz=G0+zG_z = G_0 + z in blue.

Translating Paraboloid

In the case of G0G_0 being a paraboloid ff is of the form f(x,y)=ax2+by2f(x,y)=ax^2+by^2 (a,b∈Ra,b\in\mathbb{R}). Tangent vectors parallel to the coordinate axis to GzG_z can by found by normalizing ∂xGz\partial_xG_z and ∂yGz\partial_yG_z respectively. This intuitive choice leads to

t=(14a2x2+102ax4a2x2+1)b~=(014b2y2+12by4b2y2+1)(52)\mathbf{t}=\begin{pmatrix} \frac{1}{\sqrt{4a^2x^2+1}}\\ 0\\ \frac{2ax}{\sqrt{4a^2x^2+1}} \end{pmatrix} \qquad \tilde{\mathbf{b}}=\begin{pmatrix} 0\\ \frac{1}{\sqrt{4b^2y^2+1}}\\ \frac{2by}{\sqrt{4b^2y^2+1}} \end{pmatrix} \tag*{(52)}

along with the normal vector

n(x,y,z)=(−2ax4a2x2+14b2y2+1−2by4a2x2+14b2y2+114a2x2+14b2y2+1)(53)\mathbf{n}(x,y,z)=\begin{pmatrix} -\dfrac{2ax}{\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}}\\ -\dfrac{2by}{\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}}\\ \dfrac{1}{\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}} \end{pmatrix} \tag*{(53)}

However, t\mathbf{t} and b~\tilde{\mathbf{b}} are generally not orthogonal. This can be corrected via Gram-Schmidt, keeping t\mathbf{t} and replacing b~\tilde{\mathbf{b}} with:

b~=(−4abxy4a2x2+1(4a2x2+14b2y2+1−4abxy)4a2x2+14a2x2+14b2y2+1−4abxy2by4a2x2+1(4a2x2+14b2y2+1−4abxy))(54)\tilde{\mathbf{b}}=\begin{pmatrix} -\dfrac{4abxy}{\sqrt{4a^2x^2+1}\left(\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}-4abxy\right)}\\ \dfrac{\sqrt{4a^2x^2+1}}{\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}-4abxy}\\ \dfrac{2by}{\sqrt{4a^2x^2+1}\left(\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}-4abxy\right)} \end{pmatrix} \tag*{(54)}

Using b~\tilde{\mathbf{b}} over b\mathbf{b} has the advantage that κb~=∂yb~\kappa^{\tilde b}=\partial_y\tilde{\mathbf{b}}, while κb\kappa^b is a slightly more complicated directional derivative of b\mathbf{b}. Similarly calculating the second fundamental form needed for the shape operator is slightly easier obtained with respect to {t,b}\{\mathbf{t},\mathbf{b}\}. The price to pay for this simplification is that terms calculated using b~\tilde{\mathbf{b}} will require correction via equations (42) and (43).

Either route leads to the shape operator with respect to t\mathbf{t} and b\mathbf{b} being

(SGz)t=−2a(4a2x2+1)4b2y2+1(55)\left(SG_z\right)_{t}=-\frac{2a}{\left(4a^2x^2+1\right)\sqrt{4b^2y^2+1}} \tag*{(55)}
(SGz)tb=8a2bxy(4a2x2+14b2y2+1+4abxy)(4a2x2+1)4b2y2+1(4a2x2+4b2y2+1)(56)\left(SG_z\right)_{tb}=\frac{8a^2bxy\left(\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}+4abxy\right)}{\left(4a^2x^2+1\right)\sqrt{4b^2y^2+1}\left(4a^2x^2+4b^2y^2+1\right)} \tag*{(56)}
(SGz)bt=8a2bxy(4a2x2+14b2y2+1−4abxy)(4a2x2+1)4b2y2+1(4a2x2+4b2y2+1)(57)\left(SG_z\right)_{bt}=\frac{8a^2bxy\left(\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}-4abxy\right)}{\left(4a^2x^2+1\right)\sqrt{4b^2y^2+1}\left(4a^2x^2+4b^2y^2+1\right)} \tag*{(57)}
(SGz)bb=−2b(16a3bx2y24b2y2+1+4a2x24a2x2+1+4a2x2+1)(4a2x2+1)(4a2x2+4b2y2+1)(58)\left(SG_z\right)_{bb}=-\frac{2b\left(\frac{16a^3bx^2y^2}{\sqrt{4b^2y^2+1}}+4a^2x^2\sqrt{4a^2x^2+1}+\sqrt{4a^2x^2+1}\right)}{\left(4a^2x^2+1\right)\left(4a^2x^2+4b^2y^2+1\right)} \tag*{(58)}

The curvatures κn,κt\boldsymbol{\kappa}^{n},\boldsymbol{\kappa}^{t} and κb\boldsymbol{\kappa}^{b} are equally lengthy so that for the sake of readability we will only provide their inner products with Ω∥,b\boldsymbol{\Omega}_{\parallel},b and tt as needed.

Ω∥⋅κn=cos⁡(ω)4a2x(4a2x2+1)3/2(4b2y2+1)+sin⁡(ω)4by(−4a3x2+4a2bx2+b)(4a2x2+1)3/2(4b2y2+1)(4a2x2+14b2y2+1−4abxy)(59)\boldsymbol{\Omega}_{\parallel}\cdot\boldsymbol{\kappa}^{n}=\cos(\omega)\frac{4a^2x}{\left(4a^2x^2+1\right)^{3/2}\left(4b^2y^2+1\right)} +\sin(\omega)\frac{4by\left(-4a^3x^2+4a^2bx^2+b\right)}{\left(4a^2x^2+1\right)^{3/2}\left(4b^2y^2+1\right)\left(\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}-4abxy\right)} \tag*{(59)}

Conveniently

κt⋅b=0(60)\boldsymbol{\kappa}^{t}\cdot b=0 \tag*{(60)}

and

κb⋅t=4aby(4a2x2+1)3/2(4a2x2+14b2y2+1−4abxy)(61)\boldsymbol{\kappa}^{b}\cdot t=\frac{4aby}{\left(4a^2x^2+1\right)^{3/2}\left(\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}-4abxy\right)} \tag*{(61)}

Lastly, we can find rather immediately

b⋅∇tn=−t⋅∇bn=8a2bxy(4a2x2+1)3/24b2y2+1(4abxy−4a2x2+14b2y2+1)(62)b\cdot\nabla_t n=-t\cdot\nabla_b n= \frac{8a^2bxy}{\left(4a^2x^2+1\right)^{3/2}\sqrt{4b^2y^2+1}\left(4abxy-\sqrt{4a^2x^2+1}\sqrt{4b^2y^2+1}\right)} \tag*{(62)}

General translating Graph

As final example we shall generalize the previous example by replacing the paraboloid by a general function f∈C2(R2,R)f\in C^2(\mathbb{R}^2,\mathbb{R}). As shorthand let subscripts on ff denote derivatives with respect to the subscript, i.e. for example fx=∂f∂xf_x=\frac{\partial f}{\partial x}. We will again chose nn to be the normal vector of the graph of ff given by

n(x,y,z)=((−fxfy2+1fx2+1−fyfy2+1fx2+11fy2+1fx2+1))(63)n(x,y,z)=\left(\begin{pmatrix} -\dfrac{f_x}{\sqrt{f_y^2+1}\sqrt{f_x^2+1}}\\ -\dfrac{f_y}{\sqrt{f_y^2+1}\sqrt{f_x^2+1}}\\ \dfrac{1}{\sqrt{f_y^2+1}\sqrt{f_x^2+1}} \end{pmatrix}\right) \tag*{(63)}

Like for the paraboloid

t=(1fx2+10fxfx2+1)b~=(01fy2+1fyfy2+1)(64)t=\begin{pmatrix} \dfrac{1}{\sqrt{f_x^2+1}}\\ 0\\ \dfrac{f_x}{\sqrt{f_x^2+1}} \end{pmatrix} \qquad \tilde b=\begin{pmatrix} 0\\ \dfrac{1}{\sqrt{f_y^2+1}}\\ \dfrac{f_y}{\sqrt{f_y^2+1}} \end{pmatrix} \tag*{(64)}

are the canonical tangent vectors to GzG_z parallel to the xx and yy-axis and, also again, they generally will fail to be orthogonal. An appropriate, orthonormal choice is

b(x,y)=1fy2+1fx2+1−fyfx(fyfxfx2+1−fx2+1fyfx2+1)(65)b(x,y)=\frac{1}{\sqrt{f_y^2+1}\sqrt{f_x^2+1}-f_yf_x} \begin{pmatrix} \dfrac{f_yf_x}{\sqrt{f_x^2+1}}\\ -\sqrt{f_x^2+1}\\ \dfrac{f_y}{\sqrt{f_x^2+1}} \end{pmatrix} \tag*{(65)}

In this basis we can find the components of the shape operator by calculating the metric tensor and second fundamental form and get

(SGz)tt=−1kSfxx(66)(S_{G_z})_{tt}=-\frac{1}{k_S}f_{xx} \tag*{(66)}
(SGz)tb=fyfxfxx−fy2+1fx2+1fxykS(fy2+1fx2+1−fyfx)(67)(S_{G_z})_{tb}=\frac{f_yf_xf_{xx}-\sqrt{f_y^2+1}\sqrt{f_x^2+1}f_{xy}}{k_S\left(\sqrt{f_y^2+1}\sqrt{f_x^2+1}-f_yf_x\right)} \tag*{(67)}
(SGz)bt=(fy2+1fx2+1−fyfx)(fyfxx−(fx2+1)fxy)kS(fy2+fx2+1)(68)(S_{G_z})_{bt}=\frac{\left(\sqrt{f_y^2+1}\sqrt{f_x^2+1}-f_yf_x\right)\left(f_yf_{xx}-(f_x^2+1)f_{xy}\right)}{k_S\left(f_y^2+f_x^2+1\right)} \tag*{(68)}
(SGz)bb=fyfxx((fx2+fy2+1fx2+1+1)fxy−fyfxfxx−fy2+1fyy(fx2+1)3/2)kS(fy2+fx2+1)(69)(S_{G_z})_{bb}=\frac{f_yf_{xx}\left(\left(f_x^2+\sqrt{f_y^2+1}\sqrt{f_x^2+1}+1\right)f_{xy}-f_yf_xf_{xx}-\sqrt{f_y^2+1}f_{yy}\left(f_x^2+1\right)^{3/2}\right)}{k_S\left(f_y^2+f_x^2+1\right)} \tag*{(69)}

where

kS=fy2+1(fx2+1).k_S=\sqrt{f_y^2+1}\left(f_x^2+1\right).

Again we will only provide the necessary inner products of κn\boldsymbol{\kappa}^{n}, κt\boldsymbol{\kappa}^{t} and κb\boldsymbol{\kappa}^{b} with Ω∥\boldsymbol{\Omega}_{\parallel}, b\boldsymbol{b} and t\boldsymbol{t}.

Ω∥⋅κn=cos⁡(ω)+fyfxy+fxfxx(fy2+1)(fx2+1)3/2+sin⁡(ω)fyy(fx2+1)−fy2fxfxy+fx(fx2+1)fxy(fy2+1)(fx2+1)3/2(fy2+1fx2+1−fyfx)(70)\boldsymbol{\Omega}_{\parallel}\cdot\boldsymbol{\kappa}^{n} = \cos(\omega) + \frac{f_y f_{xy}+f_x f_{xx}}{(f_y^2+1)(f_x^2+1)^{3/2}} + \sin(\omega) \frac{f_{yy}(f_x^2+1)-f_y^2f_xf_{xy}+f_x(f_x^2+1)f_{xy}} {(f_y^2+1)(f_x^2+1)^{3/2}\left(\sqrt{f_y^2+1}\sqrt{f_x^2+1}-f_yf_x\right)} \tag*{(70)}
κt⋅b=−fy(fx2+1fy+fx2+1+fy2+1(−fy)fx)fxy(fy2+1)(fx2+1)(fyfx−fy2+1fx2+1)2(71)\boldsymbol{\kappa}^{t}\cdot\boldsymbol{b} = -\frac{f_y\left(\sqrt{f_x^2+1}f_y+\sqrt{f_x^2+1}+\sqrt{f_y^2+1}(-f_y)f_x\right)f_{xy}} {(f_y^2+1)(f_x^2+1)\left(f_yf_x-\sqrt{f_y^2+1}\sqrt{f_x^2+1}\right)^2} \tag*{(71)}

and

κb⋅t=fyfxx(fx2+1)3/2(fy2+1fx2+1−fyfx)(72)\boldsymbol{\kappa}^{b}\cdot\boldsymbol{t} = \frac{f_yf_{xx}} {(f_x^2+1)^{3/2}\left(\sqrt{f_y^2+1}\sqrt{f_x^2+1}-f_yf_x\right)} \tag*{(72)}

To wrap up we calculate the winding term as

b⋅∇tn=−t⋅∇bn=−fy(fyfxy+fxfxx)fy2+1(fx2+1)3/2(fy2+1fx2+1−fyfx)(73)\boldsymbol{b}\cdot\nabla_t n = -\boldsymbol{t}\cdot\nabla_b n = -\frac{f_y(f_yf_{xy}+f_xf_{xx})} {\sqrt{f_y^2+1}(f_x^2+1)^{3/2}\left(\sqrt{f_y^2+1}\sqrt{f_x^2+1}-f_yf_x\right)} \tag*{(73)}

It should be noted that, while the general translating graph example expectedly looks more complicated than the paraboloid, the differences, formally, are not too big. Since the curvature terms consist of first and second derivatives only the paraboloid already exposes most of the complexity that is to be expected. Further note that none of the terms depend on zz. Thus, despite none of the terms in these two example being particularly simple they all can be calculated using geometric properties of G0G_0 alone.

Conclusion

We provided a derivation of the streaming operator ∇ΩΨ\nabla_{\Omega}\Psi in equation (1) with a high level of generality and freedom concerning the choice of parametrization of Ω\Omega. We also provided various interpretations of the resulting terms relating to the curvature of manifolds, which we believe to be more intuitive than pure calculus or complicated pictures. While we provided a handful of classical examples where the correctly parametrized form of ∇ΩΨ\nabla_{\Omega}\Psi can be almost intuited, we also demonstrated the complexity can increase tremendously once the parametrization of Ω\Omega becomes even slightly more complicated. Regardless, we believe the connection to curves and surfaces sheds light on the issue at hand from a new angle and might provide intuition and guidance to practitioners when deciding if the trade-off between increased complexity versus a potential reduction in dimension when using different coordinate systems is worth it or not. Beyond that we see use as a intuitive verification tool. While actual curvatures and thus the coefficients of terms in the streaming operator can be hard to calculate, whether or not they vanish can oftentimes be decided without any calculation.

Appendix

When does ∂ωΨ\partial_{\omega}\Psi disappear?

We mentioned in previous sections that if XnX^{n} are picked to be the level sets of Ψ\Psi, i.e. ∇rΨ(r,Ω,t)∝n(r)\nabla^{r}\Psi(r,\Omega,t) \propto n(r) for all rr, Ω\Omega, then the terms ∂Ψ∂ω\frac{\partial\Psi}{\partial\omega} are also zero under rather mild conditions on XnX^{n}. These conditions are the following: Let v∈TrXnv \in T_{r}X^{n} be a tangent vector to XnX^{n} at rr and let Q∈SO(2)Q \in SO(2). We require there exists a loop γ:[0,1]→Xn\gamma:[0,1] \to X^{n} on XnX^{n} starting and ending at rr, such that parallel transport Lee [14] along $\gamma maps vv to QvQv. If this holds we say the Holonomy group of XnX^n at rr equals SO(2)SO(2), or Holr=SO(2)Hol_r = SO(2). If this condition holds then for every ω1\omega_1 and ω2\omega_2, there is a loop γ⊂Xn\gamma\subset X^n such that Ω(μ,ω1)\Omega(\mu,\omega_1) is turned into Ω(μ,ω2)\Omega(\mu,\omega_2) while moving along γ\gamma. But since XnX^n was assumed to be a level set of Ψ\Psi this implies Ψ(r,Ω(μ,ω1),t)=Ψ(r,Ω(μ,ω2,t)\Psi(r,\Omega(\mu,\omega_1),t) = \Psi(r,\Omega(\mu,\omega_2,t) and therefor ∂Ψ∂ω=0\frac{\partial\Psi}{\partial\omega} = 0. While sounding rather complicated the condition Holr(Xn)=SO(2)Hol_r(X^n) = SO(2) holds for all orientable surfaces that are not a product space or have constant intrinsic curvature. Berger [4] While spheres are a notable exception whose holonomy group happens to be SO(2)SO(2) despite having constant curvature, cylinders unfortunately are not.

Some comments on Conservation forms

In numerical applications people are often more interested in spatial and directional averages of Ψ\Psi and as such of ∇ΩΨ\nabla_{\Omega}\Psi. It is therefore convenient to rewrite ∇ΩΨ\nabla_{\Omega}\Psi in conservation form Bell and Glasstone [2]

v(r)∇ΩΨ=(∇r∂μ∂ω)⋅J(r,μ,ω),(74)v(r)\nabla_{\Omega}\Psi= \begin{pmatrix} \nabla^r \\ \partial_{\mu} \\ \partial_{\omega} \end{pmatrix} \cdot J(r,\mu,\omega), \tag*{(74)}

where v(r)v(r) is the volume element with respect to the coordinates of rr and J∈R5J \in\mathbb{R}^{5} is some vector field whose components can be interpreted as flux of Ψ\Psi. If such a form exists the integral

∫V∫S2∇ΩΨ=∫V∫S2∇ΩΨv(r) du1du2du3dμdω.(75)\int_V\int_{S^2}\nabla_{\Omega}\Psi = \int_V\int_{S^2}\nabla_{\Omega}\Psi v(r)\,du^1du^2du^3d\mu d\omega. \tag*{(75)}

is simplified tremendously via the divergence theorem.

We show through a simplified case that a reformulation such as (A1) can not be expected in arbitrary coordinates. Suppose that we have picked coordinates r(u1,u2,u3)r(u^1,u^2,u^3) such that drdu1=n=∇Ψ\frac{dr}{du^1}=n=\nabla\Psi, that is XnX^n are level sets of Ψ\Psi. Further, assume that the structure of XnX^n is sufficient for this to imply that ∂Ψ/∂ω=0\partial\Psi/\partial\omega= 0. (Compare section A.1 for details. Also take note, that this example specifically excludes cylindrical coordinates.)

We therefore have

v(r)∇ΩΨ=v(r)μ∇nΨ+v(r)((1−μ2)C(r,ω)+μ1−μ2(Ω∥⋅κn))∂Ψ∂μ(76)v(r)\nabla_{\Omega}\Psi = v(r)\mu\nabla_n\Psi + v(r)\left((1-\mu^2)C(r,\omega)+\mu\sqrt{1-\mu^2}\left(\Omega_{\parallel}\cdot\kappa^n\right)\right)\frac{\partial\Psi}{\partial\mu} \tag*{(76)}

as shown in section 2. To arrive at a conservation form of this we attempt the following Ansatz:

v(r)∇ΩΨ=v(r)μf∇n(fΨ)v(r)((1−μ2)C(r,ω)+μ1−μ2(Ω∥⋅κn))g∂(gΨ)∂μ(77)v(r)\nabla_{\Omega}\Psi = \frac{v(r)\mu}{f}\nabla_n(f\Psi)v(r) \frac{\left((1-\mu^2)C(r,\omega)+\mu\sqrt{1-\mu^2}\left(\Omega_{\parallel}\cdot\kappa^n\right)\right)}{g} \frac{\partial(g\Psi)}{\partial\mu} \tag*{(77)}

with some functions f,g>0f,g>0. For this to be in conservation form three conditions are required:

∇nv(r)μf=0⇒∇nvf=0(78)\nabla_n \frac{v(r)\mu}{f}=0 \Rightarrow\nabla_n \frac{v}{f}=0 \tag*{(78)}
∂∂μ(1−μ2)C(r,ω)+μ1−μ2(Ω∥⋅κn)g=0(79)\frac{\partial}{\partial\mu}\frac{(1-\mu^2)C(r,\omega)+\mu\sqrt{1-\mu^2}(\mathbf{\Omega}_{\parallel}\cdot\mathbf{\kappa}^{n})}{g}=0 \tag*{(79)}
μf∇nf=−(1−μ2)C(r,ω)+μ1−μ2(Ω∥⋅κn)g∂g∂μ(80)\frac{\mu}{f}\nabla_n f=-\frac{(1-\mu^2)C(r,\omega)+\mu\sqrt{1-\mu^2}(\mathbf{\Omega}_{\parallel}\cdot\mathbf{\kappa}^{n})}{g}\frac{\partial g}{\partial\mu} \tag*{(80)}

Where the last condition merely ensures equality of the right-hand sides of (A3) and (A4).

Through some manipulations and after abbreviating

K(r,μ,ω)=((1−μ2)C(r,ω)+μ1−μ2(Ω∥⋅κn))K(r,\mu,\omega)=\left((1-\mu^2)C(r,\omega)+\mu\sqrt{1-\mu^2}(\mathbf{\Omega}_{\parallel}\cdot\mathbf{\kappa}^{n})\right)

conditions (A5) and (A6) can be rewritten as

∇nvv=∇nff(81)\frac{\nabla_n v}{v}=\frac{\nabla_n f}{f} \tag*{(81)}
∂K∂μ=Kg∂g∂μ(82)\frac{\partial K}{\partial\mu}=\frac{K}{g}\frac{\partial g}{\partial\mu} \tag*{(82)}

Plugging (A9) into (A7) then leads to

∇nvv=∇nff=1μ∂K∂μ=C(r,ω)−1−2μ2μ1−μ2(Ω∥⋅κn).\frac{\nabla_n v}{v}=\frac{\nabla_n f}{f}=\frac{1}{\mu}\frac{\partial K}{\partial\mu}=C(r,\omega)-\frac{1-2\mu^2}{\mu\sqrt{1-\mu^2}}(\mathbf{\Omega}_{\parallel}\cdot\mathbf{\kappa}^{n}).

In this equation the left-hand side is independent of μ\mu and ω\omega, implying it can only hold if κn=0\mathbf{\kappa}^{n}=0 for all rr and if C(r,ω)=C(r)C(r,\omega)=C(r) is itself independent of ω\omega. The latter condition means the sectional curvature of Xn\mathbf{X}^{n} is the same in all directions, which from embedded surfaces leaves only planes and spheres. In both cases, conservation forms are known:

fplane=1gplane=1(83)f_{\mathrm{plane}}=1 \qquad g_{\mathrm{plane}}=1 \tag*{(83)}
fsphere=ρgsphere=1.(84)f_{\mathrm{sphere}}=\rho\qquad g_{\mathrm{sphere}}=1. \tag*{(84)}

All other cases fail to produce a conservation form via the chosen assumptions and Ansatz. Note that, while cylindrical coordinates, technically do not fulfill all assumptions used here, there still are conservation forms in the literature. These are however also based on a choice of nn where Xn\mathbf{X}^{n} are planes. Lastly note that this is not proof that conservation forms do not exist in general, but merely that, if they do, finding them is more complicated than using a simple Ansatz like we did.

References

  1. [1]Simon R. Arridge. Optical tomography in medical imaging. Inverse Problems, 15(2):R41–R93, 1999.DOI
  2. [2]G. I. Bell and S. Glasstone. Nuclear Reactor Theory. 1970, 10 1970. URL https://www.osti.gov/biblio/4074688.
  3. [4]Marcel Berger. Sur les groupes d’holonomie homogène des variétés à connexion affine et des variétés riemanniennes. Bull. Soc. Math. France, 83:279–330, 1955. ISSN 0037-9484. URL http://www.numdam.org/item?id=BSMF_1955__83__279_0.DOI
  4. [5]Carlo Cercignani. The Boltzmann Equation and Its Applications, volume 67 of Applied Mathematical Sciences. Springer, 1988.
  5. [6]Subrahmanyan Chandrasekhar. Radiative Transfer. Dover Publications, 1960.
  6. [8]J. Freimanis. On vector radiative transfer equation in curvilinear coordinate systems. Journal of Quantitative Spectroscopy and Radiative Transfer, 112(13):2134–2148, 2011. ISSN 0022-4073. . URL https://www.sciencedirect.com/science/article/pii/S0022407311001646. Polarimetric Detection, Characterization, and Remote Sensing.DOI
  7. [9]G. Frobenius. Ueber das pfaffsche problem. Journal für die reine und angewandte Mathematik, 1877(82):230–315, 1877. . URL https://doi.org/10.1515/crll.1877.82.230.
  8. [10]Tobias Kohn. Self-similar shrinkers in ℝ³, January 2008.
  9. [11]E. Kreyszig. Differential Geometry. Dover Books on Mathematics. Dover Publications, 2013. ISBN 9780486318622. URL https://books.google.com/books?id=Vw3CAgAAQBAJ.
  10. [12]Edward W. Larsen. The description of particle transport problems with helical symmetry. Nuclear Science and Engineering, 156(1):68–73, 2007. .DOI
  11. [13]John M. Lee. Introduction to Smooth Manifolds. 2000.
  12. [14]John M. Lee. Introduction to Riemannian Manifolds. Springer, Cham, Switzerland, second edition, 2018. . URL https://doi.org/10.1007/978-3-319-91755-9.
  13. [15]Arun Majumdar. Microscale heat conduction in dielectric thin films. Journal of Heat Transfer, 115(1):7–16, 1993.DOI
  14. [16]Dwight R. Nicholson. Introduction to Plasma Theory. John Wiley & Sons, 1983.
  15. [17]Reinhard Schlickeiser. Cosmic Ray Astrophysics. Springer, 2002.
  16. [18]J. Weingarten. Ueber eine klasse auf einander abwickelbarer flächen. Journal für die reine und angewandte Mathematik, 59:382–393, 1861. URL http://eudml.org/doc/147844.DOI

Paper details

vlnc:W_nhptq2b9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents